摘要
为了降低毫米波MIMO系统中现有交替最小混合预编码算法和基于动量梯度下降的在线学习混合预编码方案的复杂度,针对单用户通信场景,重新考虑混合预编码器的设计问题,提出一种等效的单隐藏层神经网络架构。在该架构下,数字预编码矩阵和模拟预编码矩阵的每一个组成元素可等效为单隐藏层神经网络的连接权值,其最优解可通过神经网络中的权值训练方法获得。在此基础上,结合反向传播算法,提出一种基于自适应梯度反向传播的混合预编码机制。进一步地,将所提算法扩展到多用户通信场景。仿真结果表明,在单用户场景和多用户场景下,所提算法可实现的频谱效率均接近全数字预编码,同时复杂度低于现有的基于交替最小的混合预编码算法和基于动量梯度下降的在线学习混合预编码方案。
To reduce the complexity of existing hybrid precoding algorithms based on alternating minimization(AltMin)and online learning via gradient descent with momentum in mmWave MIMO systems,aiming at the single-user scenario,the problem of designing the hybrid precoder was reconsidered and an equivalent single hidden layer neural network was proposed.Under the new architecture,the elements of the digital and analog precoder were equivalent to the connecting weights of a single hidden layer neural network,and their optimal solution could be obtained via the weights training method.Inspired by the back propagation(BP)algorithm in feed forward neural networks,an adaptive gradient(AG)-based BP algorithm for hybrid precoding was proposed.Furthermore,the proposed algorithm was extended to the multi-user scenario.The numerical results show that the proposed algorithm achieves approximately the same spectral efficiency as the fully-digital precoding in both the single-user and multi-user scenarios,while has lower complexity than the existing AltMin-based hybrid precoding algorithms and online learning hybrid precoding based on gradient descent with momentum.
作者
张煜
张治
董晓岱
ZHANG Yu;ZHANG Zhi;DONG Xiaodai(School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China;Department of Electrical and Computer Engineering,University of Victoria,Victoria V8W 3P6,Canada)
出处
《通信学报》
EI
CSCD
北大核心
2021年第10期95-105,共11页
Journal on Communications
基金
国家重点研发计划基金资助项目(No.2019YFC1511302)
国家自然科学基金资助项目(No.61629101)。
