Thanks to the low cost and power consumption, hybrid analog-digital architectures are considered as a promising energy-efficient solution for massive multiple-input multiple-output (MIMO) systems. The key idea is to connect one RF chain to multiple antennas through low-cost phase shifters. However, due to the non-convex objective function and constraints, we propose a gradient-guided meta-learning (GGML) based alternating optimization framework to solve this challenging problem. The GGML based hybrid precoding framework is \textit{free-of-training} and \textit{plug-and-play}. Specifically, GGML feeds the raw gradient information into a neural network, leveraging gradient descent to alternately optimize sub-problems from a local perspective, while a lightweight neural network embedded within the meta-learning framework is updated from a global perspective. We also extend the proposed framework to include precoding with imperfect channel state information. Simulation results demonstrate that GGML can significantly enhance spectral efficiency, and speed up the convergence by 8 times faster compared to traditional approaches. Moreover, GGML could even outperform fully digital weighted minimum mean square error (WMMSE) precoding with the same number of antennas.

翻译：得益于低成本与低功耗的优势，混合模拟-数字架构被视为大规模多输入多输出（MIMO）系统中一种极具前景的高能效解决方案。其核心思想是通过低成本移相器将单个射频链路连接至多个天线。然而，由于目标函数与约束条件的非凸性，我们提出了一种基于梯度引导元学习（GGML）的交替优化框架来解决这一具有挑战性的问题。该基于GGML的混合预编码框架具有“免训练”与“即插即用”的特性。具体而言，GGML将原始梯度信息输入神经网络，利用梯度下降从局部视角交替优化子问题，同时嵌入元学习框架内的轻量级神经网络从全局视角进行更新。我们还将所提框架扩展至包含非完美信道状态信息的预编码场景。仿真结果表明，GGML能显著提升频谱效率，且收敛速度较传统方法加快8倍。此外，在相同天线数量下，GGML甚至能够超越全数字加权最小均方误差（WMMSE）预编码的性能。