Meta-Learning for Hybrid Precoding in Millimeter Wave MIMO System

TL;DR

This paper introduces a novel gradient-guided meta learning approach for hybrid precoding in millimeter wave MIMO systems, achieving superior spectral efficiency and robustness without pre-training.

Contribution

It proposes a pre-training free, meta learning-based hybrid precoding method that improves performance and robustness over traditional algorithms and deep learning solutions.

Findings

Outperforms existing hybrid precoding methods in simulations.

Demonstrates robustness to system parameter variations.

Can surpass fully digital WMMSE precoding performance.

Abstract

The hybrid analog/digital architecture that connects a limited number of RF chains to multiple antennas through phase shifters could effectively address the energy consumption issues in massive multiple-input multiple-output (MIMO) systems. However, the main challenges in hybrid precoding lie in the coupling between analog and digital precoders and the constant modulus constraint. Generally, traditional optimization algorithms for this problem typically suffer from high computational complexity or suboptimal performance, while deep learning based solutions exhibit poor scalability and robustness. This paper proposes a plug and play, free of pre-training solution that leverages gradient guided meta learning (GGML) framework to maximize the spectral efficiency of MIMO systems through hybrid precoding. Specifically, GGML utilizes gradient information as network input to facilitate the sharing of gradient information flow. We retain the iterative process of traditional algorithms and leverage meta learning to alternately optimize the precoder. Simulation results show that this method outperforms existing methods, demonstrates robustness to variations in system parameters, and can even exceed the performance of fully digital weighted minimum mean square error (WMMSE) precoding with the same number of antennas.