Machine Learning Inspired Energy-Efficient Hybrid Precoding with Fully-Adaptive-Connected Structure

Abstract

In the millimeter wave massive multiple-input multiple-output (MIMO) system, hybrid precoding is regarded as an effective solution to reduce hardware complexity and energy consumption. Due to the fixed connections between RF chains and antennas, the energy efficiency of hybrid precoding with fully-connected and sub-connected structure is limited. In this paper, we investigate a fully-adaptive-connected hybrid precoding with low-resolution phase shifters, where the connections between RF chains and antennas are dynamically controlled by switches. We formulate the problem of energy efficiency maximization as a non-coherent combining problem. Inspired by the cross-entropy (CE) optimization in machine learning, we propose a CE-based fully-adaptive hybrid precoding (CE-FAHP) algorithm, whose complexity is linear with the number of transmitted antennas. The probability mass function of all elements in analog precoder is updated iteratively by minimizing the CE, which is utilized to generate the near-optimal analog precoder. Simulation results demonstrate that the proposed algorithm can achieve higher energy efficiency than the existing matching assisted fully-adaptive hybrid precoding (MA-FAHP) algorithm with lower complexity.