Neural Network-Based Fading Channel Prediction: A Comprehensive Overview

Wei Jiang,Hans D Schotten

doi:10.1109/access.2019.2937588

Abstract

By adapting transmission parameters such as the constellation size, coding rate, and transmit power to instantaneous channel conditions, adaptive wireless communications can potentially achieve great performance. To realize this potential, accurate channel state information (CSI) is required at the transmitter. However, unless the mobile speed is very low, the obtained CSI quickly becomes outdated due to the rapid channel variation caused by multi-path fading. Since outdated CSI has a severely negative impact on a wide variety of adaptive transmission systems, prediction of future channel samples is of great importance. The traditional stochastic methods, modeling a time-varying channel as an autoregressive process or as a set of propagation parameters, suffer from marginal prediction accuracy or unaffordable complexity. Taking advantage of its capability on time-series prediction, applying a recurrent neural network (RNN) to conduct channel prediction gained much attention from both academia and industry recently. The aim of this article is to provide a comprehensive overview so as to shed light on the state of the art in this field. Starting from a review on two model-based approaches, the basic structure of a recurrent neural network, its training method, RNN-based predictors, and a prediction-aided system, are presented. Moreover, the complexity and performance of predictors are comparatively illustrated by numerical results.

Highlights

By adapting radio transmission parameters, e.g., the constellation size, coding rate, transmit power, precoding codeword, time and frequency resource block, transmit antennas, and relays, to instantaneous channel conditions, adaptive wireless systems can potentially aid the achievement of great performance
In a frequency-division duplex system, the channel state information (CSI) is estimated at the receiver and fed back to the transmitter, where the obtained CSI might be already outdated before its actual usage owing mainly to the feedback delay
Starting from a brief review on AR and parametric models, the basic structure of a recurrent neural network, its training method based on the back-propagation algorithm, and several variants of RNN predictors applied for different scenarios ranging from frequency-flat singleantenna channels to frequency-selective multiple-input multiple-output (MIMO) channels, are presented

Summary

INTRODUCTION

By adapting radio transmission parameters, e.g., the constellation size, coding rate, transmit power, precoding codeword, time and frequency resource block, transmit antennas, and relays, to instantaneous channel conditions, adaptive wireless systems can potentially aid the achievement of great performance. Starting from a brief review on AR and parametric models, the basic structure of a recurrent neural network, its training method based on the back-propagation algorithm, and several variants of RNN predictors applied for different scenarios ranging from frequency-flat singleantenna channels to frequency-selective MIMO channels, are presented. SNt (t) T is the Nt ×1 vector of transmitted signals, n(t) stands for the vector of additive white noise, H(t)= hnr nt (t) Nr ×Nt is the matrix of continuous-time channel impulse responses, and hnr nt ∈C1×1 represents the gain of the flat fading channel between transmit antenna nt and receive antenna nr , where 1 nr Nr and 1 nt Nt. Due to feedback and processing delays, the obtained CSI at the transmitter may be outdated before its actual usage, i.e., H(t) = H(t+τ ), resulting in severe performance degradation of adaptive transmission systems [1] - [12]. It implies that the estimation process needs to be periodically and frequently conducted, resulting in extreme high complexity, which is unattractive from the viewpoint of practical implementation

AUTOREGRESSIVE MODEL

FREQUENCY-SELECTIVE MIMO PREDICTION

PERFORMANCE

CONCLUSION