Page 26 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4

and angular domain sparsity that mmWave channels ex‐ model possible block sparsity patterns among the
hibit. In this approach, the channel estimation problem is consecutive Angle Of Arrivals (AoAs) and Angle Of
formulated as a sparse recovery problem [4]. Such com‐ Departures (AoDs). irst obtain
pressive sensing based estimation techniques were irst time‐domain channels from the provided training
developed for frequency‐ lat hybrid mmWave MIMO sys‐ dataset inverse Discrete Fourier
tems [5, 6]. Recently, frequency‐selective channels with Transform (DFT) and remove the channel taps
OFDM‐based communications leading to a more complex with small magnitude. Then, apply
estimation problem have also been considered, with dif‐ ground truth time‐domain
ferent approaches to exploit the sparse channel charac‐ obtain sparse representations.
teristics [4, 7, 8]. Several model‐based signal processing Using joint angular distribution learned from
techniques for mmWave channel estimation under vari‐ training data, we re ine the grids and pattern‐
ous system settings can be found in [9–23]. coupling r t t improve
channel quality This appr is
Machine Learningand Arti icialIntelligence (ML/AI)have
called “Pattern‐Coupled Bay Learning
been shown to be powerful tools in diverse areas such as
for Channel Estimation with Dominating Delay Taps
natural language processing, speech processing, and im‐
(PCSBL‐DDT)” in the paper.
age recognition, where it is challenging to design speci ic
model‐based algorithms. However, the impact of ML/AI 3. The third approach, Projection Cuts Orthogonal
on the design and optimization of communication sys‐ Matching Pursuit (PC‐OMP), is based on the Orthogo‐
tems is yet to be extensively studied, especially under re‐ nal Matching Pursuit (OMP) algorithm. This method
alistic and practically meaningful settings. We aim to ad‐ makes use of the sparsity of the mm‐wave channel
dress some of the aspects of ML/AI in wireless communi‐ to extract channel components. At each iteration of
cations here. theOMPalgorithm, acoarseestimateofthestrongest
path parameters (AoA, AoD, and delay) is obtained
In this paper, we study the potential advantage of us‐ by a low resolution grid search. Then, each of the
ing data‐driven approaches for channel estimation in hy‐ three parameters is re ined alternately, assuming the
brid MIMO systems. The model‐cum‐data driven algo‐ other two to be known. In this way, we keep the algo‐
rithms we develop in this paper were selected as the top rithm’s complexity low without compromising on its
three solutions in the “ML5G‑PHY Channel Estimation accuracy. At the end of each iteration, a path detec‐
Global Challenge 2020” organized by the International tion hypothesis is tested, and, if successful, the path
1
Telecommunication Union (ITU) . Our main goal in this is subtracted from the channel. This process is re‐
paper is to present and contrast these three algorithms
peated until no additional path is detected.
for estimating an mmWave channel in a hybrid MIMO
system. We compare the Normalized Mean Squared Er‐
Notation
ror (NMSE) performance of these approaches and discuss
∗
the machine learning techniques relevant for the chal‐ The operator (⋅) represents the conjugate transpose or

̄
lenge at hand. These approaches utilize the channel train‐ conjugate for a matrix or a scalar, respectively. A, A , and
†
ing datasets generated using the Raymobtime tool to cus‐ A denote the conjugat tr and Moore‐Penrose
tomize the algorithms so that they perform well for a test pseudoinverse of a matrix A, respectively The multivari‐
dataset generated in a similar environment [24]. at comple vector
and covariance matrix C is denoted by ( , C) and its
We provide a brief overview of the three solutions below:
probability density function (pdf) of a random vector x is
1. greedy search high‐ denoted by (x| , C) blkdiag(⋅) represents the block‐
performing inference method irst diagonal part of a matrix. diag(X) or diag(x) represents
approach. Multi‐Level Greedy Search a vector obtained by the diagonal elements of the matrix
sparsifying virtual beamspace X or diagonal of
x diagonal, respectively A ⊗ B denotes Kro‐
dictionary that reduces of the

problem and use the learned dictionary to estimate necker product of the matrices A and B ||A|| denotes
the Frobenius norm of a matrix A. ⟨a, b⟩ is the inner prod‐
the channel using a Sparse Bayesian Learning (SBL)
method. We inally exploit the delay‐domain sparsity uct of the two vectors a and b The trace of a matrix A is
to de‐noise the estimated channels. We name the al‐ denoted by tr(A) Tx and Rx denote the transmitter and
gorithm as MLGS‐SBL. receiver, respectively. We use (A) to vectorize the ma‐
trix A column‐wise. [⋅] denotes the expectation.
2. second propose SBL‐
based algorithm to exploit the sparsity of the chan‐ 2. SYSTEM MODEL
nel. We utilize the pattern-coupling concept to
W consider sing cell mmWav hy MIMO‐OFDM
1
https://www.itu.int/en/ITU‐T/AI/challenge/2020/Pages/default.aspx system with antennas at the transmitter (Tx) and
2
The order in which the algorithms are presented is unrelated to their antennas at the receiver (Rx), as shown in Fig.
ranking in the ITU ML5G‐PHY channel estimation challenge. The or‐
dering is based on ease of presentation and readability of the paper.

10 © International Telecommunication Union, 2021