Professional Documents
Culture Documents
Digicomm Paper - 2
Digicomm Paper - 2
Digicomm Paper - 2
7, JULY 2022
Abstract— Hybrid precoding is a cost-efficient technique for mmWave MIMO system [2]–[5]. In the hybrid precoding
millimeter wave (mmWave) massive multiple-input multiple- architecture, expensive fully-digital precoder is replaced by
output (MIMO) communications. This paper proposes a deep a network of low-cost phase shifters, called analog precoder,
learning approach by using a distributed neural network for
hybrid analog-and-digital precoding design with limited feed-
followed by a low-dimensional digital precoder.
back. The proposed distributed neural precoding network, called However, the design of hybrid precoding is a challenge
DNet, is committed to achieving two objectives. First, the DNet even when the base station (BS) obtains perfect channel state
realizes channel state information (CSI) compression with a dis- information (CSI). Nonconvex constant modulus constraints
tributed architecture of neural networks, which enables practical are imposed on the analog precoding parameters, which come
deployment on multiple users. Specifically, this neural network is from the physical constraints of the phase shifter network.
composed of multiple independent sub-networks with the same
structure and parameters, which reduces both the number of Due to this nonconvexity, the optimization of hybrid precoding
training parameters and network complexity. Secondly, DNet is difficult to solve by traditional convex optimization tools.
learns the calculation of hybrid precoding from reconstructed Iterative algorithms, e.g., [2]–[8] by using alternating opti-
CSI from limited feedback. Different from existing black-box mization were proposed for calculating the hybrid precoding,
neural network design, the DNet is specifically designed according which aimed at achieving near-optimal performance in the
to the data form of the matrix calculation of hybrid precoding. case of narrowband or wideband. Alternatively, researchers
Simulation results show that the proposed DNet significantly
improves the performance up to nearly 50% compared to tried to solve the hybrid precoding design by exploiting
traditional limited feedback precoding methods under the tests deep learning (DL) technologies [9]–[13]. In [11], the authors
with various CSI compression ratios. transformed the channel matrix to an image and then designed
Index Terms— Distributed neural network, hybrid precoding, convolutional neural networks (CNN) for the precoding design
limited feedback, millimeter wave (mmWave) MIMO. in massive MIMO. The study in [13] introduced an unsuper-
vised DL-based approach to calculate the hybrid precoding and
I. I NTRODUCTION achieved performance improvement compared to the super-
vised DL-based methods.
M ILLIMETER wave (mmWave) massive multiple-input
multiple-output (MIMO) communication systems uti-
lize the wide bandwidth of mmWave to achieve high transmis-
The above methods all require the BS to know perfect CSI.
However, in a frequency division duplexed (FDD) system, CSI
is estimated by receivers and then fed back to the BS via a
sion rates. It has been a key technology in the fifth-generation
limited feedback link. As the number of antennas increases,
mobile communication system [1]. One of the major chal-
the channel matrix dimensions and the number parameters to
lenges faced by mmWave massive MIMO is that, with the
feedback rise sharply. Therefore, it is of practical interest to
rapid increase in the number of antennas, equipping each
design hybrid precoding with limited feedback. The authors
antenna with a separate radio-frequency (RF) chain can
of [6] proposed an analog precoding design from a given CSI
result in high power consumption and hardware cost. This
codebook so that only a low-dimensional equivalent matrix in
challenge makes the traditional fully-digital precoding archi-
the form of a codeword is fed back. An alternative solution
tecture difficult to implement in practice. To address this
is to apply compressed sensing (CS) [14] or DL-based data
issue, hybrid analog-and-digital precoding was proposed for
compression methods [15]–[17] to compress and reconstruct
Manuscript received 29 March 2022; accepted 10 April 2022. Date of the CSI. Deep reinforcement learning (DRL) based beam
publication 18 April 2022; date of current version 12 July 2022. This work pattern design was presented in the [19], which learned the
was supported in part by the NSFC under grants 62022026, 61871109, beam pattern without the explicit knowledge of the channels.
6211101429, and 61771431. The associate editor coordinating the review of
this letter and approving it for publication was A. Li. (Corresponding authors: The authors of [20] combined channel estimation with hybrid
Jindan Xu; Wei Xu.) precoding design and proposed a quantized network to calcu-
Kai Wei and Jindan Xu are with the National Mobile Communications late quantized analog precoding.
Research Laboratory (NCRL), Southeast University, Nanjing 210096, China
(e-mail: kaiwei@seu.edu.cn; jdxu@seu.edu.cn).
In this study, we propose a distributed neural network,
Wei Xu is with the National Mobile Communications Research Laboratory called DNet, for the hybrid precoding design with reduced
(NCRL), Southeast University, Nanjing 210096, China, and also with the number of channel feedback parameters. The design of DNet
Henan Joint International Research Laboratory of Intelligent Networking and is based on a distributed architecture of neural networks,
Data Analysis, Zhengzhou University, Zhengzhou 450001, China (e-mail:
wxu@seu.edu.cn). which can be distributively deployed at the users and BS
Ning Wang is with the Department of Electrical Engineering, Zhengzhou after the network training is completed to achieve implicit CSI
University, Zhengzhou, Henan 450001, China (e-mail: ienwang@zzu.edu.cn). compression and feedback. The network that performs the CSI
Dong Chen is with Xiaomi Corporation, Beijing 100085, China (e-mail:
chendong7@xiaomi.com). compression is composed of multiple sub-networks with the
Digital Object Identifier 10.1109/LCOMM.2022.3167712 same structure and parameters, so that the number of training
1558-2558 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on April 06,2023 at 05:52:09 UTC from IEEE Xplore. Restrictions apply.
WEI et al.: DISTRIBUTED NEURAL PRECODING FOR HYBRID mmWave MIMO COMMUNICATIONS 1569
Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on April 06,2023 at 05:52:09 UTC from IEEE Xplore. Restrictions apply.
1570 IEEE COMMUNICATIONS LETTERS, VOL. 26, NO. 7, JULY 2022
neural network based techniques for CSI compression, e.g., in of following neural network. The following layers include
[15], [16], the proposed DisNet is based on a distributed two residual networks, named ResNet. Each of the ResNet
neural network architecture which is therefore suitable for consists of four Conv layers. From the first Conv layer to
deployment in multiuser mmWave MIMO. In the DisNet, each the fourth Conv layer, the number of convolutional kernels
encoder-decoder network pair works for one UE. The channel are respectively set to 4, 8, 16, and 2. The corresponding
matrix between the k-th UE and BS, denoted by hH k , is a dimensions of these convolutional kernels are, respectively,
1 × Nt -dimensional complex vector. Thus, in our proposed chosen as 2 × 3, 4 × 3, 8 × 3, and 16 × 3. We use x and y to
encoder-decoder network, we choose a one-dimensional (1D) respectively represent the input and output of the ResNet while
convolutional (Conv) layer and a fully-connected (FC) layer let Res(·) denote the operations of the four Conv layers in the
as the basic unit to compress and recover the CSI. And the ResNet. Then the output of each ResNet can be expressed as
activation function is chosen as y = x + Res(x). (9)
x, x≥0 Then, all the outputs of the K decoders are combined at the BS
σ(x) = LeakyReLU(x) = (8)
0.3x, x < 0, to yield the H = [hH , hH , · · · , h
H ]H , as depicted in Fig.2,
1 2 K
to provide non-linearity and avoid dead net problem [21]. H
where hk regarded as the the implicit version of CSI of the
Concretely for the kth encoder-decoder network pair, the k-th UE.
encoder consists of one 1D Conv layer and one FC layer. Note that the proposed encoder-decoder network pairs are
The former performs CSI feature extraction, while the latter set to the same structure as described above, in terms of not
performs CSI feature data compression. Thus, hk would be only the architecture but also the value of parameters. This
reshaped to [R(hk ), I(hk )]H , which is a 2 × Nt tensor, where same setting for each sub-network helps reduce the complexity
R(hk ) and I(hk ) are the real and imaginary parts of hk , as the of the proposed DNet. In particular, it is found that the DisNet
inputs of the first Conv layer. The convolutional kernels in the component of the DNet actually learns the distribution of the
first Conv layer are of the number 4 and of the size 2×3, which channel data set, which allows the DisNet to be extendable.
is used to extract features of the channel, hk , and generates a
4×Nt tensor as the output. The output of the Conv layer is then B. PreNet for Hybrid Precode
flattened into a 4Nt -dimensional vector before fed to the FC The second component network at the BS is PreNet.
layer. This layer compresses the CSI feature data and generates As shown in the right part of Fig. 2, it learns to calculate
a feedback vector qk , which is a 2Mt -dimensional vector, the precoder according to the implicit CSI, H, obtained by
for limited feedback. Thus, the compression ratio can be also the DisNet. Specifically, the PreNet consists of two parts,
calculate as γ = Mt /Nt . The K encoder parts are trained i.e., RFNet and BBNet, generating analog-precoding and
for once and then distributively deployed at the corresponding digital-precoding, respectively.
UE. Each UE employs its encoder sub-network to generate
Consider that the input of RFNet, i.e., the implicit CSI, H,
limited feedback parameters, and sends them back to the BS. is a K × Nt -dimensional complex matrix. We then choose a
Once the BS receives those codewords, i.e., {q1 , · · · , qK }, two-dimensional (2D) Conv layer as the first layer in RFNet to
through limited feedback links from the UEs, we propose The real and imaginary parts of
extract internal features of H.
a subnetwork of decoder to recover the CSI vectors for
the H, i.e., [R(H); I(H)], are marked with different colors
all users. To account for feedback or quantization errors, in the image and are input of the 2D Conv layer. The two
we superimpose a noise on the input to the decoder. The convolutional kernels of this Conv layer are with the size of
first layer of the decoder is an FC layer, which converts a 3 × 3. And we apply zero padding to make sure that the size
2Mt -dimensional vector to a 2Nt -dimensional vector. Then of the output feature map remains to be K × Nt . The output is
this vector is reshaped to a 2 × Nt tensor, setting as the input flattened to a vector of size 2KNt × 1 and then sent to the FC
Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on April 06,2023 at 05:52:09 UTC from IEEE Xplore. Restrictions apply.
WEI et al.: DISTRIBUTED NEURAL PRECODING FOR HYBRID mmWave MIMO COMMUNICATIONS 1571
layer. The output of this FC layer is a KNt -dimensional vector sequentially by applying self-supervised learning and super-
and is then reversely reshaped into a K × Nt -dimensional vised learning. The loss function of DisNet is chosen as
matrix Θ. Then the analog precoding matrix from RFNet is 1
L1 = H − H F, (14)
expressed as KNt
= √1 (cosΘ + jsinΘ),
F (10) and the loss funtion of RFNet is
Nt 1 F.
L2 = F − F (15)
where j 2 = −1. Considering that the phase shifter is quan- KNt
tized, we quantize the output F in four bits after the network Note that when the DisNet is trained, the perfect CSI matrix H
training. is fed into the DisNet to generate the imperfect CSI matrix H,
Inspired by the PZF precoder [2], given the CSI H and as the input of RFNet. The training label F of RFNet is
analog precoding F, the digital precoding is performed as calculated from the perfect CSI using the previous method
D = Heq (Heq HH −1 in [2].
eq ) Λ, (11)
After the pre-training of DisNet and RFNet, we build
where Heq HF is the equivalent channel and Λ is a the whole DNet, including DisNet, RFNet, and BBNet, and
diagonal matrix for column power normalization. The issue initialize the DisNet part and the RFNet part of this new DNet
is that, from the DisNet, only implicit CSI H is available, with the learned parameters. The BBNet of DNet is randomly
which means we are not safe to directly use the PZF algo- initialized. We choose the sum rate of the MIMO system as
rithm or other traditional precoding calculations as in [3]–[7]. the optimization objective of DNet. It follows
To address this issue, the last part of PreNet, namely BBNet as
K P H 2
shown in Fig. 2, is designed to calculate the digital precoding K |hk Fdk |
L3 = − log2 1 + , (16)
from the output of RFNet, F, and the implicit CSI, H. 1+ P d
|hH F j |2
k=1 j=k K k
We notice that in the PZF algorithm, Heq is a diagonally i for i = 1, . . . , K is the ith column of D.
During
where d
dominant matrix when the channels of the UEs are nearly
the joint training stage, we not only train the BBNet part but
orthogonal. Inspired by this feature, we define a non-linear
also modify the parameters of the DisNet and RFNet of the
function diag(·) that extract all the (i, i)-th element of the
proposed DNet.
Heq , where i = 1, . . . , K. If Heq is a diagonal matrix, then
equation (11) is simplified to D = diag(H−1 −1
eq )Λ = Heq Λ.
Even if Heq is not a diagonally dominant matrix when the D. Extension of DNet
UEs are densely clustered, all the (i, i)-th elements of Heq When the number of UEs changes, we can easily change
are also the special elements of Heq , because the physical the number of these encoder-decoder pairs without additional
meaning of analog-precoding lies in beam alignment. It may training, as for a given subcarrier, where the distribution of
lead to a different distribution of diag(Heq ). According to channels remains the same. Also, when the number of data
this conjecture, we divide BBNet into two smaller neural streams is less than the number of RF chains, i.e., K < NRF ,
networks, i.e. Net a and Net b in Fig. 2. Net b inputs diag(Heq ) we suggest training the multiple groups of PreNet with dif-
and performs a rough estimate of the digital-precoding, while ferent numbers of UEs in advance, then switching the network
Net a inputs Heq and calculates corrections for the digital- based on the actual number of UEs. Meanwhile, although the
precoding. In this way, the digital-precoding from BBNet, D u, system model and the corresponding problem formulation and
is calculated as solution are described for a narrowband system, which can
u = Da + Db directly extend to a wideband orthogonal frequency division
D
multiplexing (OFDM) system, like in [22].
= σ(Wa H F + b) + σ(Wb diag(H F)
+ bb ), (12)
where σ is the activation functions in (8), and Wa , ba , IV. N UMERICAL R ESULTS
Wb , and bb respectively represent the equivalent weights and A. Dataset and Training
equivalent biases of Net a and Net b. Before ending BBNet,
we added a normalization module, which is defined as We choose an open-source MIMO channel data set, called
√ DeepMIMO [23], which is an OFDM system with 1024 sub-
KD u
D= (13) carriers, as the source of training data set for making simu-
D
F u F lation experiments under more realistic channels. The number
to meet the transmitting power constraint. of UE is 4 while the number of antennas at the BS is 64.
Considering that our model is based on a flat subband of
C. Network Training an OFDM, which corresponds to a flat narrowband fading
The proposed neural network realized the channel infor- channel [23]. Without loss of generality, we randomly sample
mation compression which is a typical encoder-decoder prob- the flat fading channel data on these subcarriers of the OFDM
lem. We cannot simply use supervised learning strategies to form the training data set of our proposed DNet.
like [18]. Inspired by the work in [9], we propose a joint
supervised-unsupervised learning approach. The entire train- B. Achievable Sum Rates
ing process is divided into two parts: pre-training and joint In Fig. 3, we study the relationship between the sum
training. rate and the SNR for different precoding schemes under
During the pre-training stage, we build the pre-training various channel parameter compression ratios. We see that the
network, DisNet and RFNet. DisNet and RFNet are trained proposed scheme outperforms the other schemes, including
Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on April 06,2023 at 05:52:09 UTC from IEEE Xplore. Restrictions apply.
1572 IEEE COMMUNICATIONS LETTERS, VOL. 26, NO. 7, JULY 2022
Authorized licensed use limited to: SHIV NADAR UNIVERSITY. Downloaded on April 06,2023 at 05:52:09 UTC from IEEE Xplore. Restrictions apply.