Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Liu C, Liu X, Wei Z, et al.

Deep learning-empowered predictive beamforming for IRS-assisted multi-


user communications[C]//2021 IEEE Global Communications Conference (GLOBECOM). IEEE, 2021:
01-07.

This paper implements deep learning to learn historical channel features and predict IRS phase shifts for the
next time slot, to maximize the average achievable sum-rate IRS multi-user system. The method can reduce
the channel estimation (CE) overhead.

In the system, the AP has M-element antenna array. The AP serves 𝐾 ≥ 1 single-antenna users with the
assistance of an IRS with 𝑁 = 𝑁𝑥 × 𝑁𝑦 passive reflecting elements.

The set the user index is: The k-th user is represented as:

The 3D coordinates of AP and IRS are:

The corrdinate of user 𝑈𝑘 at the 𝑡-th time sloe is:

The location of user in the next time slot is:

Where is the average velocity of 𝑈𝑘 at 𝑡-th time slot. The amplitude and phase:

and follow uniform distribution: ,

∆𝑇 is time slot duration, models environment uncertainty, where


denotes uncertainty offset on each axis.
For each time slot, only historical locations are known, which wil be adopted by the DL for predictive
beamforming.

1. Channel Model
Rician fading model is adopted for all the channels. The channel of the AP-to-IRS is formulated as:
Where 𝛽 𝐴𝐼 ≥ 0 is the Rician factor of the AP-to-IRS link. denotes the NLoS component.

The LoS component:


Where:

𝜃 and 𝜉 are the horizontal and vertical angle-of-arrivals from AP to IRS, which are indicated in Fig.1:
|ℎ𝐼 −ℎ𝐴 | |𝑦 | |𝑥𝐴 |
sin(𝜃 𝐴𝐼 ) = , cos(𝜉 𝐴𝐼 ) = 𝑑𝐴𝐼𝐼 , sin(𝜉 𝐴𝐼 ) = .
𝑑𝐴𝐼 𝑑𝐴𝐼

Where , 𝜆𝑐 denotes the wavelength of the carrier frequency,


∆𝑑𝐴 is the distance between two adjacent antenna elements of AP, ∆𝑑𝐼𝑦 and ∆𝑑𝐼𝑧 are the distances between
two adjacent IRS elements on y-axis and z-axis.
The channel of the IRS-to-Uk link at the 𝑡-th time slot is:

Where Rician factor is , NLoS component is

LoS component is:


Where:

Where ,

2. Signal Model
The IRS can reflect incident signals to the desired user via controlling the phase shift matrix:

𝜑𝑛,𝑡 denotes the phase shift of the n-th IRS element at the t-th time slot.
The received signal in 𝑈𝑘 at the 𝑡-th time slot is:

Where denote the beamforming vector and the transmitted data for 𝑈𝑖

at the t-th time slot. is the additive Gaussian noise in 𝑈𝑘 at the t-th time slot.
The received signal-to-interference-plus-noise ratio (SINR) is:

3. Transmission protocol
The CE in the system incurs large overhead because the IRS-associated cascaded channel has high dimension:

If the IRS phase shifts are preset, the instantaneous CSI (ICSI) reduce to a low-dimensional MISO channel:

The paper propose a hierarchical transmission protocol, the shift matrix at IRS adopted in next time slot has
been properly predicted and fixed in advance, thus only a MISO CE is required for optimizing the transmit
beamforming.

The transmission protocol consists of three phases. The channel coefficients and the locations of users change
over time slots, while they remain static within each time slot.

The predictive beamforming design for the 𝑡-th time slot consists (i) Phase 𝐼𝐼𝑡−1-B and (ii) Phase 𝐼𝑡

In (i), the AP obtains historical locations: through precious iplink transmission

and assembles the historical LoS channels: where 𝜏


is the number of available historical time slots. Then the AP predicts the LoS-based IRS phase shift matrix

to be adopted in time slot 𝑡.


In (ii), the IRS first sets the phase shift matrix as: for CE to acquire the effective ICSI

Then the information is exploited to optimize the ICSI-based transmit beamforming matrix
By the method, the protocol reduces the CE overhead while maintain the beamforming performance.

Problem Formulation
The paper aim to maximize the average sum-rate by jointly optimizing the LoS-aware phase shift matrix at
IRS and the ICSI-aware beamforming matrix at AP subject to the power constraint at AP and the phase shift
constraint at IRS.

For any time slot the optimization problem is formulated as:

Where and denote the LoS-aware


predictive phase shift matrix and the ICSI-aware beamforming matrix for time slot t. The inner maximization
is over the ICSI-aware 𝑊𝑡 with the optimized from the outer maximization. The outer maximization is

over the LoS-aware predictive . The expectation is taken over all random realizations of ,
given the historical LoS channels

Where
The problem is challenging.

Predictive beamforming for IRS-MUC system


The paper proposed an unsupervised DL-based predictive beamforming scheme. An LA-CLNet is designed

to address the outer maximization for acquiring the optimized and an IA-FNN is developed to address

the inner maximization to obtain optimized exploiting the ICSI with


1. LA-CLNet for LoS-aware predictive beamforming design
The cascaded LoS channel of 𝑈𝑘 at time slot 𝜆 is denoted by:

Let the diagonal elements of denoted by:


The outer maximization can be solved sub-optimally through:
Where is taken over all random realizations of given the historical cascaded LoS

channels: where
It is intractable because only the historical locations are available. The paper proposed the LA-CLNet to learn
the channel features and directly predict the beamforming matrix.

The LA-CLNet consists 1 input layer, 1 CNN module, 1 flatten layer, 1 concatenate layer, 1 LSTM module,
1 fully-connected layer, 1 output layer. The hyperplane is:

The input is: where is the mapping


function. The LA-CLNet is formulated as:

Where denotes the common layers (input layer to LSTM module) with common parameter 𝜔0 , and

is the FC layer with parameter 𝜔𝑖 :


Given unlabeled training set:

Where is the 𝑖-th training sample of 𝒳.


The cost function of LA-CLNet is:

With

Where is the network parameters. denote the complex-valued

vectors based on the output of LA-CLNet with an input of


After training via BP algorithm, the well-trained LA-CLNet is obtained:

And

Where and are the well-trained phase shifts and the beamforming matrix. are the

well-trained network parameters. is an auxiliary parameter for optimizing

2. IA-FNN for ICSI-aware Instantaneous beamforming design

Given we only need to obtain the effective ICSI:

Where is an equivalent MISO channel of AP-IRS-Uk , and can be acquired via


existing CE methods at Phase I of each time slot. The aim is to design an ICSI-aware beamforming matrix

to further enhance the sum-rate of the system. The inner maximization problem can be formulated as:

An IA-FNN is proposed to optimize exploiting the effective ICSI . The IA-FNN consists of 1
input layer, several hidden layers, and 1 output layer. The input is:

The IA-FNN can be formulated as: . ς is network parameter.


Given the effective ICSI as the training example, the cost function of IA-FNN is:

With:

Where is a complex-valued vector based on the output of IA-FNN with an input of . After

online training via the BP algorithm, the well-trained beamforming matrix can be expressed as:

Where is well-trained parameters.


Wei Y, Zhao M M, Zhao M J. Model-Driven GAN-Based Channel Modeling for IRS-Aided Wireless
Communication[C]//2021 IEEE Global Communications Conference (GLOBECOM). IEEE, 2021: 1-6.

This paper proposed a GAN-based channel modeling framework for learning the reflected channel distribution
in IRS system. The IRS-GAN framework characterizes the distribution of the reflected channel for an IRS-
aided MISO system, and the special structure of the reflected channels is taken into consideration in the
generative model. The WGAN is applied during the training process.

System Model

N-element IRS is deployed to improve communications from M-antenna BS to single-antenna user. The
baseband equivalent channels of BS to user, IRS to user, and BS to IRS, are denoted by:

The reflection coefficient vector is:


The 𝑛-th element 𝜃𝑛 satisfies:

The 𝑁 × 𝑁 diagonal reflection coefficient matrix is:


The received signal is:

Where is AWGN, is the complex transmitted signal, 𝑠 is information symbol


and 𝑤 is active precoding vector. The BS-IRS and IRS-user channels are difficult to get. The reflected
channel from BS to user via IRS is defined as:

Then the received signal can be rewritten as:

Based on the knowledge of the instantaneous CSI of the BS-user and BS-IRS-user links, the active transmit
precoding vectors 𝑤 and passive phase shifts 𝜃 can be separately or jointly optimized to improve the system.

In actual, the actual channel may not follow classical channel models, such as Rayleigh, Rician, and so on.
The paper considered Rician fading channel first. The channels are modeled as:

Where , with 𝛽 denoting


the Rician factors of the BS-user, BS-IRS, and IRS-user channels, 𝑙 are corresponding pass loss.
The LoS components are:

The NLoS components are:


For accurate channel model is difficult to obtain, especially for MIMO system, the paper proposes the model-
driven GAN-based channel modeling method.

IRS-GAN
GAN has been used for distribution learning.

In IRS-GAN framework, the cascaded BS-IRS-user channel is taken into consideration to learn the channel
distribution. ℱ𝐷 and ℱ𝐺 are the discriminative model and generative model. ℱ𝐺 transforms random noise
into a channel sample, and ℱ𝐷 accept the generated sample or a true channel sample to produce a real value
denoting the probability that the input sample is from the true channel distribution or not.

In the paper, the model-driven DL technique is proposed. It is combined with GAN to solve the IRS channel
modeling problem. The model-driven DL can accelerate the training process of neural networks. The IRS-
GAN framework is based on the prior knowledge that the reflected channel is the cascade of the BS-IRS and
IRS-user channels.

The n-th column of the reflected channel 𝐻 is the scalar-vector multiplication of the n-th entry of ℎ𝑟 and
the n-th row of 𝐺:

Thus, the paper proposed a three-node generative model ℱ𝐺 , which consists of a BS-IRS (BI) node 𝒩𝐵𝐼 , an
IRS-user (IU) node 𝒩𝐼𝑈 and a Cascading (C) node 𝒩𝐶 , as shown in Fig.3.

Node 𝒩𝐵𝐼 transform the input random noise vector 𝑧𝐵𝐼 from a fixed distribution (uniform or Gaussian
distribution) into a matrix 𝐺̃ , which approximates the BS-IRS channel 𝐺, CNN is used for extracting feature.
The BS-IRS channel matrix 𝐺 is generally high-dimensional and spatially correlated.

In the paper, the channel 𝐺 is treated as a CNN network: For the complex elements, the
paper regards the real and imaginary parts of 𝐺 as two channels with real entries. For the LoS component in

𝐺 does not change much over time, an additional bias term: is introduced into 𝒩𝐵𝐼 to
represent the LoS component. The BBI is independent of the input noise vector 𝑧𝐵𝐼 . The output of 𝒩𝐵𝐼
denoted by 𝐺̃ is:
The 𝒩𝐼𝑈 is designed to represent the IRS-user channel ℎ𝑟 by passing a random noise vector 𝑧𝐼𝑈 through a

multi-layer fully connected network then adding a bias term :

Where ℎ̃𝑟 denotes the output of 𝒩𝐼𝑈 and 𝜃𝐼𝑈 is the parameters of the FNN. Since the IRS-user channel
ℎ𝑟 is a vector, the FNN structure is implied (CNN can also be used).
Finally, the node 𝒩𝐶 is designed to first concentrate the output of 𝒩𝐵𝐼 and 𝒩𝐼𝑈 , then pass the concentrated
results through a few fully-connected layers FNNC to further approximate the reflected channel and output the
generated reflected channel samples {𝐻 ̃ }:

Where , 𝜃𝐶 is network parameter.

Let denote the overall learnable parameters in the proposed generative model
ℱ𝐺 . Then ℱ𝐺 can be expressed as:

For the discriminative model ℱ𝐷 , for the reflected channel 𝐻 can be viewed as a matrix whose columns are
scaled versions of those of the BS-IRS channel 𝐺 and thereby inherits the spatial correlation in 𝐺, the multi-
̂ } to determine whether 𝐻
layer CNND is employed to extract the features in the input channel samples {𝐻 ̂ is
from true channel distribution or not. Let 𝜃𝐷 denote the network parameters, ℱ𝐷 can be expressed as:

The 𝐻 ̂ can be the true channel sample 𝐻 or the generated channel sample 𝐻
̃ , and the output 𝑝̃ represents
̂ is drawn from the true channel distribution.
the probability that 𝐻

Learning strategy
The training dataset is constructed by generating reflected channel samples according to the known Rician
channel model before. The reflection coefficient matrix is set as discrete Fourier transform matrix. Gaussian
noise vectors 𝑧𝐵𝐼 and 𝑧𝐼𝑈 fed into ℱ𝐺 follow complex normal distribution:

Biases in 𝒩𝐵𝐼 and 𝒩𝐼𝑈 : are randomly generated from the


noise variances are hyper-parameters.

The loss functions are generated by Wasserstein distance for training the generator and discriminator:

Where represents the probability that ℱ𝐷 decides that a true channel sample comes from the

true channel dataset, and represents the probability that ℱ𝐷 decides that a
generated fake channel sample is from the true channel dataset. The max⁡(. ) function confines the output of
ℱ𝐷 in [0,1]. The generator and discriminator are expected to play a two-player minimax game:

(11) and (12) are based on the Wasserstein distance and thus the IRS-GAN is easier to train and converge.

The training process of the IRS-GAN framework consists of two loops, in each of the outer loop, three inner
loops including a generative loop, a discriminative loop and a testing loop are executed sequentially.
In the generative and discriminative loop, ℱ𝐺 and ℱ𝐷 is trained with a fixed number of iterations 𝐼𝐺 and
𝐼𝐷 . In the testing loop, the coarse estimate of the Wasserstein distance between the true channel distribution
and the generated channel distribution is calculated:

The distribution is saved for the next testing loop. The is related with the convergence and the generated
sample quality of ℱ𝐺 .
Let 𝛼𝐼 , {𝜖1 , 𝜖2 } and 𝑙 denote the initial learning rate, two threshold values and the outer iteration index
respectively.

When , the training process is terminated.

If , the learning rate is set to 𝜏𝛼𝐼 with 𝜏 denoting a decaying factor.


Otherwise, the next outer loop starts with 𝛼𝐼 .

To speed up the training procedure, the three schemes is employed in learning strategy:

First, train Θ𝐷 by minimizing the sum of ℒ𝐷 and gradient penalty term :

Where M is batchsize, denoting a random coefficient which is fixed


before training. The gradient penalty is an alternative to clipping weights and can enable more stable training.

Second, a moment penalty term is added to ℒ𝐺 such that ℱ𝐺 is forced to focusing on learning the statistical
information contained in the true channel samples. In the generative loop, the corresponding network

parameters are learned by minimizing

Where

is the moment penalty term, 𝜇 denotes the moment penalty coefficient. As such,
ℱ𝐺 is expected to produce a distribution that has a similar expectation as the true channel distribution. It is
helpful for accelerating training process especially when the channel has a large LoS component.
Finally, the initial learning rate for training the bias terms is set to be larger than the

learning rates for other parameters . It enables the IRS-GAN to learn the
LoS component quickly and then gradually approximate the distribution of the NLoS component.

In the simulation, the paper shows that the distribution of the generated channel samples approximates the
distribution of the proposed network gradually during the training process.

The proposed channel modeling is used in other beamforming scheme. For the beamforming task, the channel
models are assumed to be known. The simulation shows that if the generated channel samples is used for
beamforming scheme, the average achievable sum-rate can approximate that got by the beamforming scheme
based on the true channel samples.

You might also like