Content uploaded by Bo Yang

Author content

All content in this area was uploaded by Bo Yang on Mar 09, 2021

Content may be subject to copyright.

1

Intelligent Spectrum Learning for Wireless Networks

with Reconﬁgurable Intelligent Surfaces

Bo Yang, Xuelin Cao, Chongwen Huang, Chau Yuen, Fellow, IEEE,

Lijun Qian, Senior Member, IEEE, and Marco Di Renzo, Fellow, IEEE

Abstract

Reconﬁgurable intelligent surface (RIS) has become a promising technology for enhancing the relia-

bility of wireless communications, which is capable of reﬂecting the desired signals through appropriate

phase shifts. However, the intended signals that impinge upon an RIS are often mixed with interfering

signals, which are usually dynamic and unknown. In particular, the received signal-to-interference-plus-

noise ratio (SINR) may be degraded by the signals reﬂected from the RISs that originate from non-

intended users. To tackle this issue, we introduce the concept of intelligent spectrum learning (ISL),

which uses an appropriately trained convolutional neural network (CNN) at the RIS controller to help

the RISs infer the interfering signals directly from the incident signals. By capitalizing on the ISL, a

distributed control algorithm is proposed to maximize the received SINR by dynamically conﬁguring the

active/inactive binary status of the RIS elements. Simulation results validate the performance improvement

offered by deep learning and demonstrate the superiority of the proposed ISL-aided approach.

Index Terms

Reconﬁgurable intelligent surface, intelligent spectrum learning, convolutional neural network.

I. INTRODUCTION

Due to the dynamic nature of the wireless environment that results in severe signal ﬂuctuations caused

by multipath fading and the presence of large obstacles, the wireless link between a desired cellular

B. Yang, X. Cao, and C. Yuen are with the Engineering Product Development Pillar, Singapore University of Technology and

Design, Singapore 487372 (e-mail: bo yang, xuelin cao, yuenchau@sutd.edu.sg).

C. Huang is with Zhejiang Provincial Key Lab of information processing, communication and networking, Zhejiang University,

No.38 Zheda Road, Hangzhou, 310007, P.R. China (e-mail: chongwenhuang@zju.edu.cn).

L. Qian is with the Department of Electrical and Computer Engineering and CREDIT Center, Prairie View A&M University,

Texas A&M University System, Prairie View, TX 77446, USA (e-mail: liqian@pvamu.edu).

M. Di Renzo is with Universit´

e Paris-Saclay, CNRS, CentraleSup´

elec, Laboratoire des Signaux et Syst`

emes, 3 Rue Joliot-Curie,

91192 Gif-sur-Yvette, France. (marco.di-renzo@universite-paris-saclay.fr)

2

BS

Distributed

RIS Controller 1

Reecting element

Reecting signal

CSI feedback

Reecting element

RIS 1

RIS K

CSI feedback

Distributed

RIS Controller K

U2

U1

UL

...

Reecting signal

λ

BS

Reecting element

CSI feedback

Reecting element

RIS 1

RIS K

CSI feedback

U1

UL

...

Reecting signal

Distributed RIS Controller 1

deployed with a trained CNN model

Distributed RIS Controller K

deployed with a trained CNN model

ON/OFF

control and

phase shift

RF signal

RF signal

ON/OFF

control and

phase shift

U2

No reecting signal

Other user signals

Desired user signals

（a）（b）

ISL

λ

Fig. 1: In (a), a traditional multiple-user uplink RIS-assisted wireless communication system with KRISs is shown, where we

assume that Lusers transmit to the BS at same time and frequency, and all RISs serve one user at a time. So, the other users

are considered as interferers. In (b), the proposed ISL-enabled RIS-assisted wireless communication system is shown. Each RIS

controller is deployed with a trained CNN model to identify the interfering users from the incident RF signals, so as to optimize

the operation of the RISs in a distributed way.

user and a base station (BS) may not be reliable enough or may even undergo a complete outage. To

tackle this issue, reconﬁgurable intelligent surfaces (RISs) have been proposed to improve the received

signal-to-interference-plus-noise ratio (SINR) at the users by appropriately reﬂecting the incident signals

and generating directional beams [1].

In the literature, some preliminary works investigated the optimization of RIS-assisted wireless com-

munications. In [2], a joint transmit power allocation and phase shift design was developed to maximize

the energy efﬁciency. In [3], the authors considered a downlink RIS-assisted multiuser communication

system and studied a joint transmission and reﬂection beamforming problem to minimize the total transmit

power. In [4], the channel estimation problem was investigated for an RIS-aided wireless communication

system by jointly optimizing the training sequence of the transmitter and the reﬂection pattern of the RIS.

Furthermore, an RIS-assisted anti-jamming solution was proposed for securing wireless communications

via reinforcement learning [5]. In [6], the authors introduced RISs in mobile edge computing systems,

where a joint design of computing and communications was developed to minimize the computational

latency.

However, most of the existing works assume either that no interference exists, which rarely occurs

in practice, or that the interference is known and can be taken into account, which is not trivial to

3

estimate since the interference is usually dynamically changing. These issues are exacerbated in RIS-

aided systems, since they are nearly-passive surfaces with no active sensing capabilities for channel and

interference estimation [7]. In order to elucidate the problem at hand, let us consider the case study

depicted in Fig. 1(a), in which we consider a practical scenario where all the RISs serve one desired user

(e.g., U1). In the analyzed case study, the other users (e.g., Ul,l∈[2, L]) are considered as interfering

users for U1at each RIS. As a result, the signal reﬂected by each RIS is a mixture of the desired signal

from U1and the interfering signals from the other users. In this context, the received SINR at the BS

via an RIS that does not account for the interfering signals could be even worse than the SINR of the

direct link. This impact of the interference is, in particular, more severe if the interfering devices are

close to the desired user, e.g., in Fig. 1(a) U2may cause severe interference to U1at RIS1when the

angle between them (i.e., λ) is small.

To overcome these challenges, in this paper, we empower a conventional RIS-assisted wireless system

with ‘intelligent spectrum learning (ISL) capabilities’, by leveraging appropriately trained convolutional

neural networks (CNN) at the RIS controller in order to predict/estimate the interfering devices from the

incident signals, as highlighted in Fig. 1(b). In the proposed system, the active-inactive (or ON-OFF) status

of each RIS1and the corresponding phase shifts need to be carefully optimized, since the interference

distribution at each RIS is, in general, different. The corresponding SINR maximization problem turns

out to be a mixed-integer nonlinear program (MINLP), which is usually difﬁcult to solve. To tackle this

issue, we decompose the original problem into two subproblems, which are solved independently and in a

distributed manner. The proposed solution equips a conventional RIS with the capability to dynamically

‘think-and-decide’ whether reﬂecting or not the incident signals through the proposed distributed ISL

principle.

II. SY ST EM MODEL AND PRO BL EM FORMULATION

We consider an RIS-assisted uplink wireless system that consists of one BS, a set Kof KRISs, and

a set Lof Lusers, where K≥1and L>Kusually hold. We assume that the BS allocates all the

RISs to serve one user at a time in order to improve the quality of the wireless link. Also, the direct

links between the users and the BS are available. Under these assumptions, all the other users act as

interferers for the intended user either through the direct links or through the links reﬂected by the RISs.

Each RIS operates as a nearly-passive surface, i.e., the RIS elements are passive but the RIS controller

1In this paper, the ON-OFF status is referred to having the entire RIS ON or OFF, i.e., either all the elements of the RIS are

turned ON or all the elements of the RIS are turned OFF.

4

may consume power [8]. In particular, the RIS controller is equipped with a CNN, as shown in Fig. 1(b).

The CNN is discussed in the next sections.

A. RIS-Assisted Communication Model

Each RIS, denoted as Rk,k∈ K, is equipped with Nkreﬂecting elements, which can be appropriately

conﬁgured by the RIS controller to reﬂect the signals of the users towards the BS. In general, the RISs can

be appropriately deployed so that line-of-sight (LoS) links can be established with the BS and, possibly,

the users. We assume that the channel state information (CSI) of all channels involved is perfectly known

at the BS2, which, in turn, can feed back the CSI to the RIS controller via a dedicated control channel [3],

[6].

Each RIS can be in two possible states: ON and OFF. We introduce a binary variable βk∈ {0,1}

to indicate the ON-OFF state of Rk.βk= 1 indicates that Rkis ON, which means that it reﬂects the

incident signals, while βk= 0 indicates Rkis OFF, which means that it does not reﬂect any signals.

As for the kth RIS that is ON, the amplitude reﬂection coefﬁcient is assumed to be equal to one for

all the Nkreﬂecting elements and the phase reﬂection matrix is

Φk= diag ejθk

1, ejθk

2, ..., ejθk

Nk,(1)

where Θk=θk

1, θk

2, ..., θk

Nkdenotes the vector of phase shifts that can be optimized by Rk.

B. Wireless Channel Model

We consider an RIS-aided uplink wireless system, where the channels from the desired user (denoted as

Ul) to Rk, from Rkto the BS, and from the mth interfering user (denoted as Um,m∈ L, m 6=l) to Rk, are

hl,k ∈CNk×1,gk∈C1×Nk, and hm,k ∈CNk×1, respectively. The channel gains of the direct links from

Ulto BS and from Umto BS are denoted by hd,l and hd,m, respectively. These channels are assumed to

be perfectly estimated and quasi-static, hence remaining nearly-constant during the transmission time [3].

Without loss of generality, we assume that the users are randomly distributed and have time-varying

trafﬁc demands. This implies that the users may not be all active during the considered transmission

time3. In particular, the total number of interfering users for Ulis given by ωl=Pm∈L,m6=lαm, where

αm∈ {0,1}is a binary variable that indicates that Umis active (αm= 1) or inactive (αm= 0) and

therefore can cause or not interference to Ul, respectively.

2Many papers in the literature have tackled the issue of estimating and reporting the CSI, e.g., [7], [9]. Therefore, this problem

is not addressed in this paper and it is left to a future research work.

3We assume that the active interfering users remain unchanged during the transmission time of the desired user.

5

Accordingly, the set of active interfering users is W={Um|αm= 1}, where ∀m6=land ∀m∈ L.

Since the number of active interferers is dynamic, it is not easy to estimate the interference distribution

over time. The received signal at the BS depends on the desired signal sent from Ul(including the direct

link and the link reﬂected by the RIS), the interference from the interfering users in W, and the white

Gaussian noise:

yl= hd,l +

K

X

k=1

βkgkΦkhl,k!√plsl

| {z }

Desired signal from Ul

+X

m∈W hd,m +

K

X

k=1

ξk

m,lβkgkΦkhm,k !√pmsm

| {z }

Interference from other users

+nl,(2)

where pland pmdenote the transmit power of Uland Um, respectively, sland smare the unit-power

information signals sent from Uland Um, respectively, and nl∼ CN(0, σ2)is the white Gaussian noise.

In addition, ξk

m,l ∈[0,1] accounts for the impact of the interference caused by Umto Uland that is

associated to the kth RIS, as clariﬁed in Assumption 1.

Assumption 1. We assume that the impact of the strength of the interference reﬂected by an RIS is

inversely proportional to the difference of the angles of incidence between the desired user and the

interfering users at the RIS (see λin Fig. 1(b) for an example).

C. Problem Formulation and Analysis

Let B={β1, β2, ..., βK}denote the RIS binary decision vector that collects the binary variables that

identify the ON-OFF status of the RISs. The set of RISs that are active is R={Rk|βk=1},∀k∈ K.

Accordingly, the SINR at the BS for the intended user Ulis

γl=

plhd,l +PK

k=1 βkgkΦkhl,k

2

Pωl

m=1 pmhd,m +PK

k=1 ξk

m,lβkgkΦkhm,k

2+σ2

.(3)

Our objective is to maximize the SINR in (3), by optimizing the RIS binary activation vector (B), and

the phase shifts matrix of the active RISs which is denoted as Θ={Θ1,Θ2, ..., ΘK}. To this end, we

need to solve the following optimization problem:

P1: max

B,Θγl(4a)

s.t. βk∈ {0,1},∀k∈ K,(4b)

ejθi

n= 1,∀n∈[1, Ni],∀i∈ R.(4c)

Constraint (4b) indicates that the kth RIS can only be ON (i.e., βk= 1) or OFF (i.e., βk= 0) at one

time. Constraint (4c) indicates that each RIS reﬂecting element can only provide a phase shift θi

n∈[0,2π)

without amplifying the signals.

6

USRP2

User-1

USRP2

BS

USRP2

...

Laptop with

GNU radio

Laptop with

GNU radio Laptop with

GNU radio

USRP2

Laptop with

GNU radio

Tx Antenna

Tx Antenna Tx Antenna

User-2 User-L

I/Q le

Rx Antenna

(a)

Conv+ReLU layer

FC layer

FC layer

Softmax

…

…

Feature Extraction

Output

Trained CNN Model

Incident

Signal

Conv+ReLU layer

Interfering

users set

RIS ON/OFF state

Distributed RIS

Binary Control

Incident Signal

Trained CNN

model #1

Trained CNN

model #2

Trained CNN

model #L

...

Interfering

devices set

Inference

Model

selection

RIS2

RIS1

BS

U

I1

I2

θ

1,1

θ

2,1

θ

2,2

θ

1,2

BS

RIS

λ

100m

60m

80m

40m

120o

U2

U1

(b)

Fig. 2: RF trace collection scenario is shown in (a), where Ltransceivers (users) are scheduled for transmission towards a

receiver (BS). The proposed ISL-aided RIS control structure is shown in (b), where the input to the CNN is the incident RF

signals and the output is the set of interfering users.

Our Proposal: We observe that P1is an MINLP, which is NP-hard and whose global optimal solution

is, in general, difﬁcult to obtain. In addition, traditional optimization methods may be computationally

intensive. An emerging approach to tackle this issue is to apply deep learning methods to solve P1at

a reduced computational complexity [10], [11]. P1, however, may not be easy to solve even using deep

learning methods, since the number of interfering users is a random variable that is unknown. In addition,

during the channel estimation phase, the RISs cannot estimate on their own the active interferers because

they only reﬂect the incident signals in a passive manner. This makes the solution of P1even more

difﬁcult. To address this challenge, we propose a distributed control mechanism that solves P1with the

aid of the ISL algorithm.

III. DISTRIBUTED RIS CO NT ROL VIA INTELLIGENT SPECT RUM LEARNING

In this section, the ISL-aided distributed RIS control mechanism is introduced. The proposed approach

leverages appropriately trained CNNs at the controller of the RISs, which can identify the active interfering

users in a distributed way.

A. Intelligent Spectrum Learning

ISL is a multi-class classiﬁcation algorithm that, based on a CNN, returns the set of interfering users

for each intended user Ul. To design the ISL algorithm, three main aspects need to be discussed: 1) RF

traces collection, 2) ofﬂine CNN training, and 3) online CNN inference.

1) RF Traces Collection: As far as the RF data collection phase is concerned, historical RF traces are

collected using a universal software radio peripheral (USRP2) testbed, which is wired connected (e.g.,

Gigabit Ethernet) to a host PC with an implementation of the GNU Radio, as illustrated in Fig. 2(a). In

particular, the users are emulated through a laptop that is mainly responsible for baseband processing while

a USRP2 platform is used for the up-conversion, the digital-to-analog (D/A) conversion, and wireless

transmission of the signals. As far as the BS is concerned, another USRP2 module ﬁrst receives the

7

signals from the radio interface and then performs A/D and down-conversion. Subsequently, the laptop

receives the signals from the USRP2 via the Ethernet and executes the baseband processing. Finally, the

Inphase (I) and Quadrature (Q) sequences are stored as a ﬁle. In particular, the experimental setup for RF

data collection using the USPR2 is performed by using signals at 2.4GHz carrier frequency with 1MHz

bandwidth. In order to collect realistic RF signals in the presence of interference, we let multiple USRP2

units transmit RF signals to an USRP2 receiver. The RF traces have been collected as I/Q sequences, by

including the wireless channel, for a wide range of SNR (e.g., from 0to 20 dB with interval of 5dB)

in order to account for different interfering cases [12].

2) CNN Ofﬂine Training: The acquired RF traces have been used for training the CNN architecture

illustrated in Fig. 2(b), where each convolutional layer is followed by a rectiﬁed linear units activation

function for feature extraction. Fully-connected (FC) layers are used to classify the signals by using

the softmax activation function for the output layer [13]. The training algorithms is based on the Adam

algorithm that uses the cross entropy as the loss function. The CNN model is trained ofﬂine using

TensorFlow on a GPU cluster (NVIDIA Tesla P100-PCIE-16GB).

Even though the training of the CNN does not account for all possible channel conditions, the

generalization property of deep learning enables the trained CNN to infer channel conditions not included

in the training dataset [14]. It is also noteworthy that several methods have been proposed to scale up the

training process of deep neural networks across GPU clusters, which helps to further reduce the runtime

of the ofﬂine training. Once the CNN model is appropriately trained, it can directly infer incident signals

in near real-time. In other words, the proposed ISL-based framework moves the complexity from online

computation to ofﬂine training.

3) CNN Online Inference: The CNN is, in particular, trained in order to return the set of active

interfering users based on different input signals at each RIS. Speciﬁcally, the received RF signals ﬁrst

undergo A/D conversion and frequency down-conversion. Then the baseband I/Q sequences are fed into

the trained CNN model to perform online inference at the RIS controller.

By performing feed-forward calculation via the CNN model (i.e., online inference), the intefering users

set for the desired user Ulis obtained as

e

Il=

{Um|eαm= 1},∀m∈ L,∀m6=l, If eωl≥1,

∅,If eωl= 0,

(5)

where eαmdenotes the inferred state ﬂag of Um, and eωlindicates the inferred total number of intefering

users in the incident signal.

8

4) An Illustrative ISL Example: To better understand the proposed ISL algorithm, we illustrate an

example with only two users (denoted as U1and U2), and each user has a binary state, i.e., ‘active’/‘ON’

and ‘inactive’/‘OFF’. In this case, there exist four combinations of signals from the perspective of each

RIS: (1) ‘Idle’ (indicating that both U1and U2are inactive), (2) ‘Only U1’ (indicating that only U1is

active), (3) ‘Only U2’ (indicating that only U2is active), and (4) ‘U1+U2’ (indicating that both U1and

U2are active). Based on the superimposed incident signal(s), the RIS needs to identify the composition

of the signal(s), i.e., to identify the correct class out of the four possible classes of signals. Therefore,

this signal identiﬁcation boils down to a four-class classiﬁcation problem, as illustrated in Table I.

TABLE I: An illustrative ISL example with two users

Inferred Class Description

Class-1: Idle The collected RF traces include only the noise

Class-2: Only U1The collected RF traces include only U1

Class-3: Only U2The collected RF traces include only U2

Class-4: U1+U2The collected RF traces include both U1and U2

B. Distributed RIS Binary Control

By feeding the inferred set of interfering users into the distributed RIS binary control algorithm, the

corresponding phase shifts and the binary ON-OFF status of the RISs can be obtained, as illustrated in

Fig. 2.

1) Optimal Phase Shifts Calculation: To calculate the phase shifts at the RIS, we denote the obtained

CSI associated to the kth RIS as Ck={hl,k,hm,k,gk, hd,l , hd,m}. Based on the inferred interfering users

(including e

Iland eωl) obtained via the trained CNN, the received SINR of the signal sent from Ulis

given by

eγl=

plhd,l +PK

k=1 βkgkΦkhl,k

2

Pe

ωl

m=1 pmhd,m +PK

k=1 ξk

m,lβkgkΦkhm,k

2+σ2

.(6)

Based on (6), we ﬁrst obtain the phase shifts of each RIS under the assumption βk= 1,∀k∈ K, and

then optimize the optimum binary ON-OFF vector Bbased on the obtained phase shifts. The ﬁrst step,

in particular, can be formulated as

P2: max

Θeγl(7a)

s.t. βk= 1,∀k∈ K,(7b)

ejθk

n= 1,∀n∈[1, Nk],∀k∈ K.(7c)

9

We observe that P2is a non-convex problem, which can be tackled by using several methods, such as the

semideﬁnite relaxation (SDR) method [3] and the successive convex approximation (SCA) method [15].

The optimal solution for the kth RIS is denoted by Θ∗

k, and the corresponding reﬂection-coefﬁcient

matrix is Φ∗

k. With the obtained reﬂection-coefﬁcient matrix, each RIS ON-OFF status is optimized via

the following RIS binary control algorithm.

2) Distributed RIS Binary Control Algorithm: Assuming that only the kth RIS is ON, i.e., βk= 1,

the received SINR of the signal sent from Ulto the BS via the kth RIS is

eγk

l=pl|hd,l +gkΦkhl,k|2

Pe

ωl

m=1 pmhd,m +ξk

m,lgkΦkhm,k

2+σ2

.(8)

If, on the other hand, all the KRISs are OFF, i.e., βk= 0 for k∈ K, the received SINR of the signal

sent from Ulto the BS via the direct link is

eγD

l=pl|hd,l|2

Pe

ωl

m=1 pm|hd,m|2+σ2.(9)

Based on (8) and (9), the kth RIS decides whether to be ON or OFF as detailed in Remark 1.

Remark 1. The kth RIS should be ON if eγk

l≥eγD

lholds. This indicates, in fact, that the desired signal

from Ulcan be enhanced via the kth RIS. Otherwise, the kth RIS should be OFF to avoid the degradation

of the desired signal. In this case, the incident angle between the signals of the desired user and the

interfering signals at the RIS is in general small.

The approach for solving P2is summarized in Algorithm 1, which is executed at each RIS controller

when the incident signal is received. Speciﬁcally, upon receiving the incident signal, each RIS controller

identiﬁes the set of interfering devices by extracting the I/Q samples from a copy of the incident signal and

feeding them into the trained CNN. Based on the classiﬁcation outcome of the CNN, the RIS controller

can set the RIS ON-OFF state in a distributed manner.

3) Computational Complexity: The total computational complexity includes the online inference via

the CNN and the iterative algorithm to solve the phase shift optimization problem P2and the RIS ON-OFF

optimization problem.

•Complexity for CNN online inference: The CNN is trained ofﬂine in a supervised fashion, therefore

the complexity of training can be ignored. The trained CNN model has a quadratic time complexity

during the inference process, i.e., O(M2CK), where Cdenotes the number of layers, Mdenotes

the number of neurons, and Kdenotes the total number of RISs.

•Complexity for solving the phase shift optimization problem P2: To solve the problem P2, the

complexity lies in computing the optimal phase shift at each iteration of the optimization method,

10

Algorithm 1: Distributed RIS Binary Control

Input: e

Il,eωl, and Ck;

Output: B;

1: Initialize k= 0, all the RISs are ON;

2: while k < K do

3: k←k+ 1;

4: Calculate Θ∗

kand Φ∗

kby solving problem P2;

5: Calculate eγk

land eγD

lvia (8) and (9), respectively;

6: if eγk

l≥eγD

lthen

7: Keep the kth RIS ON, i.e., βk= 1;

8: else if eγk

l<eγD

lthen

9: Turn the kth RIS OFF, i.e., βk= 0;

10: end if

11: Add the kth RIS binary decision βkto B.

12: end while

e.g., the SCA method [15] whose complexity is O(Qz), where Q=PK

k=1 Nkdenotes the total

number of elements of all the RISs, and zis the total number of the iterations required.

•Complexity for solving the RIS ON-OFF optimization problem: Since eγk

land eγD

lneed to be

calculated via (8) and (9), respectively, the computational complexity of solving the RIS ON-OFF

optimization problem required at each RIS controller is O(K).

As a result, the total complexity is O(M2CK +Qz +K), which grows linearly with the total number

of RISs.

IV. SIMULATION RES ULTS

In this section, we ﬁrst evaluate the inference accuracy of the trained CNN model and the computational

complexity of the proposed ISL-based DRBC algorithm. Then we validate the beneﬁts of deploying the

ISL-based DRBC algorithm.

A. CNN Testing Results

We trained the CNN with the 80% of collected RF data set which contains about 800 million I and Q

samples (training set), validated it by using 10% of the dataset (validation set), and tested it by using 10%

of the dataset (testing set) each corresponding to about 100 million of the I and Q samples. The trained

CNN model consists of two convolutional (Conv) layers with ReLU activation functions, followed by

11

TABLE II: Inference accuracy of the trained CNN model.

Scenarios w= 32 w= 128 w= 512

Idle 100.00%100.00%100.00%

Only U198.35%98.04%96.21%

Only U296.09%96.12%95.64%

U1+U299.66%99.76%99.93%

two dense fully connected (FC) layers. In particular, the trained CNN model contain 256 ﬁlters (1×3) in

the ﬁrst Conv layer, 128 ﬁlters (1×3) in the second Conv layer, 256 neurons in the ﬁrst FC layer, and 9

neurons in the second FC layer (output).

The classiﬁcation accuracy of the trained CNN is analyzed in Table II, by considering a two-user

scenario. The window size (i.e., the number of time steps of the collected RF data) is 32,128, and 512,

respectively. We observe from Table II that the online inference accuracy is, in general, greater than 95%

in the considered scenario. Compared to other classes, the ‘Idle’ class has the main characteristic that

no user transmits and only background noise exists. Due to the distinguishable pattern compared to the

other three classes, the CNN model predicts the ‘Idle’ class perfectly.

B. Computation Time

The proposed ISL-based DRBC algorithm allows us to obtain the optimal ON-OFF status of each RIS.

As detailed in previous text, the optimal ON-OFF status of the RISs can be formulated as the solution

of a non-convex MINLP, which is usually challenging to solve [15]. In this section, we compare the

proposed ISL-based DRBC algorithm against the spatial branch and bound (sBB) method, which is often

employed to solve non-convex MINLP [16], in terms of computation time.

The comparison of the average computation time (deﬁned as t=Total time consumption

Total number of computations ) between

the proposed ISL-based DRBC algorithm and the traditional sBB method is conducted on the same

hardware platform that consists of an Intel Xeon(R) CPU E5-2650@2.0 GHz x 16. The obtained results

are illustrated in Table III. Compared to the traditional sBB method, the ISL-based DRBC algorithm

results in much lower computation time while still yielding the optimal ON-OFF status for each RIS.

The computation time of the proposed algorithm is less than one-thousandth of the computation time of

the sBB method when the number of RISs varies from 2to 5.

C. Performance Evaluation

1) Simulation Setting: The simulation model consists of KRISs, one desired user (U1), and one

interfering user (U2). Each RIS consists of 256 elements and all the KRISs are equally spaced by 5m

in vertical direction. The distances from BS and U1to the RIS center are 80 m and 60 m, respectively.

12

TABLE III: Computation time (ms) of the proposed ISL-based DRBC algorithm and the traditional sBB method

K

tTraditional sBB method Proposed DRBC algorithm

214.1 2.60 ×10−3

314.2 2.55 ×10−3

414.5 2.40 ×10−3

515.2 5.15 ×10−3

The incident angle between the BS and U1at the RIS is 150o, and the incident angle between U2and

U1is λ∈[0,150o]. We assume that ξk

m,l is linearly inversely proportional to λ. The channel parameters

are selected according to the 3GPP Urban Micro standard [17], which describes the path loss for both

line-of-sight and non-line-of-sight components [18]. The transmission power of U1is 20 dBm, the noise

power σ2is -94 dBm, the carrier frequency is 3GHz, and the reﬂection amplitude is equal to one.

2) Simulation Results: We evaluate the performance of the proposed ISL-aided algorithm by comparing

it with two benchmarks: ‘RIS always ON’ and ‘RIS always OFF’. Figs. 3(a)-(b) depict the achievable

SINR versus the angle of incidence (λ) for K= 1 and pm= 10,15,20 dBm. We observe that the SINR

ﬁrst gradually decreases as λincreases due to the reduction of the distance between U2and BS, and

then increases due to the perfect interference elimination at the RIS. When λis small in particular, the

RIS is prone to be OFF since the interference reﬂected via the RIS is more pronounced. If λis large, on

the other hand, the impact of the interference is reduced and it is more probable that the RIS is ON. In

general terms, however, the impact of λon the system performance is still an open issue, whose analysis

is postponed to a future research work.

In Figs. 4(a)-(c), the SINR versus Kis illustrated, where pm= 10 dBm, the distance between U2and

the RIS is 5m. We observe that the RISs are always OFF if λ= 0, since the impact of the interference

is too high. If λis very large, e.g., λ= 150oin Fig. 4(b), the impact of the interference is low and

the RISs are always ON. When λis randomly selected, e.g., λ∈[30o,120o]in Fig. 4(c), the SINR

obtained by the ISL algorithm increases with Kand outperforms the two benchmarks, by about 100%

with respect to ‘RIS always OFF’ and by nearly 300% with respect to ‘RIS always ON’ when K= 5.

From Figs 4(a)-(c), we conclude that the performance of the proposed ISL algorithm largely depends on

λ, which impacts the interference cancellation at the RISs.

13

0 30 60 90 120 150

0

0.5

1

1.5 RIS always ON

RIS always OFF

Proposed RIS control

Pm=10, 15, 20 dBm

(a)

0 30 60 90 120 150

0

1

2

3

4

5

RIS always ON

RIS always OFF

Proposed RIS control

Pm=10, 15, 20 dBm

(b)

Fig. 3: Achievable SINR vs. λfor K= 1. The distance between the interfering user (U2) and the RIS is 10 m in (a), and 5m

in (b), respectively.

12345

0

0.2

0.4

0.6

0.8

1

1.2

RIS always ON

RIS always OFF

Proposed RIS control

(a)

12345

0

1

2

3

4

5

RIS always ON

RIS always OFF

Proposed RIS control

(b)

12345

0

0.5

1

1.5

2RIS always ON

RIS always OFF

Proposed RIS control

(c)

Fig. 4: SINR vs. Kfor pm= 10 dBm. λ= 0 is shown in (a), λ= 150ois shown in (b), and λrandomly selected in [30o,120o]

is shown in (c).

V. CONCLUSION AND FUTURE WO RK

In this paper, we introduced an ISL algorithm that uses appropriately trained CNNs for the interference

management in RIS-aided multi-user uplink networks. With the aid of the ISL algorithm, the RISs

are capable of inferring the interfering signals directly from the incident signals. A distributed control

algorithm was proposed to maximize the received SINR by dynamically conﬁguring the binary status

of the RIS elements. Simulation results validated the performance improvement offered by the proposed

ISL-aided RIS approach. Ofﬂine training is only a candidate way to train a CNN, which may need to

be retrained when RF data distribution changes signiﬁcantly. This issue may be avoided by using online

training methods, such as federated learning, which is a promising method for application in dynamical

wireless environments.

14

REFERENCES

[1] M. Di Renzo et al. “Smart radio environments empowered by reconﬁgurable AI meta-surfaces: An idea whose time has

come,” EURASIP Journal on Wireless Communications and Networking, vol. 2019, no. 129, pp. 1-20, May 2019.

[2] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconﬁgurable intelligent surfaces for energy

efﬁciency in wireless communication,” IEEE Transactions on Wireless Communications, vol. 18, no. 8, pp. 4157-4170,

Aug. 2019.

[3] Q. Wu, and R. Zhang, “Intelligent reﬂecting surface enhanced wireless network via joint active and passive beamforming,”

IEEE Transactions on Wireless Communications, vol. 18, no. 11, pp. 5394-409, Aug. 2019.

[4] J. M. Kang, “Intelligent reﬂecting surface: joint optimal training sequence and refection pattern,” IEEE Communications

Letters, vol. 24, no. 8, pp. 1784-1788, Aug. 2020.

[5] H. Yang, et al. “Intelligent reﬂecting surface assisted anti-jamming communications based on reinforcement learning,” In

Proc. of IEEE Global Communications Conference (GLOBECOM), Taipei, Dec. 2020.

[6] T. Bai, C. Pan, Y. Deng, M. Elkashlan, A. Nallanathan and L. Hanzo, “Latency minimization for intelligent reﬂecting surface

aided mobile edge computing,” IEEE Journal on Selected Areas in Communications, vol. 38, no. 11, pp. 2666-2682, Nov.

2020.

[7] M. Di Renzo, et al. “Smart radio environments empowered by reconﬁgurable intelligent surfaces: how it works, state of

research, and the road ahead,” IEEE Journal on Selected Areas in Communications, vol. 38, no. 11, pp. 2450-2525, Nov.

2020.

[8] C. Huang et al., “Holographic MIMO surfaces for 6G wireless networks: opportunities, challenges, and trends,” IEEE

Wireless Communications, vol. 27, no. 5, pp. 118-125, Oct. 2020.

[9] A. Zappone, M. Di Renzo, F. Shams, X. Qian and M. Debbah, “Overhead-aware design of reconﬁgurable intelligent

surfaces in smart radio environments,” IEEE Transactions on Wireless Communications, vol. 20, no. 1, pp. 126-141, Jan.

2021.

[10] B. Yang, X. Cao, J. Bassey, X. Li, and L. Qian, “Computation ofﬂoading in multi-access edge computing: a multi-task

learning approach,” IEEE Transactions on Mobile Computing, pp. 1-1, Apr. 2020.

[11] C. Huang, R. Mo and C. Yuen, “Reconﬁgurable intelligent surface assisted multiuser miso systems exploiting deep

reinforcement learning,” IEEE Journal on Selected Areas in Communications, vol. 38, no. 8, pp. 1839-1850, Aug. 2020.

[12] B. Yang, X. Cao, Z. Han, and L. Qian, “A machine learning enabled MAC framework for heterogeneous Internet-of-Things

networks,” IEEE Transactions on Wireless Communications, vol. 18, no. 7, pp. 3697-712, May 2019.

[13] A. Zappone, M. Di Renzo, and M. Debbah, “Wireless networks design in the era of deep learning: Model-based, AI-based,

or both?” IEEE Transactions on Communications, vol. 67, no. 10, pp.7331-7376, Oct. 2019.

[14] Y. S. Abu-Mostafa, M. Magdon-Ismail, and H. T. Lin, “Learning from data,” New York, NY, USA:: AMLBook, vol. 4,

Mar. 2012.

[15] Z. Yang, et al., “Energy-efﬁcient wireless communications with distributed reconﬁgurable intelligent surfaces,”

arXiv:2005.00269, [Online]. Available: https://arxiv.org/abs/2005.00269.

[16] P. Belotti, J. Lee, L. Liberti, F. Margot, and A. Wchter, “Branching and bounds tightening techniques for non-convex

MINLP,” Optimization Methods &Software, vol. 24, no. 4-5, pp. 597-634, Aug. 2009.

[17] “Further advancements for E-UTRA physical layer aspects (Release 9),” 3GPP Standard TS 36.814, Mar. 2010.

[18] H. Zhang, B. Di, L. Song, and Z. Han, “Reconﬁgurable intelligent surfaces assisted communications with limited phase

shifts: How many phase shifts are enough?” IEEE Transactions on Vehicular Technology, vol. 69, no. 4, pp. 4498-4502,

Apr. 2020.