You are on page 1of 11

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
1

Deep Learning for An Effective Non-Orthogonal


Multiple Access Scheme
Guan Gui, Senior Member, IEEE, Hongji Huang, Student Member, IEEE, Yiwei Song,
and Hikmet Sari, Fellow, IEEE

Abstract—Non-orthogonal multiple access (NOMA) has been an essential enabling technology, allowing next generation
considered as an essential multiple access technique for enhancing wireless networks to meet the heterogeneous demands of
system capacity and spectral efficiency in future communication low latency, high reliability, massive connectivity, improved
scenarios. However, the existing NOMA systems have a fun-
damental limit: high computational complexity and a sharply fairness, and high throughput [2]. Some fundamental work on
changing wireless channel make exploiting the characteristics the rate region, bit error ratio (BER) performance, and resource
of the channel and deriving the ideal allocation methods very allocation schemes based on the NOMA principles have been
difficult tasks. To break this fundamental limit, in this paper, carried out [3]–[5] to improve the reliability and the efficiency
we propose a novel and effective deep learning (DL)-aided of NOMA systems to a large extent.
NOMA system, in which several NOMA users with random
deployment are served by one base station (BS). Since DL is Most recently, many novel applications have been proposed
advantageous in that it allows training the input signals and based on the NOMA scheme. In [6], the performance of
detecting sharply changing channel conditions, we exploit it a downlink NOMA scheme with randomly deployed users
to address wireless NOMA channels in an end-to-end manner. was well investigated. A NOMA unicast-multicast system,
Specifically, it is employed in the proposed NOMA system to which can achieve diverse orders equal to the number of
learn a completely unknown channel environment. A long short-
term memory (LSTM) network based on DL is incorporated multicast users, was developed in [7]. Additionally, a fair
into a typical NOMA system, enabling the proposed scheme to NOMA approach was proposed and employed in the case of
detect the channel characteristics automatically. In the proposed pairing a near base-station user and a cell-edge user, where
strategy, the LSTM is first trained by simulated data under the ergodic capacity can always be improved for each user
different channel conditions via offline learning, and then the [8]. For the purpose of optimizing the spectrum and energy
corresponding output data can be obtained based on the current
input data used during the online learning process. In general, efficiency, a new spectrum and energy-efficiency mmWave
we build, train and test the proposed cooperative framework to transmission scheme that integrates the concept of NOMA
realize automatic encoding, decoding and channel detection in an into beamspace multiple-input multiple-output (MIMO) was
additive white Gaussian noise (AWGN) channel. Furthermore, proposed in [9]. Additionally, the authors in [10] derived
we regard one conventional user activity and data detection closed-form expressions for both exact and asymptotic outage
scheme as an unknown nonlinear mapping operation and use
LSTM to approximate it to evaluate the data detection capacity probabilities as well as delay-limited throughput based on their
of DL based on NOMA. Simulation results demonstrate that proposed NOMA cooperative system.
the proposed scheme is robust and efficient compared with Practically, channel state information (CSI) has great impact
conventional approaches. In addition, the accuracy of the LSTM- on the performance of the NOMA system, and a lot of works
aided NOMA scheme is studied by introducing the well-known have been devoted to realize channel estimation based on the
10-fold cross-validation procedure.
NOMA scenarios. In [11], in order to detect the CSI, a novel
Index Terms—Non-orthogonal multiple access (NOMA), long channel estimation scheme was proposed by designing new
short-term memory (LSTM), deep learning. type of linear estimator that aims at maximizing the average
effective signal-to-interference and-noise ratio (SINR) of the
I. I NTRODUCTION strong user with bounded SINR guaranteed for the weak user.
Meanwhile, due to the CSI is hard to be obtained through tra-
Non-orthogonal multiple access (NOMA) is an effective ditional methods, some scholars turn to investigate the NOMA
solution for enhancing spectral efficiency and has drawn sig- based strategies in different CSI scenarios. The performance
nificant interest from researchers because of its very promis- of NOMA under a non-ideal setup and two power allocation
ing application in fifth generation (5G) networks [1]. It is algorithms were provided under imperfect CSI [12]. Also,
based on the uplink NOMA systems, the authors revealed
This work was partially supported by the National Natural Science Foun-
dation of China (61571240), Jiangsu Specially Appointed Professor Grant that imperfect CSI leads to not only extra interference on
(RK002STP16001), Innovation and Entrepreneurship of Jiangsu High-level the desired signal but also the incorrect decoding order [13].
Talent Grant (CZ0010617002), and ”1311 Talent Plan” of Nanjing University Hence, how to derive an efficient way to acquire perfect CSI
of Posts and Telecommunications.
G. Gui, H. Huang, Y. Song, and H. Sari are with Nanjing University is an important issue in the field of NOMA based means and
of Posts and Telecommunications, Nanjing 210003, P. R. China. (e-mail: new methods should be exploited to resolve this problem.
guiguan@njupt.edu.cn, b14111829@njupt.edu.cn, b15080234@njupt.edu.cn, Although many recent papers have derived differen-
hikmet@njupt.edu.cn).
Corresponding author: Guan Gui t schemes for sum data rate and reliability optimization, these
Manuscript received XXX, XX, 2018. methods require high computational complexity due to the

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
2

nonlinear optimization and fail to derive a corresponding pow-


er allocation scheme based on given CSI. Generally speaking,
all the potential gains of NOMA systems depend largely on
CSI; thus, existing works have proposed many methods for
further improving the performance of channel estimation [14].
However, since the channel characteristics in multiple-user
systems are very complicated [15], conventional methods are
Fig. 1. LSTM-aided NOMA communication framework.
incapable of capturing the change in the channel condition
in real time. In other words, CSI acquirement is disturbed
and the NOMA system performance is degraded because of uplink analysis in NOMA networks.
the sharply varying channel characteristics. Meanwhile, as 3) We provide a performance analysis of the proposed
the channel sparsity patterns are often assumed unknown in DL-based methods in the NOMA system for different
the existing works, nonlinear reconstruction procedures are cases. Specifically, the sum data rate and block error rate
ineluctable. Thus, conventional techniques are not efficient and (BLER) are well investigated. Also, extensive simulation
reliable enough to achieve super-resolution direct of arrivals results and a comparison demonstrate the efficiency and
(DOA) estimation and signal detection. In recent decades, robustness of the proposed schemes.
the promising technique of machine learning (ML) has been
incorporated into the NOMA system to realize auto-detection The remainder of this paper is organized as follows. In
of the CSI. The deep learning (DL) [16] concept (a typical Section II, we construct a typical NOMA system with multiple
branch of ML), which was proposed in 2006, is a very users. Then, a scheme that incorporates LSTM into the NOMA
powerful tool for handling big data and solving complex system is proposed in Section III, in which different activation
nonlinear problems. Some previous works incorporated DL functions are employed in different layers to carry out specific
into communication with regard to channel coding, MIMO, mapping relationships. The numerical results used to evaluate
and the physical layer [17]–[19]. In [20], a DL-aided Sparse the performance of the proposed schemes are provided in
code multiple access (SCMA) where the codebook that mini- Section IV, which is followed by conclusions in Section V.
mizes the BER is adaptively developed, which requires smaller Notations: Vectors are denoted by boldface small letters,
computation time than conventional methods. The interesting while matrices are denoted by boldface capital letters. Su-
system that integrates the DL into the orthogonal frequency- perscripts (·)? , (·)T , (·)H and k · k represent the conjugate,
division multiplexing (OFDM) scenario has been exhibited in transpose, Hermitian and Frobenius norm operators, respec-
and Cm×n denote the Kronecker product and
L
[21] and its excellent performance has been corroborated in tively. Also,
terms of channel estimation and signal detection. Furthermore, the vector space of all m × n complex matrices, respectively.
DL has been adopted in traffic control systems, the encourag- Furthermore, E[·] is the expectation operator. Additionally,
ing performance of which has been demonstrated [22]–[25]. diag(x) represents a diagonal matrix with a main diagonal
In addition, some advantages of the DL-based communication of x.
schemes have been exhibited.
Motivated by the aforementioned considerations, a compre- II. S YSTEM M ODEL
hensive study is conducted to optimize the NOMA system
based on the DL technique in terms of effectiveness and reli- We consider a communication system with a single source
ability. The main contributions of this paper are summarized (i.e., base station (BS)) and n users, where all the users
as follows. are deployed at different distances from the BS in the same
frequency band. To be specific, all the users are NOMA users,
1) To the best of our knowledge, by establishing a NOMA which are denoted as U1 , U2 , · · ·, Ui , · · ·, Un . In this scenario,
system with the aid of DL, we first attempt to use deep we assume that a superposed signal transmitted by the BS,
learning methods instead of traditional online learning to and the distance from the BS to Ui is defined as di . Note that
analyze the complex channel characteristics of NOMA. increasing the distance from the BS to the user will deteriorate
To promote the performance of the DL framework, the the channel condition (i.e., consider a NOMA-far user U1 and
pre-training process of the DL network is modified. a NOMA-near user U2 ; the channel condition of U2 is superior
2) In our work, taking advantage of the long short-term to that of U1 ). Based on the principle of NOMA, a user with
memory (LSTM) (i.e., a typical branch of deep neu- a better channel condition will be allocated lower transmitted
ral network (DNN)) and NOMA, a framework that power. Additionally, successive interference cancellation (SIC)
integrates the LSTM into NOMA is provided. Here, can be employed at some of the users with good channel
the proposed scheme can capture the system statistics conditions, while the signals of the users with bad channel
automatically. Meanwhile, to enhance the performance conditions are first decoded and subtracted, and then the
of a novel user activity and data detection approach in desired information is decoded for the good-channel-condition
terms of reliability and complexity, we use the LSTM users. In particular, in our proposed scenario, the mathematical
network to address this nonconvex problem based on model of the wireless communication channel is completely
the uplink NOMA system. Thus, it is verified that DL unknown, and the distance from the BS to the user is random.
can be introduced into data detection optimization and Assuming that the wireless channel is corrupted by additive

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
3

white Gaussian noise (AWGN), the signal received at Ui can In other words, the RBM is used as the pre-training structure
be formulated as for promoting the performance of the LSTM. Generally speak-
n
X ing, messages are converted into a sequence of transmitted
yi = hi xi + hj xj + ni , (1) symbols that represent different transmitted signals in digital
j6=i communication systems. Based on our analysis above, let the
where hi is denoted as the unknown channel coefficient from unknown signal vector xi = {si,1 , si,2 , ..., si,q } be a finite
the BS to Ui . Also, xi can be denoted as the unknown sequence of all transmission symbols, where q denotes the
transmitted signal at Ui , and ni is assumed as the AWGN with amount of samples in the sequence at Ui . Meanwhile, the
zero mean and variance σi2 . It is noted that the transmitted symbols will be corrupted by complex noise and
PBSn
transmits
degraded by the polytropic environment, and the correspond-
signals to n users synchronously, and the item j6=i hj xj and
the AWGN ni will be subtracted in subsequent operations. ing output yi can be detected at the receiver. Specifically, this
Via an algebraic transformation, Eq. (1) can be expressed relation can be represented by a channel model, which can
as be illustrated by the probability density function (PDF) as
p(yi |xi ). In this procedure, each RBM independently trains
yi = xhT + ni . (2) the framework based on its local unknown input signal by a
stochastic gradient descent (SGD) algorithm. Thus, for the first
Here, h, and x are regarded as the channel vector, and the
sample, we can work out the training data set as follows:
signal vectors space, respectively, which can be expressed as  1 1
), (s2i,1 , yi,1
2
), ..., (sK K

h = [h1 , h2 , · · ·, hn ], and x = [x1 , x2 , · · ·, xn ], respectively. (si,1 , yi,1 i,1 , yi,1 ) , (5)
To be specific, assuming the original transmitted signal as zi
at Ui , x can be further given by where K is the number of iterations. This data set is applied to
pre-train the DL scheme and then employed to train the LSTM
x = zi p, (3) network (the data generation procedure is depicted below).
K
√ √ √ Also, {(sti , yit )}t=1 is formed as a (label, feature) pair of
where p = [ p1 , p2 , · · ·, pn ] is the power vector. After-
the DL network. Since the NOMA based network is very
wards, applying the power allocation coefficients at the BS,
complex, we use the well-known Pearson Correlation mean
the constraints of the transmitted power are expressed as
 n [27] to extract the features automatically.
X √ LSTM is a special type of RNN architecture that is capable
k pi k2 = p2


of remembering values over arbitrary intervals [28]. During



i=1
n (4) the learning proceedings, stored values are not modified. In
addition, one typical advantage of an RNN is that forward and
 X
2
kαi k = 1,



backward connections can be realized between neurons. The

i=1
RBM-based pre-training method boosts LSTM performance,
where αi is the power allocation coefficient of Ui . Concretely,
especially when the training set is not sufficiently large.
in Eq. (4), pi = αi kpk2 , indicating the power served for Ui
Each RBM has the same number of hidden units in the
in the NOMA system.
framework. In the pre-training process, the weights of the
LSTM are initialized to a proper range (a detailed illustration
III. LSTM-A IDED NOMA S CHEME of the process is given in [26]), which contributes to better
In this section, we consider a novel scheme that adopts optimization and regularization.
the LSTM network into the NOMA framework. Among all
the frameworks based on DL, the recurrent neural network B. LSTM scheme based on the NOMA system
(RNN) has the advantage of realizing internal memory to
process arbitrary sequences of inputs, which is very appro- In this subsection, we propose a novel framework that
priate for high-data-rate sequence analysis. Lately, to solve integrates the improved LSTM network into the NOMA model,
the vanishing gradient problem and long-term dependencies as shown in Fig. 1. This LSTM network with different layers is
existing in an RNN, LSTM has been developed, and it is very used for complex data processing. For the purpose of keeping
superior in terms of signal processing, e.g., the promising the training set small, we extend the LSTM with additional
natural languages processing (NLP). Thus, after developing layers (i.e., hidden layers) to promote the representation and
a strategy based on restricted Boltzmann machines (RBM) learning capacities of the network. It is indicated that these
[26] to train the initial input, LSTM is employed to learn layers have no trainable parameters, which perform a certain
the channel characteristics of NOMA systems, and its end-to- action, such as mixing distortion, and propagate only one
end performance is simulated via offline training and online symbol to the node of the next layer with the same form.
training. Specifically, this LSTM network consists of 8 layers, in which
6 layers are hidden layers that are employed to carry out
training and recognition. Each hidden layer of this branch of
A. Pre-training based on the RBM the DNN is equipped with multiple neurons, and an output is
For the sake of enhancing the generalization ability of the the weighted sum of these neurons with a nonlinear function.
network and reducing the dimension of the input data, the The input layer is processed by the RBM network, while
RBM is implemented to train the original input of the LSTM. the output layer is processed by the sigmoid function. Also,

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
4

all the hidden layers are processed by the Rectified Linear as


Unit (ReLU) function. Here, in our LSTM framework, we n
X
denote the length of each training sequence as L, which is the yi = hi fi (si ) + ni , (6)
dimension of the input layer. This layer is a fully connected i=1
layer that includes 128 neurons, serving as the input of the where fi (·) is the mapping function of the transmitted symbol
transmitted signal vectors and conveying important features to si based on the DL framework.
the first hidden layer. The second hidden layer and the third Thereafter, in the DL-based multiple-user NOMA system,
hidden layer are dense layers that consist of 500 neurons and since a bias is corrupted optionally and the weighted initial
256 neurons, respectively. With these two hidden layers, the input is added to a neuron of the network, we often introduce
encoding procedure is conducted based on the output of the appropriate activation functions to propagate the results. As
input layer. Then, the next layer is designed as a noise layer indicated above, assuming z as the independent variable of
with 200 neurons, and the processed signal is corrupted by the the function, two typical activation functions are introduced
AWGN in this layer. Afterwards, the remaining two hidden in this paper:
layers comprise 128 neurons and 64 neurons, which act as
a decoder. Additionally, the output layer is the linear layer, 1
fsigmoid (z) = , (7)
providing the estimated output signal vectors based on the 1 + e−z
NOMA system.

In the channel estimation field, it is very difficult to collect fReLU (z) = max {0, z} . (8)
enough labeled training data for the field of ML, but using DL Then, noting that M and N are defined as the number
for channel coding is special because man-made signals are of input layers and number of hidden layers, respectively,
what we address. In past decades, scholars have developed all RN N
j−1 → Rj represents the mapping carried out by the jth
kinds of channel models that well describe the real channels layer. Hence, the form of the output oj of the jth layer can
in terms of channel characteristics, which make it possible be expressed as
for us to develop a simulator that can simulate different
channel statistics. Note that the proposed scheme is affected oj = fsigmoid (Wj oj−1 + bj ), (9)
by unknown signals. The training data, which map almost all
where Wj and bj denote the weight parameter and bias
the cases, can be introduced in the simulation because the
parameter, respectively. With the aid of the activation function
transmitted signals are obviously known only if noisy code-
fsigmoid (·), a nonlinearity transform can be realized. This
words are generated. In each simulation, we generate a random
procedure is essential to maximize the expressive power of
transmitted data sequence, and the corresponding NOMA
LSTM. The goal of our model is to minimize the difference
frame is constituted with pilot symbols; the channel models
between the output of the LSTM network and the transmitted
for multiple cases are then utilized to generate the CSI. At the
symbol; thus, the loss function can be expressed as
signal receiver (i.e., the final output of the LSTM network), the
NOMA signal is obtained based on the NOMA frames being n
1X
distorted by the current channel, particularly by the AWGN loss(x) = l(x, x
b(h, p, yi )). (10)
n i=1
and other unknown distortions. The transmitted signals and
the output signals are formed into a set, which is called the Here, according to the maximum likelihood estimator (MLE)
training data set. In particular, for the sake of realizing the theorem, a typical method can be formulated as
power allocation policy automatically and appropriately, the
samples at different users are allocated in correspondence with b(h, p, yi ) = arg min kyi − hT pxT k2 .
x (11)
x,h,p
the power allocation factor α according to the principles of the
classical ε - greedy algorithm, contributing to obtaining the However, this rule can achieve high accuracy in terms of es-
best power allocation policy with the aid of the training stage. timation at the cost of computational complexity O(2n ), which
The above procedures constitute the offline training process. induces unintended computational complexity. Therefore, we
In parallel, when we deploy this communication system, the modify Eq. (11) according to the cross entropy theory, and a
system will employ the online learning method to process the new estimation method can be formulated as
input data according to the current input data and the simulated l = xi log(xbi ) + (1 − xi )log(1 − xbi ), (12)
channel statistics generated by the offline learning, as shown
in Fig. 2. In the following, we explain the technical details of where xbi is the prediction. In this case, the computational
the LSTM-aided NOMA system. complexity reduces to O(logn).
Then, we focus on which new message in the cell is worthy
Consider the NOMA system. The channel conditions from of saving (i.e., the memory of the cell); this procedure is as
different users to the BS are varied, which is associated with follows.
the power allocation scheme illustrated in Eq. (4). Thanks to Step 1) Since the RBM is deployed as the input layer, the
the DL theory, we observe that each power allocation and output signals of the RBM are processed and their character-
signal detection policy of the NOMA can be simplified as a istics are captured. Then, the data are trained in the hidden
mapping of the LSTM network, and Eq. (1) can be rewritten layers with the aid of the SGD algorithm.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
5

Fig. 2. RBM-aided LSTM network.

jth hidden layer is hej and the corresponding condition is C


fj ,
hj the updating process can be given by
hej = fsigmoid (Wj · [e
hj−1 , xi ] + bi ), (14)

fj = tanh(Wj · [e
C hj−1 , xi ] + bi ). (15)
tanh During this procedure, the condition of the cell is updated
from Cj−1 to Cj . For a practical NOMA system, if the data
rate is lower than the channel capacity, this signal vector
should be eliminated. However, when the channel condition is
suddenly disturbed, the channel capacity will be degraded, and
its relation with the data rate is very likely to change. Hence,
some messages should be discarded, while other messages will
likely be added in the process of updating. This procedure can
be formulated as
Fig. 3. Typical structure of a cell of the LSTM.
Cj = oj × Cj−1 + hej × C
fj . (16)

Step 2) In the hidden layer, a tanh(·) function is always Specifically, by examining Eq. (16), we can observe that
employed to create new selective vectors Cj , which are the LSTM scheme is capable of avoiding the case in which the
conditions that can be added in the LSTM algorithm and can state suddenly changes to 0 via its form of continued addition.
be obtained as By contrast, considering that of the initial RNN method, Eq.
(16) should be rewritten as
1 − e−2x
tanh(x) = . (13) Cj = tanh(xi ) × U + Wj × Cj−1 , (17)
1 + e−2x
Afterwards, we integrate Step 1) into Step 2) to update the where U represents the updating coefficient of the RNN. This
condition. As shown in Fig. 3, assuming that the output of the procedure is based on the continuous multiply operation, and

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
6

Algorithm 1 LSTM-aided SISD training algorithm.


Input: x[1] , x[2] , · · ·, x[J] , representing the structured sparsity
in J continuous time slots, environment simulator.
Output: LSTM network.
1: Initialize the model. The number of iterations is initially
k = 0, and the support set I = ∅. The weight w = 0.
Also, initialize the error threshold as τ = 10−7 .
2: Generate a set of data sequences from the training data
Fig. 4. The proposed LSTM-aided SISD algorithm. set of each user.
3: Sample a mini-batch of data based on step 2.
4: Start the environment simulator to generate the wireless
the state may change to 0 if only one signal vector approaches channel, and mix specific man-made noise or distortion
0 during the updating process. Thus, the LSTM can outperform into the channel.
the initial RNN based on the NOMA model. 5: Compute the complementary set A(k) of the support I (k) .
After all the iterations, we obtain the output of the last layer while Card(I (k) ) < n − S, do
of the LSTM network, which can be given by
A(k) ← (I (k) )c := {1, 2, · · ·, n} \ I (k) . (20)
(M +N −1) (M +N −2) 1
o = f (v, W ) = f (f (· · ·f (si ))), (18) 6: Denote the loss function according to the mean squared
error (MSE) principle as
where o denotes the output signals of the last layer of the
network. J
X
l2 = kHs[j] x[j] − y[j] k22 . (21)
j=1
C. Approximation of SISD Algorithm via LSTM network 7: Calculate the new support by
To solve the multiple-user detection problem of uplink (k)
I (k+1) ← {i : kxav,i k > τ k }, (22)
NOMA systems for 5G, in which user activity must be
perceived at the BS, methods for detecting user activity and (k)
where xav,i denotes the absolute value of xi .
(k)
data during the same time slot (a typical advantage in 5G) 8: Train the LSTM network with the aid of the SGD algo-
must be proposed. In [33], to utilize the structured sparsity rithm.
of the NOMA system, a novel structured iterative support 9: Update the weight w and the output of each layer of the
detection (SISD) algorithm was proposed to jointly detect LSTM.
the transmitted signal and user activity in some continuous 10: end while
time slots, as shown in Fig. 4. By adopting the LSTM, 11: return LSTM network.
the computational complexity and the performance can be
optimized. Initially, based on the NOMA system developed
above, the signal vector received in J continuous time slots is
given as

y[j] = H[j]
s x
[j]
+ n[j] , 1 < j < J. (19) Algorithm 2 LSTM-aided SISD testing algorithm.
Input: LSTM network, environment simulator.
Here, y[j] and H[j] are defined as the received signal and the
Output: Reconstructed sparse signals: x̂[1] , x̂[2] , · · ·, x̂[J] .
equivalent channel matrix in the j-th time slot, respectively.
1: Load the LSTM network.
Also, n[j] represents the AWGN in the j-th time slot. Specifi-
[j] 2: Start the environment simulator to generate the wireless
cally, Hs ∈ Cn×S is the channel matrix, where S represents
channel, and mix specific man-made noise or distortion
the number of orthogonal frequency division multiple (OFDM)
into the channel.
subcarriers. It can be demonstrated that each iteration of the
3: Process the LSTM network.
SISD represents a continuous mapping transformation in the
4: Update the output of the LSTM network.
LSTM. Hence, we merge the LSTM into the SISD scheme,
5: Transform J estimated signal vectors into absolute values:
which is depicted in Algorithm 1 and Algorithm 2.
J
Then, to evaluate the performance of the DL-based SISD X
scheme in terms of data detection, we introduce the well- x[j]
av = kx[j] k, (23)
j=1
known mean squared error (MSE) principle, which can be
given as [j] [j]
where kx[j] k = [kx1 k, kx2 k, · · ·, kxn k]T .
[j]

J 6: return x̂[j] = x[j] , j = 1, 2, · · ·, J.


1 X kx̂[j] − x̂[j] k2
MSE = . (24)
J j=1 kx[j] k2

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
7

D. Encoding scheme for NOMA communication system where j is an even integer and xji,E represents a DL decoder.
Apparently, the proposed LSTM-aided NOMA communi- Similarly, for odd numbers, Eq. (25) can be rewritten as
  
cation system is an end-to-end communication system, which
1 Y
comprises a transmitter and multiple receivers, as well as a xji,E = tanh−1  Wj lv + Wj xj−1
i,E 0
 .
random channel. According to the DL theory, this system 2 0 0 0 0
E =(v ,c ),v 6=v
can be regarded as a black box (i.e., an autoencoder) and is (26)
capable of completing the encoding process by the interaction A straightforward but extremely effective way to enhance
of the neurons. To simplify the description, we introduce the the reliability of the NOMA system is to constrain the BER.
indicator function, which can be expressed as 1si,1 . For this Thus, bj is expected to be eliminated at the output unit.
function, all the symbols (except si,1 ) are equal to zero; si,1 is According to min-sum approximation theory, Eq. (25) and Eq.
equal to one. Based on the proposed LSTM scheme provided in (26) can be rewritten as
section III-B, we propose an encoding and decoding algorithm  Y
0
j j−1 j
for the NOMA network, which is illustrated in Algorithm 3. xi,E = max min 0
|xi,E | − bi , 0 sign(xj−1
i , E ). (27)
E
E0
Algorithm 3 Encoding method for NOMA system.
Here, max(···, 0) is used to hinder the subtraction from flipping
Input: xi = {si,1 , si,2 , ..., si,q }, for all i, environment simu- the symbols of the signals, and sign(·) is the well-known sign
lator. function.
Output: ybi = {b si,1 , sbi,2 , ..., sbi,q }, for all i. Unfortunately, the typical gradient descent method is not the
1: Initialize the model. best choice for optimizing these ”min-sum” decoders, as the
2: Generate a set of data sequences from the training data function computed by the check nodes has non-differentiable
set. kinks at certain points. However, the good characteristic of the
3: Start the environment simulator to generate the wireless SGD is that it can select the subgradient during backpropaga-
channel, and mix specific man-made noise or distortion tion. Thus, the SGD is introduced to train the DL decoders.
into the channel. Additionally, for the sake of assessing the performance of
4: si,1 is encoded as a one-hot vector 1si,1 , and the encoded the proposed DL-based scheme in terms of encoding and
symbols are corrupted by the unknown channel environ- decoding, we introduce the BLER metric, which is expressed
ments generated by the simulator. as
5: Set the learning rate and the loss rate of the network. Kn
Meanwhile, initialize the weight of the LSTM as w = 0. 1 X
Prb = Pr(s̄i 6= si |si ), (28)
6: Train the LSTM network based on the given sequence Kn i=1
according to the proposed offline learning scheme by the where s̄i denotes the symbols after the encoding and decoding
SGD algorithm. process.
7: Update the weight w and the output of each layer of the
LSTM. IV. N UMERICAL R ESULTS AND A NALYSIS
8: The last layer of the LSTM network uses a softmax
A. Simulation results for the proposed LSTM scheme
activation to generate a probability distribution Pqn over all
In this part, we discuss a numerical analysis of our proposed
possible signals, which can be given as m=1 pm =
1, ∀pm ∈ {0, 1}. Also, the estimated symbol sbi,1 is paired LSTM-aided NOMA scheme with different parameters. Python
with the pm that has the highest probability. 3.6 and MATLAB are used in our performance analysis. To
9: return y bi . implement the proposed DL-based schemes on a GPU, the
total mini-batch is treated as a “big training example”, in
which the operation of the proposed approach is based on
the averaged values of all training samples in the mini-batch.
E. Decoding scheme for NOMA communication system
To establish the DL framework and process the training pro-
In the proposed scenario, the DL decoder is employed with cedure, the powerful Keras is employed to train the proposed
two hidden layers with 64 and 32 neurons each. Assuming system. We investigate the sum data rate in different cases, and
that there is a processing element in the jth hidden layer, this the BLER performance is discussed to assess the performance
layer is linked to a single input node in the previous layer and of the proposed scheme. Also, to investigate the proposed DL-
corresponds to the edge E = (v, c). Here, v and c are defined based SISD approach, an extensive simulation is conducted.
as the variable node in the input layer and the connected For the sake of simplicity, binary phase shift keying (BPSK)
node in the connected layers, respectively. The preceding part modulation is employed. It is noted that the samples are
illustrates the generality of the weight updating process. Here, generated randomly with Gaussian distribution. Additionally,
note that the hidden layer includes Iv , v = 1, 2, ..., Q nodes. the wireless channel is designed as the new radio model (WIN-
0
E denotes the edge of the j−1th layer. The signal propagating NER II) [29] and Rayleigh fading channel is deployed. To be
process of the ith user can be formulated as specific, the detailed simulation environment and parameters
 
Y in the proposed DL based scheme are presented in TABLE I.
xji,E = 2tanh−1  xj−1
i,E 0
, (25) Fig. 5 shows the simulation results in terms of the sum data
E 0 =(v 0 ,c0 ),v 0 6=v rate of the NOMA system for sequence length of 10 bits, 15

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
8

bits, 20 bits, and 25 bits. It can be seen from Fig. 5 that the sum
rate improves as the training sequence length increases. The
TABLE I
PARAMETER SETTING IN THE SIMULATION . reason is that total training power of the system is proportional
to the training sequences. Additionally, with the enhancement
Parameter Value or Versions of the signal-to-noise ratio (SNR), it can be observed that the
Learning framework Keras 2.0.7
Coding environment Ubuntu 14.04 LTS sum rate improves simultaneously.
Total training steps 400000 In this comparison, we investigate the simulation results in
Pre-training steps (RBM) 8000 terms of the sum rate versus SNR for our scheme and for the
Number of the training samples 250000 state-of-the-art NOMA schemes. As can be observed from Fig.
Total transmitted power 2W
Path loss exponent 2 6, the proposed LSTM scheme outperforms the conventional
Number of the NOMA users 32 schemes in terms of the sum rate. Here, the sequence length
Discount factor ε 0.99 is set as 30 bits in all the methods for fair comparison. This
result is supported by the fact that the DL technique can obtain
current and accurate CSI of the link between the BS and each
user, for which it can obtain a better power allocation strategy.
Thus, it can be further implied that the power allocation
performance can be optimized with the aid of DL, which is
one of the key principles of the NOMA concept.
In Fig. 7, the performance of the proposed DL-based
SISD strategy for user activity and data detection is well
investigated. Here, the learning rate is set to 0.01, and the
MSE performance is studied for different batch sizes. It
can be observed from Fig. 7 that the MSE performance of
the DL-based SISD scheme is optimized, as a larger SNR
contributes to a better channel environment. Also, as the
batch size increases, the performance of the proposed DL-
based SISD scheme declines, which occurs because a larger
batch size leads to slower convergence. However, it is noted
that a batch size that is too small is more likely to cause
unstable convergence behavior. Hence, batch size selection is
an important issue for the proposed DL-based SISD method,
which is a tradeoff between data detection performance and
algorithm stability.
Fig. 5. Sum rates for different sequence lengths.
Fig. 8 shows the MSE versus SNR for the proposed DL-
based NOMA scheme when the learning rate is set to 0.1, 0.02,
0.01, and 0.001. We can see from Fig. 8 that the performance
corresponding to user activity and data detection for the
proposed scheme is optimized by introducing a lower learning
rate, which is supported by the fact that a larger learning rate
induces a higher validation error. It must be noted, however,
that a lower learning rate will give rise to slower convergence,
although it can promote the performance of the DL-based
scheme. Thus, choosing an appropriate learning rate is of
great significance with regard to enhancing the performance
of the DL-based framework for accurate user activity and data
detection.
Fig. 9 depicts the BLER performance for different NOMA
schemes when the number of users is 6. As shown in Fig. 9,
the BLER performance is enhanced with increasing SNR, with
the lowest BLER being approximately 10−7 of the proposed
LSTM scheme. Also, compared with the hard decision (i.e.,
Hamming (7,4) coding) method, it can be seen that the
BLER performs better when implementing the proposed LSTM
scheme. For comparison, the BLER of uncoded data is also
provided. This result indicates that the DL autoencoders and
Fig. 6. Sum rate versus SNR for different NOMA-based communication
schemes (JSAC [30], TVT [31]). decoders have learned the channel characteristics without any
prior knowledge regarding the NOMA system. Additionally,
our training is conducted at a rate R = 4/7 with a learning

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
9

Fig. 9. BLER versus SNR for the LSTM and other baseline communication
schemes.
Fig. 7. MSE versus SNR for different batch sizes based on the validation
data set for the proposed scheme.

1
10-

2
10-
3-
1
0
江山T3

0
4
1

5
10-

�- Hard decision
6 0 Theory - Uncoded data
10- -Ar- Proposed LSTM scheme

口 Proposed LSTM+RTN scheme


1

4
2

6
3
SNR(dB)

Fig. 10. BLER versus SNR for different NOMA-based communication


schemes with NLOS propagation.

Fig. 8. MSE versus SNR for different learning rates based on the validation
data set for the proposed scheme. ilarly, uncoded data are introduced in the simulation for
comparison.
In Fig. 11, we study the relationship between the BLER
rate of 0.001.
and the training epochs in two different cases when the SNR
Paper [32] came to an initial conclusion that introducing is set to 1.2 dB. As the epoch increases, the BLER decreases
radio transformer networks (RTNs) to integrate expert knowl- initially before finally becoming stable. In particular, it can
edge into the DL model can further improve the end-to-end be observed from Fig. 11 that the BLER of the RTN-aided
training performance. Here, we deploy RTNs in our proposed LSTM scheme is lower than that of the LSTM scheme. Also,
NOMA system to elevate its performance. In Fig. 10, the we can further see that the proposed RTN-aided LSTM scheme
BLER of the proposed LSTM scheme, the proposed RTN-aided converges quickly compared with the proposed LSTM scheme,
LSTM scheme, and the typical hard decision coding scheme as its curve becomes smooth and initially remains unchanged.
are investigated under non-line-of-sight (NLOS) propagation. Hence, we come to an initial conclusion that RTNs can be
All the schemes were carried out at a rate R = 4/7. Clearly, used to optimize DL-based communication systems.
all the schemes based on LSTM achieve a lower BLER than
the hard decision scheme, which indicates that the NOMA
scheme based on DL achieves better performance in terms B. Performance evaluation of the proposed scheme
of reliability. Also, it can be seen that the RTN-aided LSTM This section evaluates the performance of the proposed
scheme outperform the LSTM scheme without RTNs, which NOMA scheme. Specifically, according to the principles of
occurs because the RTNs exploit expert domain knowledge the 10-fold cross-validation procedure, all the data used in
and simplify the complexity of the training procedure. Sim- section A are randomly divided into 10 subsets. Each time,

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
10

100 TABLE II
0 Proposed LSTM+RTN scheme F- SCORE FOR F OUR DL S CHEMES .
D Proposed LSTM scheme
1
10- Method LSTM CNN LNN DBN
F-score 0.76 0.68 0.60 0.74
2
10-
3
1
0

of the proposed LSTM scheme and other typical DL networks


正UJ
W3

(convolution neural network (CNN), linear neural network


4
0
1

(LNN), and deep belief network (DBN)). Also, the activation


function and number of the hidden layers are designed as
5
10- the proposed LSTM based algorithm. The results show that
the proposed scheme outperforms all others, as illustrated in
6
10- TABLE II.

10 20 30 40 50 60 70 80 90 100 V. C ONCLUSIONS
Training epoch
In this paper, we propose an effective DL-aided NOMA
Fig. 11. BLER versus training epochs with and without RTNs. scheme, the performance of which is well investigated. Taking
advantage of the RBM, the input signals are processed by
the RBM and then trained by the LSTM network. Different
from previous methods, the proposed scheme can model the
CSI by learning the environment automatically via offline
learning. Also, it is shown that the proposed LSTM-aided
NOMA system can achieve better performance in terms of
the BLER and sum data rate, and its robustness and high
precision are verified. Additionally, by incorporating the DL
into the SISD algorithm, it is proven that the DL-based
framework is very suitable for user activity and data detection
based on NOMA because of its superior performance. Our
work suggests that further research on NOMA systems should
consider integration with deep learning because of its powerful
learning ability. Additionally, we firmly believe that this is
the beginning of a field that incorporates DL or ML into
NOMA systems and that this study will attract great interest
Fig. 12. AUC curve for the proposed LSTM scheme. from the academic and industrial communities to investigate
the application of DL-based techniques in next generation
communication systems.
nine subsets are used as training data, while the remaining
subset, as testing data. Also, since the area under the receiver
operating characteristic (AUC) curve is a common index for R EFERENCES
assessing an algorithm, the AUC curve of the LSTM-aided [1] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “Systemlevel
NOMA scheme is provided. In addition, the false positive rate performance evaluation of downlink non-orthogonal multiple access
(NOMA),” in Proc. IEEE Annu. Symp. Pers. Indoor Mobile Radio
is calculated as the ratio between the number of negative events Commun. (PIMRC), London, U.K., Sep. 2013, pp. 611–615.
wrongly categorized as positive (false positives) and the total [2] L. Dai, B. Wang, Y. Yuan, S. Han, C. l. I, and Z. Wang, “Non-
number of actual negative events, while the true positive rate orthogonal multiple access for 5G: solutions, challenges, opportunities,
and future research trends,” IEEE Commun. Mag., vol. 53, no. 9, pp.
is used to measure the proportion of positive events that are 74-81, Sep.2015.
correctly identified. [3] H. Huang, J. Xiong, J. Yang, G. Gui, and H. Sari, “Rate Region Analysis
Fig. 12 shows the simulation results of the proposed LSTM- in a Full-Duplex-Aided Cooperative Nonorthogonal Multiple-Access
System,” IEEE Access, vol. 5, pp. 17869-17880, Aug. 2017.
aided NOMA scheme. It is observed that the AUC is 0.98, [4] Y. Chen, L. Wang, Y. Ai, B. Jiao, and L. Hanzo, “Performance Analysis
indicating that this scheme achieves a good performance in of NOMA-SM in Vehicle-to-Vehicle Massive MIMO Channels,” IEEE
terms of robustness and accuracy. This result is obtained J. Sel. Areas Commun., vol. PP, no. 99, pp. 1-1,Jul. 2017.
[5] Z. Wei, D. W. K. Ng, J. Yuan, and H. M. Wang, “Optimal Resource
because of the low requirements in terms of the data size Allocation for Power-Efficient MC-NOMA With Imperfect Channel
and data distribution. Furthermore, the F-score is commonly State Information,” IEEE Trans. Commun., vol. 65, no. 9, pp. 39443961,
used to evaluate the accuracy of an algorithm and can be Sep. 2017.
[6] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of
interpreted as a weighted average of the precision and recall. nonorthogonal multiple access in 5G systems with randomly deployed
Specifically, the F-score has a maximum (best) value of 1 and users,” IEEE Signal Process. Lett., vol. 21, no. 12, pp. 1501–1505,
a minimum (worst) value of 0. Here, to gain further insight Dec.2014.
[7] L. Yang, J. Chen, Q. Ni, J. Shi, and X. Xue, “NOMA-Enabled Coop-
into the performance of the proposed scheme with other DL erative Unicast-Multicast: Design and Outage Analysis,” IEEE Trans.
frameworks, we use the same data set to formulate the F-score Wireless Commun., vol. PP, no. 99, pp. 1-1. Sep. 2017.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2848294, IEEE
Transactions on Vehicular Technology
11

[8] J. A. Oviedo, and H. R. Sadjadpour, “A Fair Power Allocation Approach [32] Timothy J. O’Shea, and Jakob Hoydis ,“An Introduction to Deep
to NOMA in Multiuser SISO Systems,” IEEE Trans. Veh. Technol., vol. Learning for the Physical Layer ,” ArXiv preprint arXiv:1702.00832.,
66, no. 9, pp. 7974-7985, Sep. 2017. 2017.
[9] B. Wang, L. Dai, Z. Wang, N. Ge, and S. Zhou, “Spectrum and Ener- [33] B. Wang, L. Dai, T. Mir, and Z. Wang, “Joint User Activity and Data
gyEfficient Beamspace MIMO-NOMA for Millimeter-Wave Communi- Detection Based on Structured Compressive Sensing for NOMA,” IEEE
cations Using Lens Antenna Array,” IEEE J. Sel. Areas Commun., vol. Commun. Lett., vol. 20, no. 7, pp. 1473-1476, Jul. 2016.
35, no. 10, pp. 2370-2382, Oct. 2017.
[10] X. Yue, Y. Liu, S. Kang, A. Nallanathan, and Z. Ding, “Outage
performance of full/half-duplex user relaying in NOMA systems,” in
Proc. IEEE Int. Conf. Commun. (ICC) , Paris, May 2017, pp. 1-6.
[11] Y. Tan, J. Zhou, and J. Qin, “Novel Channel Estimation for Non- Guan Gui received the Dr. Eng degree in Informa-
orthogonal Multiple Access Systems,” IEEE Signal Process. Lett., vol. tion and Communication Engineering from Univer-
23, no. 12, pp. 1781-1785, Dec. 2016. sity of Electronic Science and Technology of China
[12] K. Senel, and S. Tekinay, “Optimal Power Allocation in NOMA Sys- (UESTC), Chengdu, China, in 2012. From October
tems with Imperfect Channel Estimation,” in Proc. 2017 IEEE Global 2009 to March 2014, he joined the wireless signal
Communications Conference, Singapore, 2017, pp. 1-7. processing and network laboratory (Prof. Adachi’s
[13] Y. Gao, B. Xia, Y. Liu, Y. Yao, K. Xiao, and G. Lu, “Analysis of the laboratory), Department of Communications Engi-
Dynamic Ordered Decoding for Uplink NOMA Systems with Imperfect neering, Graduate School of Engineering, Tohoku
CSI,” IEEE Trans. Veh. Technol., vol. PP, no. 99, pp. 1-1, Jan. 2018. University as for research assistant as well as post-
[14] D. Fan, F. Gao, G. Wang, Z. Zhong, and A. Nallanathan, “Angle Do- doctoral research fellow, respectively. From April
main Signal Processing-Aided Channel Estimation for Indoor 60-GHz 2014 to October 2015, he was an Assistant Professor
TDD/FDD Massive MIMO Systems,” IEEE J. Sel. Areas Commun.,vol. in Department of Electronics and Information System, Akita Prefectural
35, no. 9, pp. 1948-1961, Sep. 2017. University. Since November 2015, he has been a professor with Nanjing
[15] Timothy J. O’Shea, Tugba Erpek, and T.Charles Clancy, “Deep Learning University of Posts and Telecommunications (NUPT), Nanjing, China. He
Based MIMO Communications,” Arxiv preprint arXiv:1707.07980,2017. is currently engaged in research of deep learning and compressive sensing for
[16] Hinton, Geoffrey E., S. Osindero, and Y. W. The, “A Fast Learning physical-layer wireless communications. He is a senior member of IEEE.
Algorithm for Deep Belief Nets,” Neural Computation, vol.18, 2006.
[17] Tobias Gruber, Sebastian Cammerer, Jakob Hoydis, and Stephan ten
Brink, “On Deep Learning-Based Channel Decoding,” ArXiv preprint
arXiv:1701.07738,2017.
[18] T. Wang, C. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, “Deep Hongji Huang is currently pursuing his bachelor’s
Learning for Wireless Physical Layer: Opportunities and Challenges,” degree of communication engineering at Nanjing
ArXiv preprint arXiv:1706.01151, 2017. University of Posts and Telecommunications, Nan-
[19] S. Shen, C. J. Chang, and L. C. Wang, “A cellular neural network and jing China, from 2014. His research interest is 5G
utility-based radio resource scheduler for multimedia CDMA communi- communications based on full duplex communi-
cation systems,” IEEE Trans. Wireless Commun., vol. 8, no. 11, pp.5508- cations, convex optimization and power allocation.
5519, Nov. 2009.
[20] M. Kim, N. I. Kim, W. Lee, and D. H. Cho, “Deep Learning Aided
SCMA,” IEEE Commun. Lett., vol. PP, no. 99, pp. 1-1, Jan. 2018.
[21] H. Ye, G. Y. Li, and B. H. Juang, “Power of Deep Learning for Channel
Estimation and Signal Detection in OFDM Systems,” IEEE Wireless
Communications Lett., vol. 7, no. 1, pp. 114-117, Feb. 2018.
[22] Z. M. Fadlullah et al., “State-of-the-Art Deep Learning: Evolving
Machine Intelligence Toward Tomorrow’s Intelligent Network Traffic
Control Systems,” IEEE Commun. Surveys & Tutorials, vol. 19, no. 4,
Yiwei Song is currently pursuing his bachelor’s
pp.2432-2455, May 2017.
degree of Computer Science at Nanjing University
[23] F. Tang et al., “On Removing Routing Protocol from Future Wireless
of Posts and Telecommunications, Nanjing China,
Networks: A Real-time Deep Learning Approach for Intelligent Traffic
from 2015. His research interest is data mining, deep
Control,” IEEE Wireless Commun., vol. PP, no. 99, pp. 1-7, Oct. 2017.
learning based communications.
[24] N. Kato et al., “The Deep Learning Vision for Heterogeneous Network
Traffic Control: Proposal, Challenges, and Future Perspective,” IEEE
Wireless Commun., vol. 24, no. 3, pp. 146-153, Jun. 2017.
[25] B. Mao et al., “Routing or Computing? The Paradigm Shift Towards
Intelligent Computer Network Packet Transmission Based on Deep
Learning,” IEEE Trans. Comput., , vol. 66, no. 11, pp. 1946-1960, Nov.
2017.
[26] Salakhutdinov R, Mnih A, and Hinton G, “Restricted Boltzmann ma-
chines for collaborative filtering,” in Proc. IEEE Int. Conf. Mach.
Learn.(ICML), Oregon, 2007, pp. 1-8.
[27] Y. Sung, L. Tong, and H. V. Poor, “Neyman-pearson detection of gauss- Hikmet Sari is currently Professor of Nanjing U-
Markov signals in noise: closed-form error exponentand properties,” niversity of Posts and Telecommunications (NUP-
IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1354-1365, Apr. 2006. T), and also Chief Scientist of Sequans Communi-
[28] F. A. Gers, and E. Schmidhuber, “LSTM recurrent networks learn sim- cations. His distinctions include the IEEE Fellow
ple context-free and context-sensitive languages,” IEEE Trans. Neural Grade (1995), the Andr Blondel Medal (also in
Networks, vol. 12, no. 6, pp. 1333-1340, Nov. 2001. 1995), the Edwin H. Armstrong Achievement Award
[29] P. Kyosti, “IST-4-027756 WINNER II D1.1.2 v.1.1: WINNER II Chan- in 2003, the Harold Sobol Award in 2012, as well
nel Models,” 2007, [online] Available: http://www.ist-winner.org. as election to Academia Europaea (the Academy
[30] P. Xu, and K. Cumanan, “Optimal Power Allocation Scheme for Non- of Europe) and to the Science Academy of Turkey
Orthogonal Multiple Access With α -Fairness,” IEEE J. Sel. Areas in 2012. Prof. Sari has served as an Editor of
Commun., vol. 35, no. 10, pp. 2357-2369, Oct. 2017. the IEEE Transactions on Communications (1987 -
[31] Y. Liu, Z. Ding, M. Elkashlan, and J. Yuan, “Nonorthogonal Multiple 1981), a Guest Editor of the European Transactions on Telecommunications
Access in Large-Scale Underlay Cognitive Radio Networks,” IEEE (1993) and of the IEEE JSAC (1999), and an Associate Editor of the IEEE
Trans. Veh. Technol., vol. 65, no. 12, pp. 10152-10157, Dec. 2016. Communications Letters (1999-2002). He served as a Distinguished Lecturer
of the IEEE Communications Society in 2001 -2006, as a member of the
IEEE Fellow Evaluation Committee in 2002 - 2007, and as a member of the
Awards Committee in 2005 - 2007.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like