You are on page 1of 5

The Role of Deep Learning in NOMA for 5G and Beyond

Communications
Moh. Khalid Hasan*, Md. Shahjalal, Md. Mainul Islam, Md. Morshed Alam, Md. Faisal Ahmed, and Yeong Min Jang**
Dept. of Electronics Engineering, Kookmin University, Seoul, South Korea
Email: *khalidrahman45@ieee.org, **yjang@kookmin.ac.kr

Abstract— In the coming future, it is obvious that the wireless In NOMA, a multiuser signal is multiplexed using
networks will be congested with massive amounts of data traffic superposition coding in the transmitter and then sent to the users
with the increasing number of users. Current multiple access using different power levels in a non-orthogonal basis. Usually,
techniques will certainly not have the capability to efficiently serve the users with bad channel conditions are allocated higher
in the massively congested scenarios. In recent times, non-
power than the users with good channel conditions. Therefore,
orthogonal multiple access (NOMA) has been recognized as an
immensely potential technique for 5G and beyond
having proper channel state information (CSI) in the transmitter
communications that can increase spectral efficiency to a greater has a profound significance in NOMA. When the signal is
extent serving a vast number of users. However, several collected in the receiver, the user with the stronger channel gain
circumscriptions are observed in NOMA such as the requirement retrieves its signal immediately. On the other hand, the user
of a perfect channel state information in the transmitter and high with the weaker channel recognizes the other signals as
computational complexity in the receiver. The use of deep learning interference and performs successive interference cancellation
(DL) techniques is a great resolution to deal with the challenges. (SIC) technique to retrieve its original signal. In the
This paper discusses the applications of the DL methods in NOMA conventional orthogonal multiple access systems, the
for 5G and beyond communications. Firstly, the deep neural
orthogonality between channels is achieved using guard periods
networks that are employed in NOMA are listed up and discussed.
After that, their functions are studied specifically focusing on how to avoid interference, which particularly decreases the spectral
to improve the NOMA performance. Finally, the possible future efficiency to a great extent. This problem is completely
challenges and research issues are identified at the end of the remediated using NOMA, where no guard period is used and so
paper. signals overlap with each other. NOMA provides fantastic
throughput with maintaining an excellent level of user fairness
Index terms— Non-orthogonal multiple access (NOMA), deep
[9].
learning (DL), and 5G.
In terms of theoretical perspective, SIC can be considered
I. INTRODUCTION as an optimal scheme for both uplink and downlink
communication [10]. With the perfect SIC, the performance of
Next-generation wireless networks will require an
NOMA is superior compared to the existing OFDMA scheme
extensive capacity to support an immense amount of advanced
[11]. Perfect CSI is very crucial for performing SIC in the
multimedia applications [1, 2]. The fifth-generation (5G)
receiver. It is pretty difficult to design an efficient power
technologies face significant challenges with the explosive
allocation scheme without having the exact CSI at the
growth of internet-of-thing-based large scale heterogeneous
transmitter. Acquiring a perfect or near-perfect CSI is a
networks. The uplink/downlink transmission of future massive
challenging task. Fortunately, deep learning (DL) techniques
incremental data to a huge number of users will face numerous
are already being proved to be a useful resolution to this
hindrances using the current multiple-access techniques due to
problem. It can be incorporated in NOMA to detect a
their incapability to support efficient spectrum usage [3, 4]. The
completely unknown and sharply changing channel condition.
necessity of a new method for 5G and beyond communication
In this paper, we enlist the functions of DL to improve the
in handling a large quantity of data is realized by researchers
performance of NOMA systems. We first provide a short
over the last few years [5].
overview of the DL methods that are useful in NOMA and other
Non-orthogonal multiple access (NOMA) is a promising wireless networking applications. Afterward, we specifically
technique that has drawn significant research interest in recent discuss the applications of DL in NOMA. Finally, we provide
years due to its potential applications in 5G and beyond a short overview of the challenges that can be observed using
networks [5, 6]. It brings numerous advantages to the current DL methods in NOMA.
wireless networks such as an increase in spectral efficiency,
reliability, and connectivity reducing overall latency. NOMA II. DEEP LEARNING IN NOMA
exploits the existing resources in a more efficient basis than the
The current improvements in DL algorithms have brought
conventional multiple access techniques [7, 8].
significant benefits to the area of wireless communications.

978-1-7281-4985-1/20/$31.00 ©2020 IEEE 303 ICAIIC 2020


Input layers Encoder layer Decoder layer

x xˆ ( H , P, yi )

Hidden layers Noise layer Output layer

Fig. 1. DL-based NOMA communication architecture.

Data in wireless systems, such as CSI, are increasingly real-time and the input signals can be trained with the help of
heterogeneous and display intricate correlations [12]. Efficient that data. This type of learning is called online learning. A block
training of these data can exhibit better performance. In NOMA, diagram of the NOMA system with the integration of both
the employment of DL is extensively researched over the last online and offline training for the auto-detection of CSI is
few years. Different kinds of learning algorithms are utilized in depicted in Fig. 2. DL can also be used for estimating the
different aspects and demonstrate better performance compared channel considering adverse weather. For example, in [14], a
to traditional schemes. In general, three types of learning DL technique, called long-short term memory (LSTM) is used
methods are developed in the literature, such as supervised, for detecting dynamically fluctuating channel. Herein, the
unsupervised, and reinforcement. Supervised learning is hidden layers can act as a memory for the network state, which
extensively utilized to estimate the CSI. In addition, it can also particularly permits the DL to save, recall, and process the
be used in other applications such as spectrum sensing, preceding complex data.
localization, and throughput prediction. Unsupervised learning
DL can also be employed in NOMA for other purposes as
is mostly used in user clustering and congestion control. Deep
follows.
reinforcement learning is a different type of DL method that is
considered to have immense significance in wireless a. Resource allocation
communications. Until now, the most recognized application of
Scheduling can be very important in the cases where
reinforcement learning is resource management. It uses a
NOMA subchannels cannot provide sufficient bandwidth to the
special type of learning algorithm, called Q learning, and
users. A proper resource allocation algorithm can maximize the
produces the output without having any preliminary knowledge
overall sum rate. Therefore, the performance of the NOMA
about the learning environment.
system greatly depends on the resource allocation strategies
In NOMA, DL is employed in different applications. DL [11]. DL is a useful method that can be applied to overcome the
can be extremely useful for complex data processing and to resource allocation problem in the case of multiple users [15,
acquire perfect CSI. Two types of training can be identified in 16]. In addition, dynamic resource allocation is a challenge for
NOMA, such as online and offline training. A DL-based a fast-changing channel condition. Optimizing energy
NOMA model [13] is depicted in Fig. 1. The hidden layer is consumption with efficient resource allocation is a significant
employed for training and recognition [13]. Hidden layers are research issue using NOMA. Intelligent learning methods are
particularly equipped with multiple neurons. In Fig. 1, one of extensively developed to deal with this challenge. In [17], a DL
the layers is named as noise layer, where the processed signal method, called deep belief network, is proposed for resource
can be corrupted using additive-white Gaussian noise. The CSI allocation by minimizing the total transmit power.
can be acquired automatically by extensive training of the input Reinforcement learning can also be useful in solving the
data using existing channel models. Arbitrary sequences of the dynamic resource allocation problem [15]. In reinforcement
input signals are extensively trained using CSI of different learning, a deep deterministic policy gradient is used for
environments collected from simulations. This type of training deriving the resource allocation decision. In addition, the
is called offline training. In addition, the output signals can be subcarrier assignment decision is evaluated with the help of Q
fed into the training data to accurately predict the channel learning.
environment. Pilot signals can be utilized to carry the CSI in

304
Transmitter Channel Receiver

P1
Signal 1 Signal 2
Signal 1 detection detection
Signal 1 Power
Fading
Modulation allocation ∑ AWGN
channel
Signal 2 deployment P1>P2
Signal 1 Subtraction of
Signal 2 detection signal 1
P2

Offline Real-time
Simulated data CSI attainment Online training
training channel data

Channel estimation using deep learning

Fig. 2. Block diagram of the NOMA system employing online and offline DL mechanisms.

b. Signal decoding neural networks employed in NOMA provides a high


computational cost. Although NOMA itself is a complex
The conventional SIC method has several disadvantages.
networking system, DL techniques have enhanced its
In particular, it is very difficult to apply the SIC method to
performance but also increases the design complexity at the
decode the data perfectly when the number of users is increased.
same time. Therefore, reducing the computational complexity
The SIC method is also affected by the propagation error. The
is a significant research issue for the future. Moreover, using a
deep neural network can be useful to recover a discrete
specific neural network, different researchers used different
sequence from an impaired signal utilizing signal classification
architecture. The number of layers and the quality and quantity
[18]. It is also possible to optimize the SIC method by using DL
of the training dataset is different for a different architecture,
[19]. In [20], an online learning detection method for NOMA
however, the optimized system is yet to be discovered.
uplink is proposed to enhance spectral efficiency.
In addition, some of the DL methods, such as deep
c. Signal constellation design reinforcement learning is employed in NOMA in very few
literature. It is envisioned that the reinforcement learning
Efficient constellation design can be another useful
mechanism can optimize the performance of wireless networks
application of DL in NOMA [21]. The constellation of multiple
to a greater extent in terms of throughput, latency, and resource
users is superimposed while transmitting the data. The main
management. Extensive employment of reinforcement learning
theory is the mapping of the input source bits into symbol
in future NOMA systems is an essential research challenge.
sequences using DL after applying modulation techniques. To
increase energy efficiency, a new layer, called the
IV. CONCLUSIONS
normalization layer is added to connect the transmitter [21].
Another DL method to optimize the constellation design is NOMA is recognized to have immense significance in 5G
discussed in [22] based on autoencoders. The developed and beyond communication systems for its high spectral
mechanism is useful for multi-user access under various efficiency and low-latency characteristics. DL can enhance its
channel environments. performance to a considerable extent. In this study, the specific
roles of the DL methods in different NOMA applications are
DL can also be used in other applications in NOMA, such
briefly discussed. The concept of how DL techniques improve
as subchannel matching [23], power minimization [17], and
NOMA performance is explained. In addition, the specific DL
latency optimization [24]. The main applications of DL in
methods that are used in literature are listed up and their
NOMA are summarized in Table 1.
functions are elaborated. Finally, a short discussion on the
possible future research issues is conducted.
III. CHALLENGES AND FUTURE ISSUES
DL methods are widely employed in wireless ACKNOWLEDGMENT
communications to improve performance. Extensive
This research was financially supported by the Ministry of
development of DL in NOMA is already observed in recent
Trade, Industry and Energy(MOTIE) and Korea Institute for
times. However, there are still some challenges that are yet to
Advancement of Technology(KIAT) through the
be resolved by researchers. The efficient design of the core
InternationalCooperative R&D program.
neural networks is still a headache. Most of the current deep

305
Table 1: Applications of DL-methods in NOMA
Reference Type of learning Deep learning model Application
[11] Supervised Recurrent neural network Rapid and optimized resource allocation
[14] [13] Supervised LSTM Channel estimation
[15] Reinforcement Deep Q-network Subcarrier assignment and power allocation
[16] Reinforcement Attention-based neural network Joint power allocation and channel assignment
[17] Supervised and Deep belief network Power optimization
unsupervised
[18] Supervised Deep neural network Automatic realization of the CSI
[21] Unsupervised Deep neural network Signal constellation design
[25] Supervised Deep neural network Reliability improvement of grant-free access
[26] Supervised learning Deep neural network Optimization for energy-efficient scheduling
[27] Reinforcement Deep neural network [27] based on Power allocation optimization
Q-learning [28]

[8] A. Li, Y. Lan, X. Chen, and H. Jiang, “Non-orthogonal multiple


REFERENCES access (NOMA) for future downlink radio access of 5G,” China
[1] M. G. Kibria, K. Nguyen, G. P. Villardi, O. Zhao, K. Ishizu, and Communications, vol. 12, pp. 28-37, Dec. 2015.
F. Kojima, “Big data analytics, machine learning, and artificial [9] M. Aldababsa, M. Toka, S. Gökçeli, G. K. Kurt, and O. Kucur1,
intelligence in next-generation wireless networks,” IEEE “A tutorial on nonorthogonal multiple access for 5G and
Access, vol. 6, pp. 32328-32338, 2018. beyond,” wireless communications and mobile computing, vol.
[2] W. H. Chin, Z. Fan, and R. Haines, “Emerging technologies and 2018, 2018.
research challenges for 5G wireless networks,” IEEE Wireless [10] K. Higuchi and A. Benjebbour, “Non-orthogonal multiple
Communications, vol. 21, no. 2, pp. 106-112, Apr. 2014. access (NOMA) with successive interference cancellation for
[3] L. Dai, B. Wang, Y. Yuan, S. Han, C. I, and Z. Wang, “Non- future radio access,” IEICE TRANSACTIONS on
orthogonal multiple access for 5G: solutions, challenges, Communications, vol. 98, pp. 403-414, 2015.
opportunities, and future research trends,” IEEE [11] M. Liu, T. Song, and G. Gui, “Deep cognitive perspective:
Communications Magazine, vol. 53, no. 9, pp. 74-81, Sep. 2015. resource allocation for NOMA-Based heterogeneous IoT with
[4] M. K. Hasan, N. T. Le, M. Shahjalal, M. Z. Chowdhury, and Y. imperfect SIC,” IEEE Internet of Things Journal, vol. 6, no. 2,
M. Jang, “Simultaneous data transmission using multilevel LED pp. 2885-2894, Apr. 2019.
in hybrid OCC/LiFi system: concept and demonstration,” IEEE [12] M. A. Alsheikh, D. Niyato, S. Lin, H. Tan, and Z. Han, “Mobile
Communications Letters, vol. 23, no. 12, pp. 2296-2300, Dec. big data analytics using deep learning and apache spark,” IEEE
2019. network, vol. 30, no. 3, p. 22–29, 2016.
[5] S. M. R. Islam, N. Avazov, O. A. Dobre, and K. Kwak, “Power- [13] G. Gui, H. Huang, Y. Song, and H. Sari, “Deep learning for an
domain non-orthogonal multiple access (NOMA) in 5G effective nonorthogonal multiple access scheme,” IEEE
systems: potentials and challenges,” IEEE Communications Transactions on Vehicular Technology, vol. 67, no. 9, pp. 8440-
Surveys & Tutorials, vol. 19, no. 2, pp. 721-742, 2017. 8450, Sep. 2018.
[6] W. Shin, M. Vaezi, B. Lee, D. J. Love, J. Lee, and H. V. Poor, [14] M. AbdelMoniem, S. M. Gasser, M. S. El-Mahallawy, M. W.
“Non-orthogonal multiple access in multi-cell networks: theory, Fakhr, and A. Soliman, “Enhanced NOMA system using
performance, and practical challenges,” IEEE Communications adaptive coding and modulation based on LSTM neural network
Magazine, vol. 55, no. 10, pp. 176-183, Oct. 2017. channel estimation,” Applied Sciences, vol. 9, no. 15, 2019.
[7] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance [15] Y. Zhang, X. Wang, and Y. Xu, “Energy-efficient resource
of non-orthogonal multiple access in 5G systems with randomly allocation in uplink noma systems with deep reinforcement
deployed users,” IEEE Signal Processing Letters, vol. 21, no. learning,” in Proc. International Conference on Wireless
12, pp. 1501-1505, Dec. 2014. Communications and Signal Processing (WCSP), Xi'an, China,
2019.

306
[16] C. He, Y. Hu, Y. Chen, and B. Zeng, “Joint power allocation and
channel assignment for NOMA with deep reinforcement
learning,” IEEE Journal on Selected Areas in Communications,
vol. 37, no. 10, pp. 2200-2210, Oct. 2019.
[17] J. Luo, J. Tang, D. K. C. So, G. Chen, K. Cumanan, and J. A.
Chambers, “A deep learning-based approach to power
minimization in multi-carrier NOMA with SWIPT,” IEEE
Access, vol. 7, pp. 17450-17460, 2019.
[18] C. Lin, Q. Chang, and X. Li, “A Deep Learning Approach for
MIMO-NOMA Downlink Signal Detection,” Sensors, vol. 19,
2019.
[19] J. Kang, I. Kim, and C. Chun, “Deep learning-based MIMO-
NOMA with imperfect SIC decoding,” IEEE Systems Journal,
pp. 1-4.
[20] D. A. Awan, R. L.G. Cavalcante, M. Yukawa, and S. Stanczak,
“Detection for 5G-NOMA: an online adaptive machine learning
approach," arXiv:1711.00355.
[21] L. Jiang, X. Li, N. Ye, and A. Wang, “Deep learning-aided
constellation design for downlink NOMA,” in Proc.
International Wireless Communications & Mobile Computing
Conference (IWCMC), Tangier, Morocco, 2019.
[22] F. Alberge, “Constellation design with deep learning for
downlink non-orthogonal multiple access,” in Proc. IEEE
Annual International Symposium on Personal, Indoor and
Mobile Radio Communications (PIMRC), Bologna, 2018.
[23] Y. Liu, H. Zhang, K. Long, A. Nallanathan, and V. C. M. Leung,
“Energy-efficient subchannel matching and power allocation in
NOMA autonomous driving vehicular networks,” IEEE
Wireless Communications, vol. 26, no. 4, pp. 88-93, Aug. 2019.
[24] P. Yang, L. Li, W. Liang, H. Zhang, and Z. Ding, “Latency
optimization for multi-user NOMA-MEC offloading using
reinforcement learning,” in Proc. Wireless and Optical
Communications Conference (WOCC), Beijing, China, 2019.
[25] N. Ye, X. Li, H. Yu, A. Wang, W. Liu, and X. Hou, “Deep
learning aided grant-free NOMA toward reliable low-latency
access in tactile internet of things,” IEEE Transactions on
Industrial Informatics, vol. 15, no. 5, pp. 2995-3005, May 2019.
[26] L. Lei et al., “Learning-assisted optimization for energy-
efficient scheduling in deadline-aware NOMA systems,” IEEE
Transactions on Green Communications and Networking, vol.
3, no. 3, pp. 615-627, Sep. 2019.
[27] K. N. Doan, M. Vaezi, W. Shin, H. V. Poor, H. Shin, and Tony
Q. S. Quek, “Power allocation in cache-aided NOMA systems:
optimization and deep reinforcement learning approaches,”
arXiv:1909.11074.
[28] L. Xiao, Y. Li, C. Dai, H. Dai, and H. V. Poor, “Reinforcement
learning-based NOMA power allocation in the presence of smart
jamming,” IEEE Transactions on Vehicular Technology, vol.
67, no. 4, pp. 3377-3389, Apr. 2018.

307

You might also like