DNN Based Approach to Signal Decoding in Downlink SISO-NOMA

ABSTRACT

With ever increasing number of users and high data transfer speed requirements, from 2Kbps in 1G era to 10Gbps in 5G networks, the current communication and transmission systems are facing challenges in terms of data transfer speed, spectral efficiencies, low latency rates and requirement of high transmission bandwidth.

Multiple access (MA) techniques are being utilized to serve large number of users and perform data handling. Non-orthogonal multiple access (NOMA) wireless system has been introduced to enhances the spectrum efficiency and support massive connectivity. Among NOMA schemes, power-domain multiplexing is most preferred one due to ease in its implementation in existing communication infrastructure.

In a power-domain based NOMA network, the base station transmits the combined signal, which is a superimposition of the desired signals of multiple users with different allocated power levels, to all the users. At each receiver end, successive interference cancellation process is performed individually until the respective users signal is recovered. SIC decoding process involves complex algorithms. Implementation of successive interference cancellation requires accurate knowledge of both the channel model and channel state information (CSI), which are difficult to acquire. SIC performance is degraded in the presence of realistic imperfect channel state information (CSI) and with increase in number of users SIC process complexity increases linearly.

This paper presents an alternative approach to decoding of multiplexed signals utilizing Deep Neural Network where detection is performed in a data-driven manner and does not rely on complex mathematical formulation. Deep Neural Networks (DNN) are recognized for their ability to effectively deduce information in complex environments, relying solely on data utilized for their mappings. DNNs advantage is not limited to improvement in the performance gains of the communication system, but also in the reduction of reference signal overhead to boost throughput in the downlink system.

In this study, a system configuration of a base station equipped with a single antenna along with two user equipment (UE) each equipped with a single antenna, having different channel gains is considered in a downlink scheme. A SISO-NOMA detection system based on Deep Neural Networks (DNN) has been considered in the receiver block. This DNN based decoding block accepts the received superimposed signals from the base station and decodes them utilizing deep learning techniques.

In this study, a comparative assessment of the performance of this proposed DNN based decoding architecture and conventional SIC decoding is presented.

TABLEOFCONTENTS

Page

LISTOFTABLES......................................................... LISTOFFIGURES........................................................			vi vii
CHAPTER
1 Introduction.........................................................			1
	1.1 Background -Wireless Revolution...................................				1
	1.2 A Brief Overview Of Various MA schemes 2 Overview of NOMA.....				2
	1.3 A Brief Overview Of NOMA.........................................				5
	1.4 Deep Learning in Wireless Communication...................... ...				7
	1.5 Brief to Thesis Objectives.........................................				8
	1.6 ContributionstotheThesis.......................................				9
2 Introduction to Non-orthogonalMultipleAccess(NOMA)Scheme......			10
	2.1 WhatisNOMA?...............................................				10
	2.2 BenefitsofNOMAoverOMAsystems.. .. . .. . ... ....................				12
	2.3 Key Aspects Associated with NOMA. . ..... ........................				13
3Thesis Statement......................................................			22
	3.1 Constraints with Present Approach..............................				22
	3.2 Objectives.......................................................				23
	3.3 Brief on the Study Performed.....................................				24
4System Configuration...... ...... .....................................			26
	4.1 System Description...... ...... ...... ............................				26
5Deep Neural Network for Signal Decoding in a Downlink SISO NOMA System			29
		5.1 Background. . . . . . . . . . .. . . . . .. . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..				29
		5.2 Overview of Machine Learning. . . .. . . . . . . .. . . .. . . . . . . . . . . . . . . . . . . . . . .				29

	5.3 Application of DL in NOMA Systems. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .			32
	5.4 Introduction to DNN Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . .			41
	5.5 Structure and components of DNN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .			41
	5.6 Training of the DNN Model . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . .			43
	5.7 DNN Learning Processes. . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .			44
	5.8 Study DNN Model Description. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . .			45
	5.9 DNN Model Summary . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . .			50
6Results Presentation and Performance Evaluation. . . . . . . . . .. . . . . . . . . . . . . .		52
	6.1 Introduction. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .			52
	6.2 Results Presentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .			52
	6.3 BPSK Modulation Scheme Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .			53
	6.4 QPSK Modulation Scheme Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .			57
7Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .		59
8Future Work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .		60
REFERENCES. ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 61

APPENDIX

AMATLABCode - SIC Decoding 66

B TensorFlow Code DNN Decoding..67

LISTOFTABLES

Table Page

5.1 DNN Model Summary51

LISTOFFIGURES

Figure Page

• Illustrationofa Generalized DownlinkNOMAScheme19

4.1 Two User Downlink SISO-NOMA..26

5.1 BPSK Modulation Based Neural Network..46

5.2 QPSK Modulation Based Neural Network 47

6.1 BER vs SNR Curve - BPSK Modulation - Conventional SIC Decoding (User 1 = 0.7, User 2 = 0.3)....54

6.2 BER vs SNR Curve - BPSK Modulation DNN Based Signal Decoding (User 1 = 0.7, User 2 = 0.3)....54

6.3 BER vs SNR Curve - BPSK Modulation - Conventional SIC Decoding (User 1 = 0.6, User 2 = 0.4) ......55

6.4 BER vs SNR Curve - BPSK Modulation - DNN Based Signal Decoding (User 1 = 0.6, User 2 = 0.4) ...56

6.5 BER vs SNR Curve - QPSK Modulation - Conventional SIC Decoding (User 1 = 0.8, User 2 = 0.2) ......57

6.6 BER vs SNR Curve - QPSK Modulation - DNN Based Signal Decoding (User 1 = 0.8, User 2 = 0.2) ......58

Chapter1

INTRODUCTION

1.1 Background - Wireless Revolution

Wireless communication is defined as the transfer of information (voice, data, video etcetera) from one point to the other without using any physical medium, such as wires or cables. With the help of wireless communication, communication boundaries became defunct, the transmitter and receiver could be anywhere a few meters apart or thousands of kilometers apart. Today, wireless communication plays an important role in our day to day lives, right from mobile phones, WiFi, GPS, to smart household devices (IoT). The biggest advantage of wireless communication systems is mobility, the freedom to move around while still being connected to the network. Apart from mobility, there is data rate (data can be transmitted over large distances in a matter of seconds), accessibility, and constant connectivity.

Since wireless communication systems use a wireless channel medium for transmitting data, there is a high chance that one communication system may interfere with another system or network, thereby causing interference. In such instances, when data is transmitted over a wireless medium, the receiver receives a corrupted signal. This is because the wireless channel is susceptible to a variety of channel impediments such as path loss, interference and multi-path fading which affects the data rate, spectral efficiency and reliability of the wireless transmission.

Several mathematical models and multiple access schemes (MAs) have been developed by researchers in recent past to counter fading, path-loss effects and interference whilst still achieving the minimum performance metrics (such as data-rates, low bit-error rate (BER) or symbol error-rate (SER), low latency, spectral efficiency (SE), energy efficiency (SE)). With rapidly increasing number of mobile users, we see today it is important that these aforementioned performance metrics are met in order to provide the quality of service (QoS).

Spectral efficiency (SE) is a key aspect in wireless communication but there is a hard-limit to how much data can be transmitted over a channel, i.e. the Shannons Limit. Shannons Limit is defined as the maximum rate at which information can be transferred over a communications channel in the presence of noise. This introduces another challenge to current wireless systems. Specific algorithms have been formulated to attain an efficient transfer of information with low error probability and low latency. The use of machine learning is also being explored to attain data rates closer to the Shannons Limit and along with reliability and low latency (4).

1.2 A Brief Overview of Various MA schemes

With the increase in the usage of wireless communication systems beyond voice communications to data handling and transmission, the bandwidth of the current wireless system has exceeded leading to the development of various multiple access schemes to enhance the data handling capabilities of the existing networks.

Previous 1G/2G/3G telecommunication systems are suitable for voice communication only and cannot handle high data rate transmission and burst data traffic requirements. As per ITU recommendation a 4G system can handle data rates up to 100 Mbps for high mobility and up to 1 Gbps for low mobility . Even these rates are considered low with the increase in data transmission requirements, 5G system peak data handling rate is 10 Gbps and this is with improved parameter tuning.

Multiple access schemes have been used for all technologies from 1G to 5G due to the limitation in the wireless network infrastructure. In a multiple access scheme, a number of independent users share a resource. These schemes can be divided into three categories: orthogonal access schemes, controlled access schemes, and random-access schemes.

Most multiplexing schemes are based on the concept of dividing a communication link into independent communication channels. The three methods used to divide a communication link into independent channels are as follows: frequency-division multiplexing (FDM), time-division multiplexing (TDM), and code-division multiplexing (CDM). Orthogonal multiple access schemes refer to techniques that allow two or more users to share radio frequency spectrum in a manner that avoids collision. The main concern of the random-access schemes is transmission scheduling to minimize the probability of packet collision [51].

FDMA works on the principle of dividing the total bandwidth of the communication channel into a number of discrete segments, and allocating each segment exclusively to a user. Guard bands are used between each segment of the frequency band to prevent interference between users. The advantage of the FDMA system is its simplicity, since once the channel capacity is divided amongst users each can operate independently of each other. Since each user has exclusive access to its allocated bandwidth there is no contention and therefore no wastage of bandwidth or delays caused by collisions and retransmissions. The disadvantage of FDMA systems is that there is wastage of bandwidth, firstly caused by the usage of guard bands and secondly due to the fact that users can only use their own allocated frequency bands. FDMA adoption in telecommunication became more widespread in the 1970s and 1980s.

TDMA divides up data streams from multiple users into uniform time intervals and assigns them to channels, all of which share a single transmission path. Adoption of TDMA began in the 1990s, and it was used to carry mobile phone calls under the digital 2nd generation mobile communication system (2G).

CDMA scheme involves generating signals in which the data streams of users are mixed with unique identifier codes, and the signals of all the users are overlaid within the same frequency band and conveyed via a single transmission path. CDMA scheme started under the 3rd generation mobile communication system (3G) introduced in the 2000s [52]. The received signal is correlated with the users assigned spreading code to extract the user information as part of CDMA decoding. As part of decoding, a unique method for signal decoding was introduced in CDMA technology and that is successive interference cancellation (SIC) [49].

OFDMA, or Orthogonal Frequency Division Multiple Access, is a popular multiple access scheme used even today that allows multiple users to share the same frequency band simultaneously by dividing it into orthogonal subcarriers. This was introduced in 2004, however its applications came into the market in 2010s. In OFDMA, data is transmitted using different subcarriers that are orthogonal to each other, meaning they do not interfere with each other. This enables efficient and simultaneous transmission of data from multiple users within the same frequency band. OFDMA offers several advantages over other multiple access schemes such as CDMA , TDMA as it provides better spectral efficiency, improves system capacity and enhances overall network performance. 4G LTE technologies adopt orthogonal frequency-division multiple access (OFDMA) scheme to deliver twice the spectral efficiency as compared to third generation (3G) systems using CDMA technology [52].

1.3 A Brief Overview Of NOMA

All the above discussed multiple access schemes work on the principle of orthogonality. According to this principle, each users signal is assigned a particular resource element (frequency, time, or code) that is orthogonal to the other users signal resulting in minimum interference at the receiver. In OMA based schemes, there are only limited number of orthogonal resources available to the users. With increasing data transmission requirements and existing communication infrastructure, OMA based schemes cannot cope with the ever- increasing requirements of providing high-speed data-rates, low latency and massive connectivity. This has led to the development of Non-Orthogonal Multiple Access (NOMA) systems.

The main idea behind NOMA is to overcome the limited user capacity in OMA, that is, to support more users than the available orthogonal resources [2]. Thus, the resource allocation across users become non-orthogonal, and the support for a greater number of users comes at the cost of increase in receiver complexity [54]. Non-orthogonal multiple access (NOMA) techniques are envisioned as a promising solution for 5G and beyond systems [54].

Compared to orthogonal multiple access (OMA) techniques, NOMA can simultaneously transmit users information over the same time-frequency resources through a non-orthogonal resource allocation (RA) scheme, yielding a high spectral efficiency and serving a greater number of users. The basic idea of NOMA is to use non-orthogonal resource allocation (RA) schemes at the transmitter.

There are predominantly two non-orthogonal resource allocation schemes: Power Domain and Code Domain. NOMA employs strategies such as superimposed coding (SC) at the transmitter and successive interference cancellation (SIC) at the receiver to facilitate the transmission of data to multiple users. While these techniques prove to be useful in their own fields, there are constraints associated with them.

Power Domain NOMA can be visualized as a multiple access scheme in a new domain, namely the power-domain, as opposed to other resource domains such as time, frequency or code. At the transmitter side, different information bearing signals corresponding to different users are superimposed on each other with different power levels, over a single resource block, that is, over the same time, same frequency or with the same code. In the receiver, successive interference cancellation (SIC), a popular power-domain based multi-user detection algorithm is utilized to decode/separate each users respective signal.

The philosophy of Code Domain NOMA (CD-NOMA) is different from Code Domain Multiple Access (CDMA), code domain multiple access (CDMA) is primarily based upon the idea that users are separated by exploiting the differences among their spreading codes, whereas NOMA encourages multiple users to employ exactly the same code [5]. However, there is also a similarity in working of code domain NOMA and CDMA, except that the codes used in CD-NOMA are either low-density or non-orthogonal with a low cross-correlation value [54].

1.4 Deep Learning in Wireless Communication

With the increase in complexity in multiplexing of signals to achieve higher demands in data transfer rates without compromising QoS, complex algorithms and coding is required both at the transmitter and receiver ends to code and decode the signals. Conventional approaches of algorithms for coding and decoding are facing a challenge with increase in number of users, resulting in poor signal quality.

With the increasing capability of machine learning techniques in solving complex problems, researchers have gained a growing interest in Deep Learning to address the challenges faced in wireless communication. Deep Learning involves a data-driven, and self-learning approach to solve complex problems. Deep learning, a subset of machine learning involves deep neural networks and sophisticated algorithms to help with the extraction of intricate patterns and representations from complex data.

Deep Neural Networks (DNN) involve processing of the input data which passes through layers of sophisticated math modeling to predict the output in a hierarchical fashion. Each layer extracts features from the input data, and deeper layers build upon these features to make complex predictions or classifications. DNNs are fundamental to many machine learning tasks, including image recognition and natural language processing.

This capability of DNN to solve complex problems without using conventional algorithms has generated significant interest and a number of research works to improve various aspects associated with the NOMA system design such as improved throughput, reduced BER, low latencies, resource allocation, channel condition estimation, signal detection and user fairness have been introduced. Various research performed in these areas are discussed in chapter 5.

This thesis has been inspired by the advances in DNN capabilities and aims to study the potential of utilizing DNN techniques for decoding signals at the receiver instead of using conventional multi-user decoding techniques.

1.5 Brief to the Thesis Objectives

In this work, we present an alternative to conventional SIC decoding by introducing Deep Neural Network (DNN) models to decode the superimposed signal received at the receiver block in lieu of conventional SIC decoding. This technique utilizes a data-driven approach for decoding the received signal.

Furthermore, two types of modulation techniques have been studied: BPSK and QPSK signals for ascertaining the performance of this DNN based signal decoding.

A comparative performance assessment of the results obtained with this new DNN based signal decoding approach and the conventional SIC decoding approach has also been conducted, utilizing Bit Error Rate (BER) versus Signal-to-Noise Ratio (SNR) curves.

A system configuration comprising of a base station (BS) equipped with a single antenna along with two user equipment (UE) each equipped with a single antenna, having different channel gains is considered. The focus of the research is primarily in the detection system of the receiver block. For the purposes of the study, we have considered a downlink SISO-NOMA configuration.

1.6 Contributions to the Thesis

This research is inspired by previous contributions on DNN-based transceivers designed for non-orthogonal multiple access systems and includes the works of Mohamed A. Aref and Sudharman K. Jayaweeras paper Deep Learning aided Successive Interference Cancellation for MIMO-NOMA [12], Jae-Mo Kang, Il-Min Kim et al paper Deep Learning Based MIMO-NOMA with Imperfect SIC Decoding [25] and C Lin et.al paper Deep Learning Approach for MIMO-NOMA in Downlink Signal Detection [11].

Chapter 2

INTRODUCTION TO NON-ORTHOGONAL MULTIPLE ACCESS (NOMA) SCHEME

• WhatisNOMA?

To overcome the limitations associated with OMAschemesin terms of providinghighspeeddata rates,lowlatency and massive connectivity utilizing an existing communication infrastructure, Non-orthogonalMultipleAccess(NOMA)schemeshave evolved.

Non-orthogonalMultipleAccess(NOMA)schemeisamultipleaccesstechnique thatwasintroducedinthelate2000sandhasgainedsignificantattentionfromresearchers today.Unlike orthogonal multiple access (OMA) schemes where each users signalisallocatedanorthogonalresourceelementallowinglimitednumberofusers totransmitoveracommunicationchannel,non-orthogonalmultipleaccessscheme allows multiple users to transmit their signals simultaneously over a communication channel. A way to differentiate a NOMA system from an OMA system is, if the number of users is greater than the number of resource elements (RE), then it is defined as a NOMA scheme. However, if the number of users are less than the number of resource elements (RE), then it is defined as an OMA scheme.

The main intent behind NOMA is to overcome the limited user capacity experienced in OMA, that is, to support more users than the available orthogonal resources [54]. Thus, the resource allocation across users become non-orthogonal, and the support for extra set of users comes at the cost of increase receiver complexity [54].

Historically, NOMA was proposed as an interference-mitigation technique in CDMA systems, where a number of non-orthogonal sequences more than the number of chips were needed to support greater number of users [54].

NOMA schemes can be divided into the following two categories [54]:

1. Power-domain NOMA: In this NOMA scheme, different power levels are allocated to different users, depending on the quality of their respective channels. That is, different users are serviced over the same orthogonal resource (time, frequency or code), but with different power levels. The power allocation to each user/signal is inversely proportional to the user /signal strength i.e. strong channel conditions correspond to lower power allocation levels and vice versa. At the receiver end, the difference in this power level is exploited by using the successive interference cancellation technique. Broadly speaking, each user decodes the signals of other users and removes them before decoding its own signal, depending on the allocated power.

2. Code-domain NOMA: The working of code-domain NOMA is similar to that of CDMA, except that the codes used are either low-density or non-orthogonal with a low cross-correlation value.

However, the non-orthogonality of resources in NOMA systems introduces inter-user interference which necessitates the need for robust signal decoding schemes in the receiver.

In this study, we have focused on power-domain non-orthogonal multiple access (PD-NOMA) systems.

• BenefitsofNOMAoverOMAsystems

The future of wireless networks will have to support high-speed data rates, low- latency and massive connectivity. According to the International Telecommunication Union (ITU)[9],fifth generation networks will have to support several features [10]

• aminimumpeakdatarateof10Gbps(100timesmorethanthatinthe3rd Generation Partnership Project (3GPP) Long-Term Evolution (LTE))
• alatencyof1millisecond(tentimeslowerthanthatin4Gnetworks),and
• connection density of over a million cellular devices per square kilometer (which is 100 times more than 4G networks)

Spectral efficiency (SE) and user fairness is improved in a NOMA system by exploiting the capacity-achieving scheme in the downlink transmission. NOMA systems, popularly the Power-Domain NOMA (PD-NOMA) systems are employed to improve wireless communication through the following benefits:

• Massive-Connectivity:NOMAsystemsareknowntoprovidemassiveconThis is because NOMA allows multiple users to share resource elements(RE)non-orthogonallyviapowercontrolalgorithmsandtheprinciple ofsuperpositioncoding(SC).InOMAsystemshowever,thenumberofusers beingservedislimitedbythenumberofresourceelements(RE),forexample, inTDMAeachuserallottedatimeslot.Thenumberoftimeslotsarelimited and no two users can share the time slot.
• SpectralEfficiency(SE):NOMAprovidesbetterspectralefficiency(SE)anduser fairnesscomparedtoOMAsystemsdue tothefactthateveryuserinaNOMAsystemcanavailtheentirebandwidth, in an OMA system though each user is allocated a small fraction of the resourceelement(inverselyproportionaltothenumberofusers).NOMAisalso knowntobeanefficientenablerandcanbemergedwithnewtechnologiessuch asMIMOandmillimeterwave(mmWave)systemstherebyimprovingspectral efficiencyandincreasingsystem
• LowLatency:Latency requirements in 5G systems are very diverse and stringent in certain cases. For example, ITU requires a user plan latency of 4ms and 1ms for enhanced mobile broadband (eMBB) and ultra-reliable and low latency communications (URLLC), respectively[9].In OMA systems, it is very difficult to maintain such delay requirements because a user can transmit a data packet only when there is a resource element (RE) available. This results in high latencies which is considered to be inefficient in a wireless communication system. NOMA, in contrast, provides flexible scheduling since it can accommodate a number of users based on the application and perceived quality of service (QoS).

• Key Aspects Associated with NOMA

As mentioned in previous sections, a NOMA system involves superposition coding of the users signals over non orthogonal resources at the transmitter, power allocation of the users signals and multiuser detection primarily successive interference (SIC) based decoding at the receiver. These key methodologies and their associated constraints are discussed in this section.

• Superposition Coding (SC)

Superposition Coding (SC) is a technique often seen at the transmitter where signals of multiple users are superimposed, then modulated and transmitted over the wireless channel. An important aspect to note here is that superposition coding does not take place in uplink transmission.

• Power Control in NOMA Systems

Power control strategies are used in NOMA systems to mitigate interference and ensure user fairness.

In Power Domain based NOMA (PD-NOMA) systems, power coefficients are allotted to respective user based on the channel gains of each user. Assigning the right power coefficient to a users signal is in itself a part of a separate study commonly known as Power Control. This is similar to CDMA systems, where power control plays an important role in solving the near-far user problem. The near-far problem arises due to the presence of cellular devices being everywhere within the cell including the cell boundaries. Some devices may be near the base station (BS) and some may be near the edge of the cell boundary. This results in interference where the user closer to the base station (BS) overshadows the signal of the user that is farther away from the base station. Implementing a tight power control algorithm among users within the same cell is needed in order to improve the overall data handling capacity of the network and ensure user fairness.

Consider a 2-user downlink NOMA scheme where the total transmission power is considered as unity. The BS allocates a fraction of the total power to user i.e. where . User is allocated which is given by as the criteria must be met. BS performs superposition coding (SC) of the users signals before transmitting it over to the receivers. Each user receives the superimposed signal. For decoding, successive interference cancellation (SIC) takes place in the following order, the user with the strongest channel gain performs SIC decoding to cancel the interference and decode its signal free of interference. The user with the weakest channel treats the other users signal as noise and decodes its signal. By varying the values of between and , we attain different rate-pairs (, ) on the boundary of the capacity region of the downlink NOMA scheme. For each (, ) on the boundary of the capacity region there is one and only one power-coefficient such that and are the optimal powers of user and user , respectively[7].

Two popular power allocation strategies are considered in power-domain based NOMA (PD-NOMA) systems: power allocation based on channel state information (CSI) and power allocation based on pre-defined quality of service (QoS). These strategies are adapted from the works of Mohamed M. El-Sayed, Ahmed S. Ibrahim et al in the paper Power Allocation Strategies for Non-Orthogonal Multiple Access [8].

For the purposes of this study, we have focused on power allocation based on channel state information (CSI). In this technique, the power levels of each user are calculated based on the channel conditions experienced by all users. The BS assigns power levels to the users based on the information sent by the NOMA users over the control channels. The BS then groups the users based on this received information which contains channel information. The user with a better channel condition (higher channel gain) is grouped with the user with a weaker channel condition. Then each group will be assigned to a frequency block. The users signals will be multiplexed together based on their assigned powers in such a manner that the total transmitted power per resource block shall not exceed , that is

In this strategy, the relationship between the users power share and their respectivechannel gain is inversely proportional, this is given by

This means that the user with a weaker channel condition (lower channel gain) is assigned a higher power level while the lowest power level is assigned to the user with strongest channel condition. From equation,the power coefficient of the user is given by,

where

Considering equations and,we get a more generalized equation,

Equation (2.6) meets the criteria of ensuring that the powers assigned to the respective NOMA users does not exceed

Power Control is crucial for the fair and efficient operation of cellular systems.

Allocating higher power to users with weak channel conditions is motivated by supporting a certain quality of service (QoS) and this is usually quantified by ? where is the minimum data-rate needed for user . For weaker users, this usually implies allocating more power to compensate for weaker channel conditions, but in general it actually depends on the value of [7].

• SIC Decoding

A unique feature incorporated into the receiver architecture of a power-domain NOMA (PD-NOMA) system is successive interference cancellation (SIC). SIC receiver is classified as a multi-user detector (MUD). Successive interference cancellation (SIC) typically requires the channel state information (CSI). The basic idea behind SIC is to detect and omit interference (signal of other users is considered as interference apart from noise) and recover the respective users signal. An important property of PD-NOMA is power allocation (PA) which is normally done at the transmitter. The user with the weakest channel condition is allocated the highest transmission power whereas the user with the strongest channel condition is allocated the lowest transmission power. This is done so as to ensure user fairness.

In a PD NOMA system, signal sent by the Base station is a combination of all users signals, with each user signal being assigned a different power level. At the receiver end, successive interference cancellation (SIC) decoding is performed iteratively until the users signal is recovered.

SIC depends on the users channel state information (CSI). Power coefficients are assigned to the users CSI, in an inversely proportional manner. The user with a weak channel condition is allocated a high transmission power and user with a strong channel condition is allocated a low transmission power. Furthermore, since the user with the highest transmission power considers the signals of the other users as noise, it recovers its signal directly without performing SIC. However, all the remaining users will perform SIC. In SIC, each user first detects the signals that are stronger than its own desired signal. Next, those signals are subtracted from the combined received signal and this process continues till the required users signal is determined. Finally, each user decodes its own signal by treating other users with lower power coefficients as noise. The figure below presents the flow scheme of the process involved.

The superimposed signal at the base station can be expressed as follows:

where is the information of the user. is the total transmission power at the

BS and is the power coefficient allocated to the user with the constraint . Furthermore, as we scale the number of users to say , the power coefficients of each user will be in the following order

Without the loss of generality, the channel gains are assumed to be ordered as 22 22, whereis the Rayleigh fading channel coefficient of the user. The received signal at the user can be described as follows:

where is the zero mean complex additive white gaussian noise with a variance of , .

Both these techniques are not unfamiliar and have been employed in many NOMA and OMA papers such as [4 ; 5 ; 6]. SC- SIC is used popularly in NOMA particularly, in the power-domain scheme because of its capacity-achieving ability in several cases such as in a SISO or MIMO networks making OMA schemes sub-optimal.

• SIC: Implementation Issues

In the previous sections we have observed that successive interference cancellation (SIC) plays an important role for decoding the superimposed signal in a downlink NOMA scheme. The complications associated with successive interference cancellation (SIC) in a wireless system are discussed below[40]:

• ComplexityinScaling: Considering a downlink transmission,eachuserwillreceiveasuperimposedsignalfromthe Inorderforausertodecodeitsownsignal,itwillhavetoperformSIC iterativelytodecodeandcancelotheruserssignalsbeforedecodingitsown.As thenumberofusersinacellincreases, theSICdecodingbecomesmorecomplex fortheindividualuserswhichisunacceptable.
• Error-PropagationEffects:Inapracticalscenario,successiveinterference cancellation(SIC)neveroccursIntheprevioussections,weassume error-freedecodingbutofcourse,withactualcodes, errors are made.Once an error occurs for one user, all the users later in the SICdecodingorderwilllikelybedecodedincorrectly. Thus, when the number of users sharing a particular orthogonal resource is high, the error propagation due to SIC degrades the performance in terms of bit error probability.
• Analog-to-DigitalQuantizationError:Whenthereceivedpowersofthe usersarehighlyvarying,theanalog-to-digital(A/D)converterneedstohave a very large dynamic range, and at the same, enough resolution to quantize thecontributionfromtheweaksignalThisposeshardwarechallenges indesign of a SIC receiver.

Chapter3

THESIS STATEMENT

3.1 Constraints with Present Approach

Power-Domain based NOMA (PD-NOMA) systems employ strategies such as superimposed coding (SC) at the transmitter and successive interference cancellation (SIC) at the receiver to facilitate the transmission of data to multiple users. While these techniques prove to be useful in their own fields there are major constraints associated with them.

SIC decoding algorithm is channel dependent, i.e. it relies on the knowledge of the underlying statistical channel model. In particular, it requires the user to have accurate channel knowledge of the channel between the Base Station (BS) and each of the users. This affects the performance because determining accurate channel information is very challenging especially in rapidly fluctuating environments. Furthermore, in the SIC algorithm the interference signals are assumed to be subtracted out completely which is never the case. This is due to presence of non-linearities due to hardware impairments, resulting in imperfect SIC decoding causing error propagation effects. An important thing to note here is that in power-domain NOMA the users channel conditions need to be disparate in order to efficiently decode multiple users signals. Another important constraint with the SIC decoding algorithm is the mathematical complexity involved in decoding a users signal as we scale the number of users.

3.2 Objective

Inspired from the research being done in developing and addressing the various challenges associated with a conventional NOMA system through the use of deep learning techniques, we have also developed a simple deep neural network (DNN) for signal decoding. The DNN model introduced in this research proves to improve the signal decoding performance when compared with the conventional SIC decoder by achieving low bit error rates among signals having similar power allocation.

This approach does not rely on complex mathematical formulation but is based on learning the detection rule in a data-driven manner utilizing Deep Neural Networks (DNN).

In this paper we primarily focus on the receiver architecture of a wireless system and assess the proposed alternative to conventional SIC decoding technique. A feed-forward neural network architecture that takes input the received data from the receiver antenna, detects the data as output. This problem of signal detection can be considered to be a classification problem as it entails recovering a discrete signal from an impaired signal. The approach we describe in the coming sections is part of a supervised learning problem. A key assumption made in the approach is that the channel state information (CSI) is known to both the transmitter and receiver.

DNNs advantages lie not only is in the improvement of the performance gains of the communication system, but in the reduction of reference signal overhead to boost throughput in the downlink system. Furthermore, using this approach we hope to reduce the mathematical complexities involved with conventional SIC decoding particularly when involving large number of users.

3.3. Brief on the Study Performed

To achieve the research objectives, following key tasks were performed:

1. Design: a unidirectional downlink SISO-NOMA detection system in the receiver is designed utilizing deep learning (DL) techniques. This technique accepts the superimposed signal directly as input in the receiver block, eliminating the need for a conventional successive interference cancellation (SIC) block.
2. Developing a DNN model utilizing TensorFlow to harness the DNN capabilities to handle complex data and enhance detection performance through a data-driven approach. Training and Testing the DNN model to achieve improved detection results.
3. Signal Decoding: Performing signal decoding for a SISO-NOMA system utilizing conventional SIC based decoding on MATLAB and parallelly performing signal decoding using the DNN model under the same conditions and parameters.
4. Performance Evaluation: conducted a thorough evaluation of the results obtained from the proposed DNN based signal decoding system with results obtained from the traditional SIC decoding based algorithm.
5. Assessment on the impact of various parameters on the simulation results: through simulations, we examined the influence of key parameters such as modulation type and user power allocation levels.

Chapter4

SYSTEM CONFIGURATION

4.1 System Description

This study considers a two-user downlink Power Domain based NOMA (PD-NOMA) system. A single antenna is used at the base station (BS). Each user is equipped with a single antenna. The two users are located at different distances from the base station. User is assumed to be positioned farther away from the base station and User is nearer to the base station (refer to the diagram above).

Given the position of each user, we consider user channel gains to be greater than user channel gains, i.e. . The power allocation is performed based on the channel state information as described in Section . The relationship between the power share per user and the users channel gain is inversely proportional, i.e.

where user with the weakest channel gain is assigned the highest power level and the user with the highest channel gain is assigned the lowest power. The total power constraint is considered to be watt for simplicity and must be respected,

At the BS transmitter, the broadcast signal is given as

where and is the superposition coded (SC) signal given by

where signals and are considered to be the modulated symbols.

and are obtained from the equation in section Power Control in NOMA Systems. represents Rayleigh fading coefficients of the user by sampling from a complex gaussian random distribution with zero mean and variance . is the additive white gaussian noise with mean zero and variance . In this research, we have assumed a time-invariant channel (TIV) where the channel characteristics do not change over time.

At the receiver end, the received signal is of the form

A key assumption in this research is that the channel state information (CSI) is known to both the transmitter and the receiver.

In this work, to assess the decoding capabilities of the developed DNN model, two signal modulation techniques have been studied. i.e.

• BPSK (binary-phase shift keying) and
• QPSK (quadrature-phase shift keying).

Chapter5

DEEP NEURAL NETWORKS FOR SIGNAL DECODING IN A DOWNLINK SISO-NOMA SYSTEM

• Background

In the previous section we gave a brief introduction to NOMA and introduced different NOMA schemes. We also discussed the issues associated with present NOMA based systems. In this section, utilization of machine learning techniques to address issues associated with NOMA systems are discussed.

Deep Learning (DL), which is a part of the broader family of machine learning has been employed in the domain of wireless communication technology particularly in 5G and future communication networks. It is a very powerful tool for handling big data and solving complex non-linear problems. (8) discusses implementing different deep learning (DL) techniques in NOMA systems for 5G and future wireless communication systems.

• Overview of Machine Learning

Machine learning (ML) is the ability of a computer to think for itself and perform a task without any explicit instructions. It is a subset of artificial intelligence (AI). After the 80s and 90s, interest in the application of data-driven artificial intelligence (AI) techniques prompted many scientists and engineers to conduct research in a number of engineering fields, including speech and image recognition, pattern classification and wireless communications.

Let us compare the ideologies of a conventional engineering approach to a machine learning approach in the design of an algorithmic solution. In order to design a system through conventional means, one must have an in-depth knowledge of the subject before designing any given system first. This is followed by creating a mathematical model that captures the physics of the set-up under study. Lastly, based on the model, an optimized algorithm is developed that provides system performance guarantee under the assumption that the model created is an accurate

representation of reality (1). On the contrary, the machine learning approach is quite simple and does not require any prior domain knowledge. In order to design this system, a sufficient amount of data/information adhering to the problem of interest is collected, constituting what is called a training set. This set of information is then fed to a learning algorithm to produce a trained machine that carries out the desired task. This learning is made possible by the choice of a set of possible hypothesis class, from which the learning algorithm makes a selection during training. An example of a hypothesis class is given by a neural network architecture with learnable parameters. The performance of the learning algorithm is based on how well the hypothesis class matches the available data (1).

There are three main classes of machine learning: supervised learning, unsupervised learning and reinforcement learning. In supervised learning, pairs of a given input and desired output belonging to a training set are given/known, and the main objective is to map this input to the output through the use of a learning algorithm. supervised learning is the most commonly used machine learning technique and is commonly used in classification and regression models. In Unsupervised learning, the training set consists of just the input variables and no corresponding output. The goal of this learning technique is to model the distribution of data in order to learn more about the data. Unsupervised learning is popularly used for clustering and anomaly detection. Reinforcement learning (RL) is a machine learning technique that enables an agent to learn in an interactive environment through trial and error using feedback from its own actions and experiences. Though both supervised learning and Reinforcement Learning use input to output mapping functions, unlike supervised learning where the feedback provided to the agent are correct sets of action to perform the task, Reinforcement Learning uses system of rewards and punishments as signals for positive and negative behavior. Reinforcement Learning is often used in Robotics and Automation.

As S. Kannan has described in his paper (2), machine learning (ML) is primarily used when a given problem lacks one of the following: a model or an algorithm.

• Model Deficient: When a problem is model-deficit, that means there is exists no mathematical model due to there being insufficient domain knowledge on the subject. As a result, the conventional engineering approach discussed earlier becomes inapplicable.
• Algorithm Deficient: A problem is said to be algorithm-deficit, when there exists a well-defined mathematical model but the existing algorithms optimized on the basis of such model are too complex to be implemented for the given application. In such a case, we tend to machine learning to yield low-complexity solutions.

• Application of DL in NOMA Systems

There is a great interest in applying Deep Learning techniques to address the challenges associated with implementation of NOMA systems on a large scale. This section presents the ongoing research work in this field.

In NOMA systems, determining channel-state information is very important and by implementing Deep Learning techniques this hurdle can be overcome. In (9), a long short-term memory (LSTM) network based on deep-learning is proposed to detect channel characteristics automatically and resolve the issues of conventional NOMA systems. Similarly, papers (10)-(11), introduce a deep neural network (DNN) for channel estimation and signal detection in a joint manner, to mitigate error propagation effects in SIC and increase throughput. In paper (12), the authors introduce a novel deep learning based successive interference cancellation (SIC) receiver for an uplink MIMO-NOMA system. Separate DNNs are used to decode each users signal at every iteration. The DNN performs particularly 3 tasks: channel estimation, signal detection and cancelling decoded signals from the received combined signal. The introduction of DNN aims to tackle error propagation effects and high computational complexities.

Similar to previous papers, authors of (13) introduce a deep learning (DL) aided receiver to perform signal detection, channel estimation and demodulation jointly in an end-to-end manner for an OFDM-NOMA scheme. The proposed deep learning (DL) method improves on system performance and robustness with a tap-delayed line channel model.

The authors in paper (14) introduce a novel deep learning (DL) aided receiver to overcome the shortcomings of conventional successive interference cancellation (SIC) receiver in MIMO-NOMA systems. The proposed deep learning method can restore the transmitted signal from the combined received signal. In order to obtain faster convergence on the deep learning (DL) model, the authors introduce zero-forcing/matched filter to optimize the training process of the neural network.

The research in paper (15) discusses the drawbacks of conventional NOMA systems apart from the earlier mentioned, them being limited channel feedback and compatibility of NOMA system with adaptive coding and modulation schemes. The authors of the paper in regards to the existing problems propose an effective channel estimation (CE) algorithm based on the long-short term memory (LSTM) neural net-work to dynamically adapt to the changes in channel characteristics. Furthermore, a novel power coefficient allocation algorithm is proposed based on Binomial Distribution and Pascals Triangle. To further improve the bit error-rate (BER) performance, adaptive BCH codes, constellation rotation, and cyclic Q-delay are added to the new and improved NOMA system.

Many deep learning-based networks have been used to resolve the issues of receiver complexity and system performance. For example, considering an uplink-NOMA system as discussed in (16), a separate sparse neural network detector and decoder based on the message passing algorithm (MPA) and belief propagation (BP) algorithm respectively are designed and then cascaded together to form a larger neural network for multi-user detection and decoding. Weights are assigned to the edges of the factor graph and deep learning is used to train these weights to achieve high spectral efficiency and system performance.

As stated in paper (17), NOMA can be identified as a constellation-domain multiplexing system. One of the few downsides of conventional NOMA systems is that it simply superimposes several single-user constellations (in downlink-transmission) without considering the interactions between multiple data streams. In paper (17), a Deep Learning aided downlink-NOMA scheme is proposed by parameterizing the bit-to-symbol mapping and multi-user detection with deep neural networks (DNN). This scheme has shown to significantly lower the symbol error rate than the conventional downlink-NOMA system. It is very important to have a carefully designed system to recover data of a super-constellation in a NOMAs receiver system. Therefore, a joint optimization framework for a receiver is proposed in paper (18) through the use of Auto Encoders (AE), a neural network, without having to perform SIC at the receiver (not subjected to error propagation effects). Inspired by Minhoe Kims research paper, Deep-Learning Aided SCMA, Minsig Han et al in their research paper (19) have proposed an auto-encoder (AE) structure generalized for both sparse and dense CD-NOMA systems and an effective loss function for the multi-user multi-dimensional modulation (MU-MDM) design of code-domain NOMA schemes. Their loss function considers Euclidean distance and Hamming distance between input and output data of the auto-encoder through an end-to-end learning process for minimizing the bit error rate (BER).

Other probably different applications of NOMA systems utilizing deep learning techniques are seen in papers [(20), (21), (22)]. The research in paper (20) discusses the issues that NOMA downlink systems for satellite-based Internet of Things (IoT) suffer from, particularly, rate-control and power allocation. The paper addresses these issues using the Lyapunov optimization algorithm to tackle the long-term power allocation scheme by dividing it into the rate-control and power allocation subproblems. Using the weight relationship between the channel state and queue state an optimal decoding order using deep learning is derived. Lyapunov optimization algorithm is seen as an effective means here to achieve stability in this particular dynamic system. Unlike other papers, authors of papers (21) - (22) consider a grant-free NOMA system. In a grant-free NOMA system, data transmissions are autonomously activated by the users without the explicit dynamic grant, thereby significantly reducing the control/user plane latency caused by signal interaction and scheduling. A downside to this approach is, due to the decentralized nature of the grant-free NOMA system, the information received at the receiver is unknown. Furthermore, it leads to inter-user interference causing deteriorated signal transmissions. Also since each device transmits information without scheduling, a process to identify active devices in a cell is required.

To address the latency issues, (21) proposes a deep learning approach to a conventional grant-free NOMA system to achieve low-latency in Tactile Internet of Things (IoT). Its done by designing a variational auto-encoder (VAE) in an end-to-end manner to parameterize the variational optimization problem, which incorporates the random user activation, symbol spreading, and multi-user detection (MUD). To identify the active users in a cell of a NOMA grant-free system, authors of paper (22) have presented a novel Deep Neural Network (DNN) based Active User Detection (AUD) scheme.

Sparse code multiple access (SCMA) is another promising NOMA scheme capable of simultaneously providing both massive connectivity and high spectral efficiency. The performance of SCMA ,however heavily depends on the design of the codebook. To reduce the computational complexities [because the codewords in a codebook are not orthogonal to each other and are made up of multidimensional complex values] of designing a codebook, deep learning strategies are applied. Paper (7) proposes an auto-encoder structure of deep neural networks (DNN) for performing mainly two tasks: mapping of the input data to sparse resource blocks and decoding the received signal, to minimize the bit error rate (BER) of the NOMA system.

In research paper (23), a novel convolutional neural network (CNN) based decoding for SCMA is proposed where a blind decoding strategy is applied to reduce receiver complexity and improve performance. The proposed neural network will be able to decode the received symbols without prior knowledge of the channel information or signal characteristics more efficiently compared to the iterative message passing algorithm (MPA) used in conventional SCMA systems.

Earlier research on NOMA systems assumed that the SIC process at the receivers end is performed perfectly. However, in practical scenarios, error propagation during the SIC process can occur. As a result, the receiver cannot cancel the other users interference. In papers [(24), (25), (26), (27)] resource allocation with the aid of Deep Learning techniques is performed in a NOMA system with Imperfect SIC. In (24), a novel deep learning based power allocation scheme is designed with imperfect SIC by considering power allocation as a sum rate maximization problem.

In (25), the authors propose another novel scheme to design a non-linear precoder in the transmitters side and a non-linear SIC decoder in the receivers end to jointly optimize the NOMA system by minimizing the total mean square error of the users signal using deep learning. A deep forward neural network (FNN) is constructed to perform the joint optimization of this non-convex problem. In (26), a similar approach is taken however, the application is seen in heterogenous IoT Networks. A peeling based subchannel assignment and re-weighted message passing algorithm (ReMPA) are optimally derived as resource allocation schemes for mobile users (MUs) and the IoT-based users (IoTUs) respectively. Through mapping the derived weights into the neural network, the authors construct an innovative resource allocation based recurrent neural network (RNN) architecture and apply deep learning algorithms to implement the resource allocation schemes.

Finally, authors of paper (27) tackle the problem of resource management in two stages: a. power allocation, a deep neural network (DNN) is designed and optimized based on the interior point method (IPM) which is a popular optimization tool and b. User Scheduling, a user scheduling algorithm is used by the BS to select preferred matching sub-channel for the users with high channel gain.

Another key issue of 5G and future communication networks apart from efficient spectrum utilization is energy efficiency (EE). We observe that with advancements in wireless communication systems the energy consumption in systems today is almost outrageous. Papers (28)-(29) discuss an effective NOMA scheme of implementing simultaneous wireless information and power transfer (SWIPT) technology to ensure efficient energy harvesting and information recovery without any compromise in the Quality of Service (QoS) in multi-carrier NOMA (MC-NOMA) systems. Focusing primarily on minimization of transmit power in the down-link transmission, paper (28) introduces Deep Belief Networks (DBN) to solve this high-power minimization problem.

Similarly, paper (29) presents a novel scheme where the power splitting (PS) - SWIPT is applied in a multi-carrier NOMA (MC-NOMA) system with the aid of deep belief network (DBN) so as to achieve optimal data rate of the system with the constraints of transmit power and energy harvesting requirement. Similar works have been seen in MIMO-NOMA systems. For instance in paper (30), a novel communication deep neural network (CDNN) is designed to realize better power allocation performance for sum-rate and energy efficiency optimization. The whole MIMO-NOMA system is regarded as a black box and the proposed neural network works on an end-to-end performance optimization.

Another branch of machine learning which has shown promising results in wireless communication systems is reinforcement learning (RL). Reinforcement Learning involve goal-oriented algorithms aimed at maximizing a complex objective (goal/rewards) along a particular dimension over many steps. By implementing reinforcement learning to existing NOMA systems, to address issues of resource allocation (RA) and channel assignment [in real-time] in conventional NOMA systems. Works [(31), (32), (33), (34)] provide just this.

In (31), an anti-jamming MIMO NOMA transmission game is formulated using a fast Q-based NOMA power allocation scheme that combines hot-booting techniques and Dyna architecture. The proposed scheme accelerates the learning process thereby, improving the communication efficiency against a smart jammer.

(32) introduces NOMA to mobile edge computing (MEC) and uses reinforcement learning (RL) algorithm deep Q-Network (DQN) to reduce offloading latency in a multi-user scenario.

Mobile edge computing (MEC) is an up-and-coming technology seen in 5G networks where intensive computing, storage and networking resources are integrated into base stations (BS). It will encourage users to offload their intensive operations and applications to mobile-edge computing facilities (MEC server) thereby reducing latency in network and service operations. However, it is seen in mobile-edge computing that it takes more time and energy to offload the tasks to the server. Therefore, the authors in (32) take NOMA features and combine it with mobile-edge computing and propose a reinforcement learning algorithm based deep Q-network to find the optimal user combination state, minimizing the total delay of this hybrid system.

In (33), the authors propose a reinforcement learning aided power domain NOMA system to dynamically allocate the power-coefficient such that the energy efficiency (EE) of the system is maximized. The reinforcement learning (RL) algorithm chosen to optimize the strategy of power allocation is based on the actor-critic framework. Similar to the research work conducted in paper (33), the authors of paper (34) introduce a novel approach to optimizing the online power-allocation policy in a multi-user downlink NOMA system by making use of reinforcement learning in shallow neural networks. They do so by building an actor-arctic framework consisting of monotonic and partially monotonic shallow networks optimized for the policy and action value function, respectively. By doing so the system achieves better sum-rate, and most importantly a better energy harvesting transmitter.

Deep Reinforcement Learning (DRL) is a promising approach to improving existing NOMA systems. Best of both worlds, it combines artificial neural networks and reinforcement learning to solve really complex, non-linear, non-convex optimization problems. Research works such as [(35), (36), (37)] discuss resource allocation (RA) and channel assignment schemes in different types of NOMA systems by using deep reinforcement learning. (35) considers the joint channel assignment and power allocation problem of a downlink-transmission NOMA system. The authors propose a deep reinforcement learning framework, which utilizes an attention-based neural network (ANN) under maximizing sum rate (MSR) and maximizing minimal rate (MMR) metrics to improve the overall system performance of the NOMA system. Similarly, in (36), a joint channel assignment and power allocation problem is considered in an uplink-transmission NOMA system. A two-step model-free approach is taken by the authors to address this: 1. a deep Q-network (DQN) is designed for the optimal subcarrier assignment 2. A deep deterministic policy gradient (DDPG) network is constructed to dynamically allocate the transmit powers to all the users. This whole network is trained simultaneously with the returned feedback to maximize the long-term sum energy efficiency (EE) by updating the weights of the neural network also while ensuring quality of service (QoS) to all its users. (37) considers a cache-aided NOMA system for power allocation. Caching ensures there is minimal inter-user interference in the superimposed signal, also saving peak power consumption and bandwidth exploitation. The authors propose two novel methods for the power allocation problem: 1. divide and conquer method for the resource allocation policy 2. deep reinforcement learning that allows all users to share the bandwidth efficiently formulated based on the mixed-integer programming algorithms. Deep reinforcement learning (DRL) has also seen its application in mobile edge computing (MEC) as discussed in paper (38). Unlike in Yang et als paper (32) where reinforcement learning was used to solve the latency issue in offloading, authors of paper (38) discuss mobile edge computing (MEC) computation offloading scheme using deep reinforcement learning (DRL) in the absence of labelled data and channel models. Furthermore, the authors attempt to solve the problem of sub-channel allocation in MC-NOMA system using DRL algorithms and achieve higher user capacity without the need for complex optimization problems. This is done by solving computation modes i.e., local computing at the user equipment (UE) or edge execution at the mobile edge computing (MEC) server.

• Introduction to DNN Model

In this paper SIC Decoding utilizing a feedforward neural network has been developed for signal decoding at the receiver end. Feed forward network is widely utilized architecture in the realm of artificial neural networks, and is one of the primary building blocks for various machine learning applications. Essentially, a feedforward neural network processes information unidirectionally, progressing from input to output without incorporating feedback loops. The DNN model structure and components, processing mechanism and functioning are discussed in the following section.

• Structure and Components of DNN model

Deep neural network (DNN) architecture involves following primary components:

• Input Layer
• Hidden Layers
• Output Layer

1. Input Layer:

At the start of the feedforward neural network lies the input layer, which receives the raw data representing inputs. Each node in this layer corresponds to a distinct feature, and the values encapsulated within these nodes serve as the initial points for information propagation.

2. Hidden Layers:

Succeeding the input layer are the hidden layers, constituting the neural network's processing powerhouse. These layers are characterized by interconnected nodes, or neurons, which operate on the input data using weighted connections and activation functions. The weights associated with these connections are instrumental in shaping the network's ability to discern patterns and relationships within the data.

The processing mechanism involves nodes in the hidden layer which receive inputs from the preceding layer. The received inputs are multiplied by the respective weights associated with the connections. The results are summed, and an activation function is applied to introduce non-linearity. Activation functions are critical to functionality of the feed forward neural networks, as they introduce non-linearity in the model, which is imperative for the network to capture and represent complex relationships within the data. Common activation functions include the sigmoid, hyperbolic tangent (tanh), and rectified linear unit (ReLU).

The output of this process becomes the input for the subsequent layer. One important aspect to deep neural networks is in designing the optimal architecture. It involves striking a delicate balance between model complexity and computational efficiency. Determining the appropriate number of layers and nodes in each layer is a crucial decision that impacts the network's capacity and computational demands.

3. Output Layer:

The last component of a DNN network is the output layer, which produces the final processed information. The nodes in this layer generate the network's prediction or classification based on the patterns and relationships learned during the training phase.

• Training of the DNN Model

A neural networks prediction accuracy is very much dependent on training of the model. Training involves utilizing a set of data similar to the transmitted signals and adjustment of the prediction utilizing varying weight and activation functions. This involves backpropagation to optimize the errors. Loss function is utilized to measure the error in the predictions. As a general rule, in neural networks, greater the training data and efforts, better is the prediction accuracy.

Backpropagation involves training of a feedforward neural network by iterative adjustment of weights to minimize the disparity between the predicted output and the actual target values. This optimization process is achieved by propagating the error backward through the network. The gradients of the error with respect to the weights are computed, and optimization algorithms such as gradient descent are employed to update the weights incrementally. A loss function, also known as the error or objective function, is employed to measure the inconsistency between the predicted values by the model and the actual values in the training dataset. The goal during training is to minimize this discrepancy and increase the accuracy of the model's predictions. The choice of the loss function depends on the nature of the problem the model is designed to solve. For example, Mean Squared Error (MSE) is commonly used for regression tasks, while Cross-Entropy Loss is often employed for classification problems.

In order to design this system, a sufficient amount of data/information adhering to the problem of interest is collected, constituting what is called a training set. This set of information is then fed to a learning algorithm to produce a trained machine that carries out the desired task. This learning is made possible by choosing a set of possible hypothesis class, from which the learning algorithm makes a selection during training. An example of a hypothesis class is given by a neural network architecture with learnable parameters. The performance of the learning algorithm is based on how well this hypothesis class matches the available data.

• DNN Learning Processes

There are three main classes of machine learning: Supervised Learning, Unsupervised Learning and Reinforcement Learning.

In supervised learning, pairs of a given input and desired output belonging to a training set are known, and the main objective is to map this input to the output through the use of a learning algorithm. Supervised learning is the most commonly used machine learning technique and is commonly used in classification and regression models.

In Unsupervised learning, the training set consists of just the input variables and no corresponding output. The goal of this learning technique is to model the distribution of data in order to learn more about the data. Unsupervised learning is popularly used for clustering and anomaly detection.

Reinforcement Learning (RL) is a machine learning technique where an agent learns to make decisions by interacting with an environment to maximize cumulative rewards. Through trial and error, the agent learns which actions yield the maximum rewards, thereby resulting in the most desirable outcomes.

Supervised Learning is the most widely used machine learning technique and is extensively utilized for estimating the channel state information (CSI) of a communication system model and for signal decoding.

• Study DNN Model Description

The DNN block serves as the primary detection component for decoding the received signal. Optimizing weights, biases and hyperparameters of the deep neural network enables decoding of the received signal. This involves designing various parameters of the DNN model such as the number of layers, activation function and loss function.

• DNN Design

The DNN model designed for SISO-NOMA detection is primarily a feed-forward neural network. We have designed two neural network models for different modulation schemes, BPSK and QPSK.

For BPSK Modulation, a fully connected neural network is designed that consists of layers: one input layer, one output layer and hidden layers.

Figure:BPSK Modulation Based Neural Network

For QPSK Modulation, a fully connected neural network is designed with layers: one input layer, one output layer and hidden layers.

Figure:QPSK Modulation Based Neural Network

In both models, the input layer consists of two nodes which accepts the real and imaginary components of the received superimposed NOMA signal. This signal is sent to the network in one slot as a column vector. The hidden layers which are fully connected consists of nodes and nodes each for a BPSK modulation and QPSK modulation respectively. ReLU, which stands for Rectified Linear Unit function is used as an activation function in both models in the hidden layers to compute and avoid the vanishing gradient problem. The output layer of the neural network model consists of and nodes as it corresponds to the modulation order of in the case of BPSK modulation and modulation order of in the case of QPSK modulation respectively. Furthermore, in both models we have implemented the sigmoid activation function.

A softmax cross-entropy function is used in both models to minimize the cost-function in the neural networks. A softmax cross-entropy function is a combination of softmax activation and cross-entropy loss. It converts the model outputs into probability distributions and measures the difference between predicted and true distributions, commonly used in classification problems to optimize models for accurate probability estimation. Furthermore, the Adam optimization is used in both models to optimize and fine-tune the weights and biases of the neural network. Adam maintains adaptive learning rates for each parameter by computing two moving averages of the gradients: the first moment (mean) and the second moment (variance). These moving averages are calculated using exponential decay, giving more weight to recent gradients and thereby improving to minimize the cost function.

For BPSK modulation, we have assumed a learning rate of 0.0005. For QPSK modulation, we have assumed a learning rate of 0.0001. The above approaches have been selected primarily because of their abilities to reach faster convergence speeds especially in such complex classification problems.

The feed-forward neural network was developed using Tensorflow, an open-source machine learning framework developed by Google as part of Google Colab. MATLAB R2019b was utilized to develop the conventional SIC decoding algorithm.

• Studied Model - Training Stage

The DNN based NOMA detection block developed passes through two stages: training stage and testing stage.

In the case of BPSK modulation, the training stage comprises of a single iteration. This iteration consists of a batch-size of data samples going over epochs through different SNR values to generate a total of training samples.

Similarly in the case of QPSK modulation, the training stage goes through 4 iterations. Each iteration consists of a batch-size of data samples going over epochs through different SNR values to generate a total of training samples each iteration. This cumulates to total training samples.

Being a supervised learning problem, the input to the DNN detection block which is the superimposed signal impaired with channel fading effects and noise is associated with a label which is the intended users original data. The model learns to map inputs to outputs based on example pairs of input-output data. The labels here represent the correct outputs that the model aims to predict.

• Studied Model - Testing Stage

Upon completion of DNN training, the testing mode is activated. The testing block is employed to simulate real-time NOMA transmission. Here, the received signal is generated without the need for labels. Importantly, to prevent overfitting, the generated data in the training and testing blocks are independent and identically distributed (i.i.d.), ensuring the DNN's effectiveness in both phases.

For BPSK modulation based neural network, we have assumed a sample size of test samples.

Similarly, for the QPSK modulation based DNN we have assumed a sample set of test samples.

The system performance is assessed during this phase and compared with the conventional SIC based decoding algorithm, and the simulation results are discussed in the following results section.

• DNN Model Summary

This code was developed utilizing the following software and tools Google Colab and MATLAB R2019b. TensorFlow, an open-source machine learning framework developed by Google is employed to develop the deep neural network to decode the received superimposed signal from the base station (BS). MATLAB was used to develop the conventional SIC decoding algorithm.

As part of this study, a time-invariant (TIV) channel with complex distribution of Rayleigh fading coefficients is considered to assess the robustness of the DNN based decoding model.

The key parameters of the neural network model are summarized below:

Parameter	Value
Operating System	Windows 11
Framework	TensorFlow
Programming Language	Python, MATLAB
Channel	Rayleigh Fading (TIV), AWGN
Number of users	2
Modulation	BPSK, QPSK
Number of training samples	order of tens of millions
Number of testing samples	order of millions
Activation Functions	ReLU, Sigmoid

Table 5.1 DNN Model Summary

Chapter6

RESULTS PRESENTATION AND PERFORMANCE EVALUATION

• Introduction

This section presents the results of the simulations performed for signal decoding at the receiver end utilizing DNN based decoding models. The results of this simulation are compared with the conventional SIC decoding based algorithm.

As discussed in section , a two-user SISO NOMA system in a downlink transmission is considered. It is considered that the channel parameters are known between the transmitter and receiver. The power allocation levels of the users: User and User, are calculated based on their channel conditions with the total power constraint being assumed as unity. Useris considered farthest away from the base station (BS) and is assigned a higher power level compared to Userwhich is assigned a lower power level.

• Results Presentation

The signal decoding results are extracted as bit-error rate (BER) versus signal-to-noise ratio (SNR) curves. BER indicates the error ratio in the decoded signal and provides an indication of the strength of the forward error correction code. While SNR represents the ratio of signal power to noise in decibels. Thus, the generated BER Vs SNR curve helps in demonstrating the performance of the proposed DNN based signal decoding block vis-a-vis the conventional SIC based decoding algorithm. For the same SNR, lower bit-error rate represents a better performing system.

As part of this study and analysis, a bit-error rate (BER) vs signal-to-noise ratio (SNR) curves are compared for the two different detection blocks.

To assess the robustness of the proposed DNN based signal decoding scheme, signal decoding for two different modulation types: BPSK and QPSK have also been studied. Results of this assessment are presented below.

• BPSK Modulation Scheme Results

A two-user configuration has been considered as described in the previous sections. Userhas been allocated a higher power allocation of and Userhas been allocated a lower power allocation of .

The BER vs SNR curve for the conventional SIC decoding scheme is generated using MATLAB and is presented in Figure 6.1. The BER vs SNR curve of the proposed DNN model is presented in Figure 6.2.

Assuming the same channel conditions and power coefficients for both decoding approaches, we observe from the above figures that the BER vs SNR for a DNN based decoding block is better than the conventional SIC based decoding block.

To assess the sensitivity of the results for varying power allocation level of each user, the performance of the two decoding approach was ascertained and analyzed. In this case, Userhas been allocated a power allocation of and Userhas been allocated a lower power allocation of .

We observe a similar behavior here too as seen in the previous case, that given the new power allocation levels of Userand User, the BER vs SNR curve of the DNN decoding technique is better than the conventional SIC decoding approach.

In a conventional SIC decoding algorithm, as discussed in previous sections, the user with a better channel condition performs the SIC decoding. In our study, we have considered User2 to perform SIC decoding considering it has better channel gains. Observing both cases, it is evident that the error ratio of User2 is fairly lower using the proposed DNN based decoding scheme as opposed to the conventional SIC decoding approach.

• QPSK Modulation Scheme Results

In this section, we discuss the results considering the QPSK modulation scheme in a downlink NOMA system where Userhas been allocated a higher power allocation of and Userhas been allocated a lower power allocation of .

The BER vs SNR curve for the conventional SIC decoding scheme is generated using MATLAB and is presented in Figure 6.5. The BER vs SNR curve of the proposed DNN model is presented in Figure 6.6.

Chapter 7

CONCLUSIONS

The comparative assessment of the signal decoding performance at the receiver in a SISO NOMA system demonstrates that:

• For a BPSK modulation scheme, performance of a DNN based signal decoder is significantly improved compared with the conventional approach
• For a QPSK modulation scheme, performance of a DNN based signal decoder is similar or better than the conventional approach
• It is observed that DNN based signal decoding is more effective when power allocation between the users is similar. This presents an important opportunity in improving data handling capacity in a power-domain based NOMA (PD-NOMA) system.

Chapter 8

FUTURE WORKS

The findings of this study for a downlink SISO-NOMA system for different modulation techniques: BPSK and QPSK and varying power allocation levels is encouraging as similar and significant improvement over conventional SIC decoding is observed.

These findings encourage us to study similar DNN based signal decoding techniques for more complex multiuser downlink and uplink systems.

REFERENCES

• Simeone, A Very Brief Introduction to Machine Learning With Applications to Communication Systems, IEEE Transactions on Cognitive Communications and Networking, vol. 4, no. 4, pp. 648664, 2018.
• Kannan, H. Kim, and S. Oh, Deep Learning and Information Theory: An Emerging Interface, in Proc. IEEE ISIT, 2018.
• 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. 73317376, 2019.
• J. OShea and J. Hoydis, An Introduction to Machine Learning Communications Systems, arXiv preprint arXiv:1702.00832, 2017.
• Ding, X. Lei, G. K. Karagiannidis, R. Schober, J. Yuan, and V. K. Bhargava, A Survey on Non-Orthogonal Multiple Access for 5G Networks: Research Challenges and Future Trends, IEEE Journal on Selected Areas in Communications, vol. 35, no. 10, pp. 21812195, 2017.
• R. Islam, N. Avazov, O. A. Dobre, and K.-S. Kwak, Power-Domain Non-Orthogonal Multiple Access (NOMA) in 5G Systems: Potentials and Challenges, IEEE Communications Surveys & Tutorials, vol. 19, no. 2, pp. 721742, 2016.
• Kim, N.-I. Kim, W. Lee, and D.-H. Cho, Deep Learning-Aided SCMA, IEEE Communications Letters, vol. 22, no. 4, pp. 720723, 2018.
• K. Hasan, M. Shahjalal, M. M. Islam, M. M. Alam, M. F. Ahmed, and Y. M. Jang, The Role of Deep Learning in NOMA for 5G and Beyond Communications, in 2020 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), 2020, pp. 303307.
• Gui, H. Huang, Y. Song, and H. Sari, Deep Learning for an Effective Nonorthogonal Multiple Access Scheme, IEEE Transactions on Vehicular Technology, vol. 67, no. 9, pp. 84408450, 2018.
• Thompson et al., Deep Learning for Signal Detection in Non-Orthogonal Multiple Access Wireless Systems, in 2019 UK/China Emerging Technologies (UCET). IEEE, 2019, pp. 14.
• Lin, Q. Chang, and X. Li, A Deep Learning Approach for MIMO-NOMA Downlink Signal Detection, Sensors, vol. 19, no. 11, p. 2526, 2019.
• A. Aref and S. K. Jayaweera, Deep Learning-aided Successive Interference Cancellation for MIMO-NOMA, in GLOBECOM 2020-2020 IEEE Global Communications Conference. IEEE, 2020, pp. 15.
• Xie, K. C. Teh, and A. C. Kot, Deep Learning Based Joint Detection for OFDM-NOMA Scheme, IEEE Communications Letters, 2021.
• Xie, J. Xiao, and X. Peng, Detection Algorithm Based on Deep Learning for the Multi-user MIMO-NOMA System, in 2020 International Conferences on Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data (SmartData) and IEEE Congress on Cybermatics (Cybermatics). IEEE, 2020, pp. 193197.
• AbdelMoniem, S. M. Gasser, M. S. El-Mahallawy, M. W. Fakhr, and A. Soli-man, Enhanced NOMA System using Adaptive Coding and Modulation based on LSTM Neural Network Channel Estimation, Applied Sciences, vol. 9, no. 15, p. 3022, 2019.
• Sun, K. Niu, and C. Dong, Deep Learning based Joint Detection and De-coding of Non-Orthogonal Multiple Access Systems, in 2018 IEEE Globecom Workshops (GC Wkshps). IEEE, 2018, pp. 15.
• Jiang, X. Li, N. Ye, and A. Wang, Deep Learning-Aided Constellation Design for Downlink NOMA, in 2019 15th International Wireless Communications Mobile Computing Conference (IWCMC). IEEE, 2019, pp. 18791883.
• Alberge, Constellation Design with Deep Learning for Downlink Non-Orthogonal Multiple Access, in 2018 IEEE 29th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). IEEE, 2018, pp. 15.
• Han, H. Seo, A. T. Abebe, and C. G. Kang, Deep Learning-Based Multi-User Multi-Dimensional Constellation Design in Code Domain Non-Orthogonal Multiple Access, in 2020 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2020, pp. 16.
• Sun, Y. Wang, J. Jiao, S. Wu, and Q. Zhang, Deep Learning based Long-Term Power Allocation Scheme for NOMA Downlink System in S-IoT, IEEE Access, vol. 7, pp. 86 28886 296, 2019.
• 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. 29953005, 2019.
• Kim, Y. Ahn, and B. Shim, Deep Neural Network-Based Active User Detection for Grant-Free NOMA Systems, IEEE Transactions on Communications, vol. 68, no. 4, pp. 21432155, 2020.
• Abidi, M. Hizem, I. Ahriz, M. Cherif, and R. Bouallegue, Convolutional Neural Networks for Blind Decoding in Sparse Code Multiple Access, in 2019 15th International Wireless Communications & Mobile Computing Conference (IWCMC). IEEE, 2019, pp. 20072012.
• Saetan and S. Thipchaksurat, Power Allocation for Sum Rate Maximization in 5G NOMA System with Imperfect SIC: A Deep Learning Approach, in 2019 4th International Conference on Information Technology (InCIT). IEEE, 2019, pp. 195198.
• -M. Kang, I.-M. Kim, and C.-J. Chun, Deep Learning based MIMO-NOMA with Imperfect SIC Decoding, IEEE Systems Journal, 2019.
• Liu, T. Song, L. Zhang, and G. Gui, Resource Allocation for NOMA based Heterogeneous IoT with Imperfect SIC: A Deep Learning Method, in 2018 IEEE 29th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). IEEE, 2018, pp. 14401446.
• Yang, H. Zhang, K. Long, H.-Y. Hsieh, and J. Liu, Deep Neural Network for Resource Management in NOMA Networks, IEEE Transactions on Vehicular Technology, vol. 69, no. 1, pp. 876886, 2019.
• Luo, J. Tang, D. K. 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. 17 45017460, 2019.
• Tang, J. Luo, J. Ou, X. Zhang, N. Zhao, D. K. C. So, and K.-K. Wong, Decoupling or Learning: Joint Power Splitting and Allocation in MC-NOMA With SWIPT, IEEE Transactions on Communications, 2020.
• Huang, Y. Yang, Z. Ding, H. Wang, H. Sari, and F. Adachi, Deep Learning-Based Sum Data Rate and Energy Efficiency Optimization for MIMO-NOMA Systems, IEEE Transactions on Wireless Communications, vol. 19, no. 8, pp. 53735388, 2020.
• 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. 33773389, 2017.
• Yang, L. Li, W. Liang, H. Zhang, and Z. Ding, Latency Optimization for Multi-User NOMA-MEC Offloading Using Reinforcement Learning, in 2019 28th Wireless and Optical Communications Conference (WOCC). IEEE, 2019, pp. 15.
• Zhang, L. Li, J. Yin, W. Liang, X. Li, W. Chen, and Z. Han, A Dynamic Power Allocation Scheme in Power-Domain NOMA using Actor-Critic Reinforcement Learning, in 2018 IEEE/CIC International Conference on Communications in China (ICCC). IEEE, 2018, pp. 719723.
• Kim, T. Cho, J. Lee, W. Shin, and H. V. Poor, Optimized Shallow Neural Networks for Sum-Rate Maximization in Energy Harvesting Downlink Multiuser NOMA Systems, IEEE Journal on Selected Areas in Communications, 2020.
• 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. 22002210, 2019.
• Zhang, X. Wang, and Y. Xu, Energy-Efficient Resource Allocation in Uplink NOMA Systems with Deep Reinforcement Learning, in 2019 11th International Conference on Wireless Communications and Signal Processing (WCSP). IEEE, 2019, pp. 16.
• Nguyen Doan, M. Vaezi, W. Shin, H. V. Poor, H. Shin, and T. Q. Quek, Power Allocation in Cache-Aided NOMA Systems: Optimization and Deep Reinforcement Learning Approaches, arXiv preprint arXiv:1909.11074, 2019.
• Nduwayezu, Q.-V. Pham, and W.-J. Hwang, Online Computation Offloading in NOMA-based Multi-Access Edge Computing: A Deep Reinforcement Learning Approach, IEEE Access, vol. 8, pp. 9909899 109, 2020.
• M. El-Sayed, A. S. Ibrahim, and M. M. Khairy, Power Allocation Strategies for Non-Orthogonal Multiple Access, in 2016 International Conference on Selected Topics in Mobile & Wireless Networking (MoWNeT). IEEE, 2016, pp. 16.
• Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge university press, 2005.
• Vaezi, Z. Ding, and H. V. Poor, Multiple Access Techniques for 5G Wireless Networks and Beyond. Springer, 2019.
• Vaezi, R. Schober, Z. Ding, and H. V. Poor, Non-Orthogonal Multiple Access: Common Myths and Critical Questions, IEEE Wireless Communications, vol. 26, no. 5, pp. 174180, 2019.
• R. Usman, A. Khan, M. A. Usman, Y. S. Jang, and S. Y. Shin, On the Performance of Perfect and Imperfect SIC in Downlink Non Orthogonal Multiple Access (NOMA), in 2016 International Conference on Smart Green Technology in Electrical and Information Systems (ICSGTEIS). IEEE, 2016, pp. 102106.
• M. Cover, Elements of Information Theory. John Wiley & Sons, 1999.
• El Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2011.
• Series, Minimum Requirements Related to Technical Performance for IMT2020 Radio Interface (s), Report, pp. 24100, 2017.
• Sari, A. Maatouk, E. Caliskan, M. Assaad, M. Koca, and G. Gui, On the Foundation of NOMA and its Application to 5G Cellular Networks, in 2018 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2018, pp. 16.
• Deep Learning Based Successive Interference Cancellation for the Non-Orthogonal Downlink, Thien Van Luong, Nir Shlezinger, Member, IEEE, Chao Xu, Senior Member, IEEE, Tiep M. Hoang, Member, IEEE, Yonina C. Eldar, Fellow, IEEE, and Lajos Hanzo, Life Fellow, IEEE
• Verdu, Sergio.Multiuser detection. Cambridge university press, 1998.
• Fundamentals of Data Communication Network, Oliver C Ibe, Wiley Publication.
• Multiple Access Scheme for Future (4G) Communication:A Comparison Survey, International Symposium on Devices MEMS, Intelligent Systems & Communication (ISDMISC) 2011, Proceedings published by International Journal of Computer Applications (IJCA), Aasheesh Shukla,Member IEEE,
• Wireless Network Access Technologies, Vijay K. Garg, Yih-Chen Wang, in The Electrical Engineering Handbook, 2005
• FDMA, TDMA, and CDMA Multiple Access: Effective Utilization of Signals , (Bandwidth) in Wireless Communication, Murata Articles.
• Non Orthogonal Mutiple Access, ACCS Journal Quarterly Vol3 Issue 2,

APPENDIXA

MATLAB Code SIC Decoding

APPENDIXB

TensorFlow Code DNN Decoding