DeepMuD: Multi-user Detection for Uplink Grant-Free NOMA IoT Networks via Deep Learning
aa r X i v : . [ c s . I T ] F e b DeepMuD: Multi-user Detection for Uplink Grant-FreeNOMA IoT Networks via Deep Learning
Ahmet Emir, Ferdi Kara,
Member, IEEE,
Hakan Kaya, Halim Yanikomeroglu,
Fellow, IEEE.
Abstract —In this letter, we propose a deep learning-aidedmulti-user detection (DeepMuD) in uplink non-orthogonal mul-tiple access (NOMA) to empower the massive machine-type com-munication where an offline-trained Long Short-Term Memory(LSTM)-based network is used for multi-user detection. In theproposed DeepMuD, a perfect channel state information (CSI) isalso not required since it is able to perform a joint channel estima-tion and multi-user detection with the pilot responses, where thepilot-to-frame ratio is very low. The proposed DeepMuD improvesthe error performance of the uplink NOMA significantly andoutperforms the conventional detectors (even with perfect CSI).Moreover, this gain becomes superb with the increase in thenumber of Internet of Things (IoT) devices. Furthermore, theproposed DeepMuD has a flexible detection and regardless ofthe number of IoT devices, the multi-user detection can beperformed. Thus, an arbitrary number of IoT devices can beserved without a signaling overhead, which enables the grant-free communication.
Index Terms —beyond 5G, deep learning, error performance,grant-free communication, multi-user detection, uplink NOMA.
I. I
NTRODUCTION
Non-orthogonal multiple access (NOMA) scheme has agreat potential for massive machine-type communication(mMTC) framework since it allows multiple users/devicesto share the same resource block (RB). However, especiallyin Internet of Things (IoT) networks, the data streams aregenerally sparse and low-rate. Thus, allocating an RB foran IoT device is not efficient owing to the limited RBs andthousands of IoT devices to be served by the same access point(AP)/base-station (BS). Besides, to grant an RB for an IoTdevice costs a signaling overhead. Furthermore, this signalingoverhead should always be repeated since the number of activeIoT devices changes from time to time due to the sparsecommunication, hence; this revokes the spectral efficiencyprovided by the NOMA. To resolve this, the grant-free NOMAschemes in uplink IoT networks have attracted great recentattention where the IoT devices transmit their low-rate dataon the same RB (i.e., NOMA) without permission request [1].The grant-free NOMA has been analyzed in terms of outageprobability and capacity for finite and infinite alphabets [2]–[4].However, the major drawback of the NOMA is its errorperformance due to the inter user interference (IUI). Indeed,this poor performance in uplink NOMA gets worse and anerror floor occurs even in two-user networks [5]. Nevertheless,the error performance of the grant-free NOMA has not been
A. Emir, F. Kara and H. Kaya are with the Electrical-Electronics En-gineering, Zonguldak Bulent Ecevit University, Zonguldak, Turkey, e-mail: { ahmet.emir,f.kara,hakan.kaya } @beun.edu.tr.F. Kara and H. Yanikomeroglu are with the Department of Systems andComputer Engineering, Carleton University, Ottawa, K1S 5B6, ON, Canada,e-mail:[email protected]. studied well although it has been widely analyzed in terms ofinformation-theoretic perspectives. To the best of the authors’knowledge, the bit error rate (BER) performance of the grant-free NOMA schemes has not been evaluated yet.On the other hand, the machine learning algorithms havegained great credibility for the last few years [6] and theirpotentials for wireless communications have been admittedby the community [7]. Therefore, the machine learning algo-rithms, particularly deep learning (DL), have been proposed inNOMA-involved systems for modulation, constellation designand resource allocation [8]–[10]. Besides, the detector designs via DL instead of conventional detectors have also beeninvestigated for basic downlink [12]–[15] and uplink [16],[17] NOMA schemes. However, none of the previous uplinkdesigns [16], [17] considers a grant-free access where thenumber of devices is fixed to only two, which is very lowfor the IoT networks. Furthermore, in [16], three frames arerequired to detect one frame data to acquire the channelinformation of the devices. This causes a spectral efficiencydecay by . and weakens the advantage introduced bythe NOMA. To the best of the authors’ knowledge, there is nostudy in the literature which provides a DL-based multi-userdetection for uplink NOMA nor a spectral efficient practicaldesign for a grant-free NOMA even in two-user network.Motivated by the above discussions, we propose a DL-aidedmulti-user detection for grant-free IoT networks. The maincontributions of this letter are summarized as follows. • We propose a DL-aided multi-user detection (DeepMuD)which is able to detect symbols of an arbitrary numberof devices in grant-free communication. The DeepMuDhas been trained offline and implemented as an onlinedetector in the grant-free IoT NOMA networks. • With the proposed DeepMuD, the error performance ofthe uplink NOMA has been improved significantly. Thesame error performance with the conventional detectorsfor arbitrary number of users has been achieved withat least dB less power consumption which is verypromising for energy-limited IoT devices. • The proposed DeepMuD has a flexible detection structurewhere less than or equal to the number of devices in thetraining can be detected as online without performancedegradation. Thus, the DeepMuD enables the grant-freecommunication without introducing a signaling overhead. In the literature, code-domain NOMA has also been considered for thegrant-free communication. Hence, the multi-user detection and end-to-endoptimization of the code-domain NOMA empowered by DL have beeninvestigated in [11] where multiple RBs are used with preambles rather thanone RB as being in power domain-NOMA. Thus, code-domain NOMA isbeyond the scope of this paper and NOMA refers to power-domain NOMA.
Fig. 1. The frame design of IoT devices in grant-free uplink NOMA for L = 4 N d = 8 , N = L + N d = 12 . • In the proposed DeepMuD, the perfect CSI is also notrequired. The DeepMuD performs a multi-user detectionbased on pilot signals. The used pilot-to-frame ratio isalso very low compared to the existing works. Thus, theadvantage of the NOMA is ensured.The rest of the letter is organized as follows. In Section II,the uplink NOMA scheme is introduced (with the benchmarkdetector) and the signal design for the grant-free uplinkNOMA is given. In Section III, the proposed DeepMuD isgiven in detail. Then, in Section IV, the error performancecomparisons are presented between the proposed DeepMuDand the existing detectors. The results are discussed in detail.Finally, Section V concludes the letter.II. S
YSTEM M ODEL AND S IGNAL D ESIGN FOR G RANT -F REE
NOMAWe consider an uplink scheme, where an AP and L IoTdevices (e.g., sensors within a factory or agriculture area)are located. We assume that all nodes are equipped withsingle antenna and the channel fading between the l th (i.e., l = 1 , , . . . , L ) device and the AP experiences a block-flatfading Rayleigh channel. Since a grant-free communicationis considered, the active devices, which have data to convey,transmit their data to the AP on the same RB simultaneously.Hence, the received signal at the AP is given as y ( t ) = K X i =1 p P i x i ( t ) h i ( t ) + n ( t ) , K ≤ L (1)where K is the number of active devices. P i , x i ( t ) and h i ( t ) are the transmit power, transmitted symbol and channel coeffi-cient of the i th active device, respectively. n ( t ) is the additivewhite Gaussian noise (AWGN) which follows CN (0 , N / ) . A. Successive Interference Canceler-based Detector (SICD)(Benchmark)
In the SICD, the AP detects IoT devices’ symbols in asuccessive manner, where the symbols of the IoT device withthe best channel condition are detected firstly. Then, thesedetected symbols are subtracted from the received signal andthe symbols of the IoT device with the second-best channelcondition are detected. Thereafter, these secondly detectedsymbols are subtracted from the remaining received signal and this goes on until to the last device. Therefore, the detectionprocess is given as ˆ x i ( t ) = argmin j (cid:12)(cid:12)(cid:12) y SIC,i ( t ) − p P i ˆ h i ( t ) x i,j (cid:12)(cid:12)(cid:12) , where j = 1 , , . . . , M i ,y SIC,i ( t ) = y SIC,i − ( t ) − p P i − ˆ h i − ( t )ˆ x i − , where y SIC, ( t ) , y ( t ) , and ˆ x = argmin j (cid:12)(cid:12)(cid:12) y ( t ) − p P ˆ h ( t ) x ,j (cid:12)(cid:12)(cid:12) , j = 1 , , . . . , M i , (2)where ˆ x i is the detected/estimated symbol of the i th device. x i,j is the j th constellation point in M i -ary modulation orderof the i th device. ˆ h i is the estimated (imperfect and/or perfect)channel coefficient of the i th device at the AP, thus; an extrachannel estimation algorithm (e.g., LS, MMSE) should beimplemented. As seen in (2), in the SICD, an additional latencyis introduced since K times maximum likelihood detection and K − times subtraction should be implemented in order. B. Frame/Signal Design with Pilot Insertions
As explained above, in the SICD, an additional channelestimation algorithm is required for the CSI knowledge (sincethe SICD requires the channel order to perform detection). Onthe other hand, thanks to their high capacity in correlating, theDL-based detectors can perform well in joint channel estima-tion and symbol detection. However, the pilot insertion shouldbe handled carefully not to cause an erroneous correlation,particularly in NOMA, since the IoT devices will be servedsimultaneously. To this end, we design frames with pilotsignals for each IoT device. In the frame designs of L devices,we insert pilot signals and zero padding into the data frames.Hence, the total frame representations for the first and second(due to space limitations) IoT devices are given as x = N z }| { x p ( t ) , , ..., , | {z } L − x ,d ( t + L ) , ..., x ,d ( t + N − | {z } N d , x = N z }| { , x p ( t + 1) , , ..., , | {z } L − x ,d ( t + L ) , ..., x ,d ( t + N − | {z } N d , (3)where x pi and x i,d denote the pilot signal and data symbolof the i th IoT device, respectively. The zero padding (e.g., , ..., ) is inserted since during these zero padding of the i thIoT device, the pilot signals for other devices (i.e., x pj , j =1 , , ...L, j = i ) are transferred to the AP. Hence, none ofthe pilot signals is overlapped so that the DeepMuD is able tocorrelate them with data symbols correctly. The overall framedesign for L = 4 is given in Fig. 1. Hereby, we would like tonote that the pilot plus zero padding insertions into the frame During one frame, the channel coefficients are assumed to be constant(e.g., block fading), which is quite reasonable since the IoT communicationis generally performed in short-range communication with low-rate/frame size.
Fig. 2. The proposed DeepMuD network model. could be interleaved (e.g., coordinated random interleaving).Nevertheless, we placed them in the beginning of the framefor the representation simplicity. Since the transmission isassumed to be frame by frame, according to (1), the receivedframe turns out to be y = K X i =1 p P i h i x i + n , (4)where y = [ y ( t ) , y ( t + 1) , ...y ( t + N )] and n = [ n ( t ) , n ( t +1) , ...n ( t + N )] are defined.In (3) and (4), N d and N are the data symbols’ length andthe total frame size, respectively. Therefore, during an N framesize, N d symbols are transmitted for each IoT devices whereas N − N d = L symbols are used for non-data information. Thus,the ensured ergodic capacity is given as C = δB K X i =1 log M i (1 − P i ( e )) , (5)where δ = N d / N is the data-to-frame size ratio and B is thebandwidth. P i ( e ) is the error probability of the i th IoT device.We introduce P i ( e ) in capacity definition since the erroneous-detected symbols are not meaningful, thus we only considercorrect-detected symbols such that calling ensured capacity.III. P ROPOSED D EEP M U DIn the proposed DeepMuD, we use a model-driven DL todetect symbols simultaneously. The proposed model is basedon a Long Short-Term Memory (LSTM) network. The LSTMis chosen since the LSTM networks have good performancesin predicting frame size (time series) data where within theframe, the data is correlated [8]. As explained above, in theconsidered model, we assume that the channel coefficientsdo not change within a frame. Therefore, the symbols areexposed the same channel effect and have correlation. Theproposed model has 4 layers. The first layer is the inputlayer where the inputs are transferred to the further layerswith the trained-weight coefficients. In the next two layers,we use LSTM layers which have and LSTM cells, respectively. The last layer is a fully-connected layer whichproduces the estimated symbols for each IoT device. Hence, ithas L neurons to transfer the estimated symbols to the outputs. There is no theoretical way to find/select the optimum number of layersor cells in each layer [18]. Thus, as being in all DL-aided communicationsapplications [6]–[17], these parameters are empirically determined, such thatincreasing the sizes do not provide a noteworthy gain in learning performanceand the network performance converges.
Besides, we use a regression layer as a feedback in the training.The proposed model is presented in Fig. 2.
A. Dataset Generation
Since the DL models can not perform with complex data,we split the received data into two parts as real and imaginaryparts. Besides, in LSTM networks, the frame sizes of the inputand output should be the same. In the considered system, weseek detecting data symbols with N d length, thus; the outputframe size is N d . Based on these, we reform the received signaland these reformed-inputs ( inputs (real and imaginary) forreceived data symbols and L inputs (real and imaginary) forpilot responses of each device) vectors are given by y Ri,p = Re { N d z }| { [ y ( t + i − , y ( t + i − , ..., y ( t + i − } , y Ii,p = Im { N d z }| { [ y ( t + i − , y ( t + i − , ..., y ( t + i − } , y Rd = Re { N d z }| { [ y ( t + L ) , y ( t + L + 1) , ... . . . , y ( t + N − } , y Id = Im { N d z }| { [ y ( t + L ) , y ( t + L + 1) , ... . . . , y ( t + N − } . (6)In (6), we extend the pilot inputs (i.e., y Ri,p and y Ii,p ) to have N d length by adding copies of the received pilot responses.The dataset generation is given in Algorithm 1, where S is the sample size for each SNR value. Besides, to coverthe effects of different SNR values, in the dataset, we define SNR = [0 : 5 : 30] dB. The outputs of Algorithm 1 arethe X i,d , Y Ri,p , Y Ii,p , Y Rd , Y Id matrices and each rows of themare equal to the obtained vectors at Step 4 and Step 6 ofAlgorithm 1 for each iteration. Algorithm 1
Dataset Generation for Training DeepMuD Inputs ( N, L, S,
SNR ) for each SNR value in SNR for s = 1 : S Generate random N d log M i bits for L IoT devices andmaps them to x i,d by M i -ary modulation. Then accordingto (3) and Fig. 1, design x i frame for each device. Generate random Rayleigh channel coefficients ( h i ) foreach device and generate random AWGN vector n According to (4), calculate the y and based on this y ,obtain the y Ri,p , y Ii,p , y Rd and y Id by using (6). Outputs X i,d , Y Ri,p , Y Ii,p , Y Rd , Y Id TABLE IT
RAINING S ETTINGS AND O PTIMIZED P ARAMETERS
Parameter ValueProgramming MATLAB
SNR (dB) [0 : 5 : 30]
Modulation BPSKFading and Channel Rayleigh+AWGNIoT devices in a RB ( L ) , Frame Size ( N ) , Number of Samples ( S ) per scenario Optimizer ADAMLearning Rate 0.001Mini Batch Size 1000Maximum Epoch 20
B. Model Training
After obtaining the training dataset, the proposed DeepMuDis trained according to DL parameters. In the training process,the outputs of the DeepMuD are expected to be equal to x i,d where i = 1 , , ..., L . Therefore, the training optimizationproblem is given by half mean square error as { P } = min ( L L X i =1 || x i − ˆx i || ) . (7)In addition, by considering the user fairness of IoT devicesand not to cause severe error performance for any of the IoTdevices, we define optimization problem in terms of errorperformance as { P } = min { max { P i ( e ) }} , i = 1 , , ..., L. (8)To this end, we initialize DL parameters and train thenetwork according to P in (7). Then, the trained networkis implemented as a detector in the simulations and the BERperformances of each device are obtained. According to thisBER performances, we update DL parameters and retrain thenetwork unless the P in (8) is satisfied. The training andparameter optimization are given in Algorithm 2. The trainingsettings in Algorithm 1 and Algorithm 2 and the optimizedparameters are given in Table I. Algorithm 2
Training and Optimization of the DeepMuD Inputs ( X i,d , Y Ri,p , Y Ii,p , Y Rd , Y Id ) Initialize DL Parameters (mini batch size, learning rate,maximum epoch) Train the network with the DL parameters according to P in (7) do Update DL training parameters and go to Step 3 while Unless the P in (8) is satisfied Outputs
The DeepMuD network, optimized DL parame-ters
C. Complexity
We focus on the online implementation complexity (i.e.,feed-forward calculations) for the DeepMuD. The computa-tional complexity for an LSTM-based network is given by O ( W ) where W is the weight size in the hidden layer [18].We use two hidden layers, hence the computational complexityof the DeepMuD is obtained as O ( W W ) = O (80 × . On the other hand, the computational complexity of the SICDis given as O (4 M L + 2( L − . Besides, in the SICD, anadditional complexity will be introduced due to the channelestimation algorithm which is not required in the DeepMuD.It is seen that the complexity of the SICD is increased bythe modulation order and the number of IoT devices whereasthe proposed DeepMuD has the same complexity regardlessof these. On the other hand, with the lower M and L , theDeepMuD may have a bit higher complexity. Nevertheless,we should note that the detection is performed at the AP andthe AP has high computational capacity. Besides, consideringthe performance gain in the DeepMuD (see the next section),this negligible complexity increase is affordable.IV. N UMERICAL R ESULTS
In this section, we implement the offline-trained DeepMuDas an online detector and perform experiments on syntheticdataset. The link-level Monte-Carlo simulations are con-ducted to validate the performance comparisons.In Fig. 3, we present performance comparisons betweenDeepMuD and SICD for L = 4 , where BER and capac-ity comparisons are given in Fig. 3.a, 3.b and in Fig. 3.c,respectively. We assume that the number of active users areequal to the number of total users to be served (i.e., K = L ).Besides, in all subfigures, we present the results of DeepMuDfor different frame sizes ( N ). As seen in Fig. 3.a and 3.b,regardless of the frame size ( N ) and/or number of IoT devicesin a RB ( L ), the proposed DeepMuD always outperformsthe SICD significantly even though perfect CSI is assumedin SICD whereas only pilot responses are available for theproposed DeepMuD. Hereby, it is worth noting that the SICDcan not detect any symbols when the number of devices ishigher than . (In Fig. 4, the SICD works for only two-user networks. However, it still has an error floor.) On theother hand, once the DeepMuD results are compared, withthe increase of frame size ( N ), the performance is improvedin K = L = 4 in Fig. 3.a since the correlation between dataincreases so that the network learns better. Nevertheless, dueto the IUI, this gain is smaller when K = L = 6 in Fig. 3.b.Then, to present the ensured capacity as defined in (5),we provide normalized (i.e., B = ensured capacity com-parisons between the DeepMuD and SICD in Fig. 3.c. Sincethe SICD performs symbol-by-symbol detection, it requires K time slots for the CSI knowledge to detect one symbol. Thus, δ in SICD is equal to / ( K + 1) . Nevertheless, we also giveSICD results for the same δ = N d / N as in the proposed model.One can easily see that the proposed DeepMuD is superiorto SICD also in terms of ensured capacity where the SICDcan never achieve the theoretical capacity. This is explainedas follows. In uplink NOMA, it is theoretically possible toimprove the capacity; however, without proper detection, theBER performance is severe. Therefore, the achievable capacityis actually non-detectable capacity. Besides, it is noteworthythat the perfect CSI is assumed in SICD and its performancewill get worse when it is relaxed. On the other hand, the In simulations, all IoT devices have equal transmit power (i.e., ρ = ρ i = P i / N , ∀ i ) and the channel conditions are E [ | h i | ] = E [ | h i +1 | ] + 3 dBand E [ | h K | ] = 0 dB. (dB) -3 -2 -1 B E R st IoT Dev.2 nd IoT Dev.3 rd IoT Dev.4 th IoT Dev.
Dash-Dot Lines: SICDDash Lines: DeepMuD, N=16Dot Lines: DeepMuD, N=64 (a) BER, K = L = 4 (dB) -2 -1 B E R st IoT Dev.2 nd IoT Dev.3 rd IoT Dev.4 th IoT Dev.5 th IoT Dev.6 th IoT Dev.
Dash-Dot Lines: SICDDash Lines: DeepMuD, N=16Dot Lines: DeepMuD, N=64 (b) BER, K = L = 6 (dB) N o r m a li z e d E n s u r e d C a p ac i t y ( bp s / H z ) SICD, =1/KSICD, =(N-L)/(L), N=16SICD, =(N-L)/(L), N=64DeepMuD, N=16DeepMuD, N=64
Dash Lines: K=L=4Dot Lines: K=L=6 (c) Normalized Ensured Capacity, K = L = 4 , Fig. 3. Performance comparisons of DeepMuD and SICD when K = L = 4 , (dB) -3 -2 -1 B E R st IoT Dev., SICD2 nd IoT Dev., SICD1 st IoT Dev., DeepMuD2 nd IoT Dev., DeepMuD (a) K = 2 , L = 4 (dB) -3 -2 -1 B E R st IoT Dev., SICD2 nd IoT Dev., SICD3 rd IoT Dev., SICD4 th IoT Dev., SICD1 st IoT Dev., DeepMuD2 nd IoT Dev., DeepMuD3 rd IoT Dev., DeepMuD4 th IoT Dev., DeepMuD (b) K = 4 , L = 6 Fig. 4. BER comparisons of DeepMuD and SICD. proposed DeepMuD provides an increased BER performancewith low pilot-to-frame size ratio ( − δ ); hence, the achievablecapacity is ensured.Lastly, in order to reveal the flexibility of the proposedDeepMuD, in Fig. 4, we present BER performances for thescenarios in which the number of active IoT devices are equalto K = 2 and K = 4 when the training is performedfor L = 4 and L = 6 , respectively. As seen in Fig. 4,the proposed DeepMuD has performed well even though thenumber of active devices are less than the training numbers andthe DeepMuD outperforms the SICD remarkably. This provesthe power of the DeepMuD for grant-free networks since theDeepMuD can detect symbols for arbitrary number of IoTdevices once it is trained offline. Therefore, the IoT devicesdo not require a grant permission thus reducing the signalingoverhead. Hereby, we note that the minimum performance gainprovided by the DeepMuD is achieved when K = 2 wherethe DeepMuD and SICD have the same BER performanceswith ∼ dB less power consumption in DeepMuD. With thehigher K , the DeepMuD provides much more performancegain and it can be up to dB in some scenarios.V. C ONCLUSION
In this paper, we propose a DL-aided multi-user detection(DeepMuD) in grant-free uplink IoT NOMA networks. Theproposed DeepMuD has a good BER performance and out-performs existing multi-user detection schemes. Besides, withthe proposed DeepMuD, the signal detection can be performedfor arbitrary number of IoT devices (less than and/or equal tothe number in training, i.e., K ≤ L ) which enables the grant-free access. Furthermore, the DeepMuD detects signal basedon pilot responses, therefore; no additional channel estimation algorithm is needed. This reveals the power of DeepMuD injoint signal detection, and it can be applied in other promisingphysical layer schemes such as faster than Nyquist, indexmodulations, and large intelligent surfaces. These are seen asfuture research directions.R EFERENCES[1] M. B. Shahab et al. , “Grant-Free Non-Orthogonal Multiple Access forIoT: A Survey,”
IEEE Commun. Surv. Tutorials , vol. 22, no. 3, pp.1805–1838, 3rd Quart. 2020.[2] Z. Ding et al. , “Simple Semi-Grant-Free Transmission Strategies As-sisted by Non-Orthogonal Multiple Access,”
IEEE Trans. Commun. ,vol. 67, no. 6, pp. 4464–4478, Jun. 2019.[3] R. Abbas et al. , “A novel analytical framework for massive grant-freeNOMA,”
IEEE Trans. Commun. , vol. 67, no. 3, pp. 2436–2449, Mar.2019.[4] C. Zhang et al. , “Semi-Grant-Free NOMA: Ergodic Rates Analysis withRandom Deployed Users,”
IEEE Wirel. Commun. Lett. , Early Access ,2020.[5] F. Kara and H. Kaya, “Error Probability Analysis of NOMA-BasedDiamond Relaying Network,”
IEEE Trans. Veh. Technol. , vol. 69, no. 2,pp. 2280–2285, Feb. 2020.[6] H. Ye et al. , “Power of Deep Learning for Channel Estimation and SignalDetection in OFDM Systems,”
IEEE Wirel. Commun. Lett. , vol. 7, no. 1,pp. 114–117, Feb. 2018.[7] A. Zappone et al. , “Wireless Networks Design in the Era of DeepLearning: Model-Based, AI-Based, or Both?”
IEEE Trans. Commun. ,vol. 67, no. 10, pp. 7331–7376, Oct. 2019.[8] G. Gui et al. , “Deep Learning for an Effective Nonorthogonal MultipleAccess Scheme,”
IEEE Trans. Veh. Technol. , vol. 67, no. 9, pp. 8440–8450, Sep. 2018.[9] J. Zhang et al. , “Deep Reinforcement Learning for Throughput Improve-ment of the Uplink Grant-Free NOMA System,”
IEEE Internet ThingsJ. , vol. 7, no. 7, pp. 6369–6379, Jul 2020.[10] X. Miao et al. , “Grant-Free NOMA With Device Activity LearningUsing Long Short-Term Memory,”
IEEE Wirel. Commun. Lett. , vol. 9,no. 7, pp. 981–984, Jul. 2020.[11] N. Ye et al. , “DeepNOMA: A Unified Framework for NOMA UsingDeep Multi-Task Learning,”
IEEE Trans. Wirel. Commun. , vol. 19, no. 4,pp. 2208–2225, Apr. 2020.[12] N. Zhang et al. , “A Machine-Learning-Based Blind Detection on Inter-ference Modulation Order in NOMA Systems,”
IEEE Commun. Lett. ,vol. 22, no. 12, pp. 2463–2466, Dec. 2018.[13] C. Lin et al. , “A Deep Learning Approach for MIMO-NOMA DownlinkSignal Detection,”
Sensors , vol. 19, no. 11, p. 2526, Jun. 2019.[14] A. Emir et al. , “Deep Learning-Based Joint Symbol Detection forNOMA,” in
IEEE 27th Signal Process. Commun. Appl. Conf. , Sivas,Turkey, Apr. 2019.[15] J. M. Kang et al. , “Deep Learning-Based MIMO-NOMA with ImperfectSIC Decoding,”
IEEE Syst. J. , vol. 14, no. 3, pp. 3414–3417, Sep. 2020.[16] Narengerile and J. Thompson, “Deep Learning for Signal Detection inNon-Orthogonal Multiple Access Wireless Systems,” in
IEEE UK/ ChinaEmerg. Technol. , Aug. 2019.[17] A. Emir et al. , “Deep Learning Detectors with Pilot Interval Reduction inUplink Non Orthogonal Multiple Access,” in
IEEE 28th Signal Process.Commun. Appl. Conf. , Gaziantep, Turkey, Oct. 2020.[18] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”