Deep Learning for THz Drones with Flying Intelligent Surfaces: Beam and Handoff Prediction
Nof Abuzainab, Muhammad Alrabeiah, Ahmed Alkhateeb, Yalin E. Sagduyu
DDeep Learning for THz Drones with FlyingIntelligent Surfaces: Beam and Handoff Prediction
Nof Abuzainab , Muhammad Alrabeiah , Ahmed Alkhateeb , and Yalin E. Sagduyu Intelligent Automation, Inc. , email: { nabuzainab, ysagduyu } @i-a-i.com Arizona State University , emails: { malrabei, alkhateeb } @asu.edu Abstract —We consider the problem of proactive handoff andbeam selection in Terahertz (THz) drone communication net-works assisted with reconfigurable intelligent surfaces (RIS).Drones have emerged as critical assets for next-generation wire-less networks to provide seamless connectivity and extend thecoverage, and can largely benefit from operating in the THzband to achieve high data rates (such as considered for 6G).However, THz communications are highly susceptible to channelimpairments and blockage effects that become extra challengingwhen accounting for drone mobility. RISs offer flexibility toextend coverage by adapting to channel dynamics. To integrateRISs into THz drone communications, we propose a novel deeplearning solution based on a recurrent neural network, namelythe Gated Recurrent Unit (GRU), that proactively predictsthe serving base station/RIS and the serving beam for eachdrone based on the prior observations of drone location/beamtrajectories. This solution has the potential to extend the coverageof drones and enhance the reliability of next-generation wirelesscommunications. Predicting future beams based on the dronebeam/position trajectory significantly reduces the beam trainingoverhead and its associated latency, and thus emerges as aviable solution to serve time-critical applications. Numericalresults based on realistic 3D ray-tracing simulations show thatthe proposed deep learning solution is promising for futureRIS-assisted THz networks by achieving near-optimal proactivehand-off performance and more than 90 % accuracy for beamprediction. I. I
NTRODUCTION
With the unprecedented increase in the number of de-vices (e.g., IoT devices) and new applications (e.g., vir-tual/augmented reality) that require wireless connectivity [1],there is a critical need to redesign wireless networks thatcan support the ever-growing demand to improve the com-munication rates. Due to the abundance of bandwidth at theTerahertz (THz) band (ranging between 0.1 THz and 10THz) [2], [3], THz communication is envisioned to meetthe rate demands of next-generation wireless communicationnetworks (in particular, 6G [4]) by pushing the rates fromthe Gigabits per second (Gbps) to the Terabits per second(Tbps). One potential application of THz communications isin airborne networks, where unmanned aerial vehicles (UAVs)or so called drones form a network to extend the connectivityand coverage for communication and surveillance needs, e.g.,collectively transferring the overhead imagery that they take toa ground base station. However, there are several challengesto overcome before realizing the anticipated benefits of THz
This paper is based upon work funded by AFRL/SBIR program underAFRL contract No. FA8750-20-C-0505. communications. The THz channel suffers from severe pathloss and molecular absorption, which limits significantly thecommunication distance on the THz band. In addition, dueto the short wavelength of the THz signal, communication ishindered by most of the obstacles present along the commu-nication path, including human users. High network mobilitysuch as due to the mobility of drones aggregates the impactof this challenge. Thus, novel solutions are required that areaware of blockages and extend the range of the THz commu-nication, thereby increasing its feasibility and scale. Amongthese promising solutions are the use of highly directionalantennas, adaptive beamforming, and multi-hop relaying [3].Another promising solution to expand the coverage andrange of THz communication is the use of the reconfigurableintelligent surface (RIS) [5]–[7]. As an emerging technology tomove from massive antennas to antenna surfaces for software-defined wireless systems, the RIS relies on an array of unitcells/elements to control the scattering and reflection profilesof the electromagnetic signals, allowing considerable mitiga-tion of the propagation loss and multipath attenuation. TheRIS is very promising, in particular, for THz communicationthat is highly sensitive to pathloss and blockage. Also, the RIScan be used for effective beamforming of the THz signals as itis still challenging to realize digital beamforming at the THzband.
A. Prior Work
Beam selection has been considered in mmWave systemsusing deep learning techniques [8], [9] and base stationselection has been incorporated in [10]. Deep learning hasbeen also applied to beam selection for the RIS [11]. Therehave been recent works that consider incorporating RISs intoTHz communication networks. The deployment of the RISwas considered for indoor THz communication scenarios in[12]. The application of the holographic RIS in the THz bandwas considered in [13]. Closed-loop and compressed sensingchannel estimation algorithms were proposed to estimate theTHz channels. In [14], communication over the THz bandwas considered for an inter-satellite network to achieve highdata rate transmissions, and the gains of using the RIS onneighboring satellites were assessed in terms of error rate.The use of RISs was considered in [15] and [16] to assist inserving mobile users that are non line-of-sight with the servingbase station. In [15], cooperative beamtraining techniques wereproposed to estimate the channel with the RIS and low-cost a r X i v : . [ c s . I T ] F e b eamforming algorithms were then proposed based on thedeveloped channel estimation. The problem of optimizing thephase shift of the RIS was considered in [16] to serve a set ofusers that are non line-of-sight (NLOS) with the serving basestation, communicating over the THz band.However, none of these works considered the mobility ofusers or RISs for THz communications. Mobility increases thechallenges for THz communications. This is due to the factthat THz communications is extremely sensitive to mobilitydue the high directionality of THz links. Furthermore, dueto the ultra-high data rates of the THz links, any disruptionin communications will result in queue overflow and severedata loss, which will consequently lead to significant servicedegradation. Thus, novel proactive solutions that maintain thecommunication links are essential to realize reliable commu-nications in the THz band. B. Contribution
To tackle the aforementioned challenges, we consider a THzdrone network in which a mobile drone user is served by abase station and a flying RIS. We address two main problemsin this scenario: (i) prediction of the optimal beamformingvector for the base station and the RIS to serve the mobiledrone, and (ii) drone proactive hand-off from the base stationto the RIS and vice versa. As deep learning has been shownto effectively learn from and adapt to spectrum data [17], weconsider a deep neural network solution for beam predictionand proactive hand-off. We summarize our contribution in thefollowing points: • We propose an RIS-assisted drone THz-communicationnetwork, in which a mobile drone maintains a consistentcommunication link with the base station either througha direct link or an RIS-assisted link. • We formulate a deep learning algorithm to realize theconsistent communication link between a base stationand a drone. The algorithm proactively predicts the bestcommunication link (direct or RIS-assisted) and predictsthe best beamforming vector for that link. As we accountfor drone mobility, we consider a recurrent neural net-work solution based on Gated Recurrent Units (GRUs)to capture temporal correlations in the spectrum data andlearn the sequence dependency. • We have built a dataset to study and evaluate the pro-posed algorithm. First, We show that the proposed deeplearning algorithm achieves high accuracy in determiningthe optimal communication link as well as the servingbeams. Then, we show that a minor form of beamtraining can further increase the accuracy of the proposeddeep learning algorithm reaching up to when top-3accuracy is considered for beam prediction as opposed totop-1 accuracy.The rest of the paper is organized as follows. Section IIpresents the system and the channel model. Section III for-mulates the problem. Section IV presents the proposed deeplearning solution. Section V describes the dataset for perfor-
Fig. 1: Outdoor drone scenario with the RIS. mance evaluation. Section VI presents the numerical results.Section VII concludes the paper.II. S
YSTEM AND C HANNEL MODEL
The THz system adopted in this paper is depicted inFigure 1. It has a single base station equipped with M -element antenna array that serves a mobile drone user with abeamforming vector f ∈ C M × . The beamforming vectors areselected from a predefined beam codebook F of size M CB .If the LOS link between the base station and drone is lost,the base station serves the drone through a flying intelligentsurface, equipped with N antennas. Let h T,k and h R,k denotethe uplink channel matrices of the transmitter and receiverto the intelligent surface, and let h TT,k and h TR,k denote thedownlink channel matrices of the transmitter and receiver tothe intelligent surface. The received signal strength at the k -thsubcarrier is expressed as y k = h TR,k Ψh T,k s + v (1) = ( h TR,k (cid:12) h T,k ) T ψs + v, (2)where s ∈ C is a data symbol satisfying E (cid:2) | s | (cid:3) = P , P is the total transmit power, v ∼ N C (0 , σ ) , and Ψ ∈ C I × I isthe RIS interaction matrix and represents the interaction of theRIS with the incident signal from the transmitter. Note that (2)is a result of the diagonal structure of the interaction matrix Ψ , i.e., Ψ = diag ( ψ ) where ψ is the diagonal vector. Thisdiagonal structure results from the operation of the RIS whereevery element i , i = 1 , , ..., I reflects the incident signal aftermultiplying it with an interaction factor [ ψ ] i . The interactionfactor [ ψ ] i is given by [ ψ ] i = e jφ i considering that the RISelements are implemented using phase shifters only. Further,we will call the interaction vector in this case the reflectionbeamforming vector. The reflection beamforming vector ψ isselected from a reflection beamforming codebook P . Channel Model:
We adopt a wideband geometric THzchannel model [3] with L clusters. Each cluster (cid:96) , (cid:96) = { , ..., L } is assumed to contribute with one ray that has atime delay τ (cid:96) ∈ R , azimuth/elevation angles of arrival (AoA)represented by ( θ (cid:96) , φ (cid:96) ) , and complex path gain α (cid:96) (whichincludes the path loss). Further, let p rc ( τ ) represent a pulseshaping function for T-spaced signaling evaluated at seconds. RU x r
Beam Sequence
GRU GRUGRU
Standardization
Location Sequence
Concatenate z
With this model, the delay-d channel between the user and thebase station follows h d = L (cid:88) (cid:96) =1 ρα l p rc ( dT s − τ (cid:96) ) a rv ( θ (cid:96) , φ (cid:96) ) , (3)where a rv ( θ (cid:96) , φ (cid:96) ) is the array response vector of the basestation at the AoAs ( θ (cid:96) , φ (cid:96) ) . Given the delay-d channel in (3),the frequency domain channel vector at subcarrier k , h k , canbe written as h k = D − (cid:88) d =0 h d e − j πkK d . (4)Considering a block-fading channel model, { h n,k } Kk =1 areassumed to stay constant over the channel coherence time,denoted by T C [18].III. P ROBLEM D EFINITION AND F ORMULATION
The objective is to predict the best communication link andits serving beam. To formulate the two problems, we definethe following. • Beam Sequence:
Due to the mobility of the user, theserving base station frequently updates its beam f everytime instant corresponding to beam coherence time. Thisbeam coherence time depends on many factors includingthe speed of the user and the number of antennas at thebase station. To account for the beam coherence time,we define f ( t ) as the beam used by the base station toserve the mobile drone at beam coherence time t , where t = 1 , , ... and with t = 1 representing the first beamcoherence time at which the drone is first connected to thebase station. Based on this, we define a t -step sequenceof beams as B t = { f (1) , f (2) , ..., f ( t ) } . (5) • Position Sequence:
Let x ( t ) denote the position at timestep t (i.e., when beam f ( t ) was selected). Then, we definea t-step sequence of positions as X t = { x (1) , x (2) , ..., x ( t ) } . (6) • LIS Index Sequence:
Let ψ ( t ) denote the reflectionbeamforming vector selected at time step t . We definea t-step sequence of RIS beams as L t = { ψ (1) , ψ (2) , ..., ψ ( t ) } . (7) • Communication-link Sequence:
Let b ( t ) denote the in-dicator of whether the base station has a direct link tothe mobile drone at time t , or not. We define a t-stepsequence of communication links as W t = { b (1) , b (2) , ..., b ( t ) } . (8)Using the above definitions, we formulate our problemas predicting the communication link and the serving beamat time instance t + 1 given the sequence of beams ( L t and B t ) and positions ( X t ). Formally, we design a machinelearning algorithm to learn the mapping {B t , L t , W t } → (cid:8) b ( t +1) , f ( t +1) , ψ ( t +1) (cid:9) . We show that we can train a deepneural network to achieves high accuracy in predicting theoptimal communication link and serving beam.IV. P
ROPOSED D EEP L EARNING B ASED S OLUTION
As the deep learning solution, we propose a recurrent neuralnetwork based on GRUs. Such networks have proven to beeffective in learning sequence dependency, especially longsequences [19]. The proposed architecture is shown in Figure2. It has an input preparation stage that embeds the dualmodality inputs, i.e., beam and location sequences, and thenconcatenates them to form a sequence of high dimensionalvectors. This sequence is fed to a multi-layer GRU network.The network learns the sequential relation between the inputsand outputs a summary feature vector z , which is passed to aclassifier layer. This classifier is customized to predict eitherthe communication-link or the best beam. In both cases, theclassifier implements a fully-connected layer followed by asoftmax layer.To ensure efficient training of the deep learning network,the inputs of the network go through a preparation stage,in which the inputs are standardized to have zero mean andunit variance [20]. For the beam indices, an embedding layeris used to transform each beam value to an a -dimensional lying RISBasestationDrone grid x y NLOSLOS
Fig. 3: Outdoor drone-based scenario.
Left is a perspective view, and right is a top-view.
Gaussian vector, drawn from a Gaussian distribution of zeromean and unit variance Note a is the length of the embeddingvector and it is a design parameter. In the simulation results,we use a = 50 . On the other hand, the locations are stan-dardized by subtracting the mean (x,y,z) vector and dividingeach dimension by its standard deviation. These statistics arecomputed from the training dataset, which will be describedin Section V. Then, the standardized location and embeddedvector of the beam index are concatenated forming a new (3 + a ) -dimensional vector.V. D ATASET FOR T RAINING AND T ESTING OF THE P ROPOSED D EEP L EARNING S OLUTION
To evaluate the proposed solution, a dataset representing aterahertz system, such as that described in Section II, needsto be developed. We use the DeepMIMO data-generationframework [21] to build the scenario and develop that dataset.The scenario depicted in Figure 3 represents a downtown cityintersection with a THz base station placed 6 meters highand a flying RIS hovering 80 meters above the ground. Italso has a 3D drone grid with its base set 40 meters high.The 3D grid consists of 4 parallel 2D grids at 4 differentheights. The spacing between the points of the grid is 0.81meter along x- and y-axes and 0.8 meter along the z-axis. Theintersection is surrounded by several buildings with same basearea and different heights. In addition, each drone moves inthese trajectories according to the random waypoint model.More details can be found in [22].
TABLE I: Parameters for data generation.
Parameter BS RISActive BS 1 2Active user first 1 1Active user last 496 496Number of antennas (x,y,z) (64,1,1) (256,1,1)Antenna spacing 0.5 0.5Center Frequency 200 GHz 200GHzBandwidth 1 GHz 1 GHzNumber of OFDM subcarriers 512 512OFDM sampling factor 1 1OFDM limit 1 1Number of paths 1 1
Using the DeepMIMO generation scripts, a seed dataset isgenerated and processed to construct the development dataset. The seed dataset has all the channels between every droneposition in the 3D grid and both the RIS and the base stationas well as the channel between the RIS and the base station.The DeepMIMO data generation parameters are listed in TableI. With the channels in the seed dataset, a development datasetis built. It consists of 10-step sequences representing differentdrone trajectories. Each step in a sequence has a tuple ofdrone information, i.e., LOS status with the base station, LOSstatus with the RIS, best beamforming vector used by thebase station, best beamforming vector used by the RIS, andthe (x,y,z)-coordinates of the drone. A drone trajectory isformed by moving 10 steps in the 3D drone grid. Each stepis taken randomly along one dimension and it is governedby the following probabilities: 0.2 along y-axis, 0.2 alongz-axis, and finally 0.8 along +x-axis (a drone never movesbackwards on x-axis). The development dataset has a littleover 160 thousand sequences (data samples) of 10-step length.This data is shuffled and split − to form the trainingand validation datasets.VI. N UMERICAL R ESULTS
The performance of the proposed model is evaluated withthe development dataset described above. The next two subsec-tions discuss the network architecture, its training procedure,and finally its evaluation results.
A. Network Architecture and Training
The proposed deep learning algorithm in Figure 2 is exper-imentally designed to achieve a good generalization perfor-mance. The hyperparameters of the deep learning network issummarized in Table II.
TABLE II: Deep learning network hyperparameters.
Hyperparameter ValueNumber of GRU layers 2GRU unit dimension 20Percentage of dropout 20%Beam embedding space dimension 50Classifier layer dimension 256Training algorithm AdamNumber of training epochs 100
Through a sequence of experiments, the best performingarchitecture is found to have 2 layers of GRU units separatedith a dropout layer. Each GRU unit has a hidden statewith 20 dimensions and sees a sequence of length 7 (a dronetrajectory). The beam embedding space has a dimensionalityof a = 50 . The output of the last GRU layer (at the 7-th step) isfed to classifier layer with 256 dimensions (number of classes).This is the result of max {|F| , |P|} . The architecture is trainedusing Adam optimizer with a learning rate of × − andfor 100 epochs. Iteration A cc u r a cy ( % ) Training top-1(Beam)Training top-3 (Beam)Training top-5 (Beam)Validation top-1 (Beam)Validation top-3 (Beam)Validation top-5 (Beam)Training (BS-LIS)Validation (BS-LIS)
Fig. 4: Prediction accuracy of the proposed network architectureversus training iteration. The figure depicts the performance on bothtasks: beam and communication-link predictions.
B. Performance Evaluation
The proposed architecture has two tasks: (i) predicting thefuture beam, i.e., beam to be used at the 8-th step, and (ii)predicting communication link, i.e., direct connection to thebase station or RIS-assisted connection with the base station.Figure 4 shows the training and validation performance ofthe architecture for both tasks. For beam prediction, this isquantified by plotting the top-1, top-3, and top-5 accuracies[23] versus training iterations while for connection type, it isquantified using only top-1 since the prediction is binary. Forbeam prediction, top-1 accuracy shows that the architecture islikely to identify the correct beam with the accuracy of ∼ .This accuracy could be further improved with a little beamtraining by considering the top-3 and 5 beams predicted by thearchitecture. For instance, beam-training the top-3 predictionsof the architecture provides a increase on top-1 predictionaccuracy (from ∼ to ∼ ). This beam prediction per-formance is accompanied with a better prediction performancefor the communication link. The same figure shows that thearchitecture achieves a near-perfect prediction for this task,hitting ∼ . accuracy. This performance is not surprising,though, due to the nature of the link blockage; the drones onlyexperience link blockage with the base station when they arein the dark region illustrated in the top-view in Figure 3. Thisregion is a result of the corner building blocking the dronelinks to the base station. A deeper look at the performance of the architecture revealstwo important conclusions: (i) the architecture has an almosteven performance across all beams in F and P , and (ii)it could achieve reliable “drone handoff” performance. Theformer is demonstrated in Figure 6 where the average predictedvalue and its standard error are plotted against the groundtruthbeam index for both the base station and RIS codebooks. Thefigure basically shows that for almost any choices of activebeams (beams that are used in the scenario) at the base stationor the RIS, the architecture on average predicts the right beamwith almost zero standard error, e.g., whenever beam 50 is thecorrect choice for the 8-th step at both the base station and theRIS, the architecture predicts 50 with almost zero error. Thefigure also shows that the architecture only struggles whenthe groundtruth beam is close to the edge of the active beamset, i.e., beams 5-10 and beams 78-80 for the RIS and beams28-31 for the base station. The reason for those anomalies(mis-predictions) could be attributed to where these beamsare used; the left and middle images of Figure 5 show thebeam distribution over the drone x-y grid for both the basestation and RIS. They clearly show that those anomalies occurclose to the boundary between the LOS and NLOS regionswith the base station (transition region). Finally, the rightimage in Figure 5 demonstrates the second conclusion; notonly the beam prediction is accurate, the communication-linkprediction is also as good, which is a testament to how reliabledrone hand-off is. VII. C ONCLUSION
In this paper, we considered the problem of proactivehandoff and beam selection in a drone THz network employingRISs. To solve the problem, we proposed a novel deep learningbased solution of GRU-based recurrent neural network thatrelies on the history of mobile user locations as well as theserving beams. Our results showed that the proposed deeplearning solution yields high accuracy in proactive handoffand beam selection, and that with little beam training, theaccuracy of the deep learning algorithm can be further im-proved. The increase in accuracy is up to when top-3predictions is considered for beam prediction as comparedto top-1 predictions. The improvement in the performanceof beam prediction is also accompanied by an improvementof the communication-link prediction and consequently animprovement of the reliability in the drone hand-off. Theresults also show that the use of RISs is promising forproactive handoff and beam selection in a drone THz network.R
EFERENCES[1] D. You, B. Seo, E. Jeong and D. H. Kim, “Internet of Things (IoT)for Seamless Virtual Reality Space: Challenges and Perspectives,”
IEEEAccess , vol. 6, pp. 40439–40449, 2018.[2] I. F. Akyildiz, J. M. Jornet, and C. Han, “THz Band: Next Frontier forWireless Communications,” Physical Commun. J., vol. 12, pp. 16—32,Sep. 2014.[3] T. S. Rappaport, Y. Xing, O. Kanhere, S. Ju, A. Madanayake, S. Mandal,A. Alkhateeb, and G. C. Trichopoulos, “Wireless Communications andApplications Above 100 GHz: Opportunities and Challenges for 6G andBeyond”,
IEEE Access , vol. 7, pp. 78729–78757, 2019. ig. 5: Beam distribution over the x-y plane of the drone grid.
Left image shows groundtruth beams used by the RIS.
Center image showsgroundtruth beams used by the base station.
Right image shows the predicted beams of both RIS and base station as given by the proposednetwork. All beams are obtained from the validation set. A v e r age and s t anda r d de v i a t i on PredictionGroundtruth
Beam index -1001020304050607080 A v e r age and s t anda r d de v i a t i on PredictionGroundtruth
BasestationRIS
0 10 20 30 40 50 60 70 80
Fig. 6: Per-beam prediction statistics (average and standard error) forboth the base station and RIS codebooks. [4] W. Saad, M. Bennis, and M. Chen, “A Vision of 6G Wireless Systems:Applications, Trends, Technologies, and Open Research Problems,”
IEEE Network , vol. 34, no. 3, pp. 134–142, May/June 2020.[5] S. Hu, F. Rusek, and O. Edfors, “Beyond Massive MIMO: The Potentialof Data Transmission with Large Intelligent Surfaces,”
IEEE Transac-tions on Signal Processing , vol. 66, no. 10, pp. 2746–2758, May 2018.[6] M. Jung,W. Saad, Y. Jang, G. Kong, and S. Choi, “Performance Analysisof Large Intelligent Surfaces (LISs): Asymptotic Data Rate and ChannelHardening Effects,” arXiv preprint arXiv:1810.05667, 2018.[7] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C.Yuen, “Reconfigurable Intelligent Surfaces for Energy Efficiency inWireless Communication,”
IEEE Transactions on Wireless Communi-cations , vol. 18, no. 8, pp. 4157–4170, Aug. 2019.[8] T. S. Cousik, V. K. Shah, J. H. Reed, T. Erpek, and Yalin E. Sagduyu,”Fast Initial Access with Deep Learning for Beam Prediction in 5GmmWave Networks,” arXiv preprint arXiv:2006.12653, 2020.[9] T. S. Cousik, V. K. Shah, J. H. Reed, T. Erpek, and Yalin E. Sagduyu,“Deep Learning for Fast and Reliable Initial Access in AI-Driven 6GmmWave Networks,” arXiv preprint arXiv:2101.01847, 2021.[10] A. Alkhateeb, I. Beltagy, and S. Alex, “Machine Learning for ReliablemmWave Systems: Blockage Prediction and Proactive Handoff,”
IEEEGlobal Conference on Signal and Information Processing (GlobalSIP) ,Anaheim, CA, 2018.[11] A. Taha, M. Alrabeiah and A. Alkhateeb, “Deep Learning for LargeIntelligent Surfaces in Millimeter Wave and Massive MIMO Systems,”
IEEE Global Communications Conference (GLOBECOM) , Waikoloa,HI, USA, 2019. [12] X. Ma, Z. Chen, W. Chen, Y. Chi, Z. Li, C. Han, and Q. Wen, “In-telligent Reflecting Surface Enhanced Indoor Terahertz CommunicationSystems,”
Nano Communication Networks , vol. 24, May 2020.[13] Z. Wan, Z. Gao, M. Di Renzo, M. Alouini, “Terahertz Massive MIMOwith Holographic Reconfigurable Intelligent Surfaces,” arXiv preprint arXiv:2009.10963, 2020.[14] K. Tekbıyık, G. Karabulut Kurt, A. R. Ekti, A. G¨orc¸in, H.Yanikomeroglu, “Reconfigurable Intelligent Surface Empowered Ter-ahertz Communication for LEO Satellite Networks,” arXiv preprint arXiv:2007.04281, 2020.[15] B. Ning, Z. Chen, W. Chen, Y. Du, J. Fang, “Terahertz Multi-UserMassive MIMO with Intelligent Reflecting Surface: Beam Training andHybrid Beamforming,” arXiv preprint arXiv:1912.11662, 2020.[16] A. Tarable, F. Malandrino, L. Dossi, R. Nebuloni, G. Virone and A.Nordio, ”Meta Surface Optimization in 6G Sub-THz Communications,”
IEEE International Conference on Communications Workshops (ICCWorkshops) , Dublin, Ireland, 2020.[17] T. Erpek, T. O’Shea, Y. E. Sagduyu, Y. Shi, and T. C. Clancy, “DeepLearning for Wireless Communications” in Development and Analysisof Deep Learning Architectures , Springer, 2020.[18] V. Va, J. Choi, T. Shimizu, G. Bansal, and R. W. Heath, “InverseMultipath Fingerprinting for Millimeter Wave V2I Beam Alignment,”
IEEE Transactions on Vehicular Technology , vol. 67, no. 5, pp. 4042–4058, May 2018.[19] I. Goodfellow, Y. Bengio, A. Courville, and Y. Bengio. Deep learning.Vol. 1, no. 2. Cambridge: MIT press, 2016.[20] Y. A. LeCun, L. Bottou, G. B. Orr, and K.-R. Muller, “EfficientBackprop.” in
Neural networks: Tricks of the trade , pp. 9-48. Springer,Berlin, Heidelberg, 2012.[21] A. Alkhateeb, “DeepMIMO: A Generic Deep Learning Dataset forMillimeter Wave and Massive MIMO Applications,” arXiv preprint