DNN Based Beam Selection in mmW Heterogeneous Networks
DDNN Based Beam Selection in mmWHeterogeneous Networks (cid:63)
Deepa Jagyasi [0000 − − − and MarceauCoupechoux [0000 − − − X ] LTCI, Telecom Paris, Institut Polytechnique de Paris { deepa.jagyasi,marceau.coupechoux } @telecom-paris.fr Abstract.
We consider a heterogeneous cellular network wherein mul-tiple small cell millimeter wave (mmW) base stations (BSs) coexist withlegacy sub-6GHz macro BSs. In the mmW band, small cells use multiplenarrow beams to ensure sufficient coverage and User Equipments (UEs)have to select the best small cell and the best beam in order to accessthe network. This process usually based on exhaustive search may intro-duce unacceptable latency. In order to address this issue, we rely on thesub-6GHz macro BS support and propose a deep neural network (DNN)architecture that utilizes basic components from the Channel State In-formation (CSI) of sub-6GHz network as input features. The output ofthe DNN is the mmW BS and beam selection that can provide the bestcommunication performance. In the set of features, we avoid using theUE location, which may not be readily available for every device. Weformulate a mmW BS selection and beam selection problem as a classi-fication and regression problem respectively and propose a joint solutionusing a branched neural network. The numerical comparison with theconventional exhaustive search results shows that the proposed designdemonstrate better performance than exhaustive search in terms of la-tency with at least 85% accuracy.
Keywords: millimeter wave · beam selection · deep neural network · heterogeneous network · sub-6GHz. Millimeter Wave (mmW) communication is considered as a promising techniqueto solve the unprecedented challenge of increasing demand for high data ratesin future cellular networks. However, it suffers from limited coverage and in theultra-dense environment it is significantly prone to blockages such as high densityobjects like walls, glass, humans, etc. Thus, in-order to provide flexible coverageand minimize the infrastructural cost, it is proposed that mmW networks willbe deployed in a multi-tier heterogeneous network, where multiple small cellmmW base stations (BSs) coexist with multiple legacy sub-6GHz macro BSs [9]. (cid:63)
Supported by the EXTRANGE4G project with company ETELM funded by DGA.This work has been performed at LINCS laboratory. a r X i v : . [ c s . N I] F e b he legacy network operating in sub-6GHz frequencies can handle operationslike resource allocation, mobile data offloading, control signalling etc., while thepotential mmW BSs can handle massive data traffic [9, 13]. In this paper, wepropose a solution for optimal resource allocation for a heterogeneous cellularnetwork that enables reliable communication while leveraging the benefits ofhigh data rates from mmW bands.Beamforming is important in mmW systems in order to overcome the pathloss due to shorter wavelength. With the large number of antenna elements asso-ciated with mmW transceivers, multiple beams are possible, which can performdirectional beamforming and achieve high gain. Thus to ensure high perfor-mance, choosing the suitable BS to user equipment (UE) beam-pair from theset of all the possible directional beams is a crucial task. Beam selection hasbeen conventionally addressed using exhaustive search or multi-level selectionapproach as in [12, 15]. However, with these techniques, large number of beamsat mmW BSs leads to large beam training overhead and hence unacceptable la-tency to access the mmW network. Access latency is in turn significantly lowerin case of communication at sub-6GHz frequencies. To overcome this challenge,out-of-band spatial information has been used for reducing beam-selection over-head [1]. In recent years, in order to predict the optimal beam and significantlyovercome the training overhead, the use of deep learning (DL) and machine learn-ing (ML) tools has proved to be very promising in establishing mmW links [3].In this paper, we thus propose a deep neural network (DNN)-based mmW BSand beam selection for heterogeneous network by utilizing basic features fromthe Channel State Information (CSI) available only at sub-6GHz BSs.DL and ML techniques have been hugely explored for various communica-tion applications which include, channel estimation, design of auto-encoders,spectrum allocation, etc. [7]. In the context of mmW communications, suchtechniques have been reported for applications such as beam selection, block-age detection, channel estimation, or proactive handover. Various DL and MLtechniques to reduce the beam selection overhead in mmW communications uselocation information, channel information, out-of-band information or measure-ments from different sensors such as LIDAR, camera, or GPS. Specifically, au-thors in [6] and [5] have proposed the use of deep convolutional neural networksto perform beam selection task in distributed and centralized architecture re-spectively. In [11], authors have considered the use of situational knowledgeabout the environment and location of UEs and proposed the use of ensemblelearning-based classification to identify the optimal mmW beam. Later, in [4],authors proposed the applicability of deep learning techniques such as k-nearestneighbours (KNN), support vector classifier (SVC) and multi-layer perceptronby using angle of arrival information to perform the beam-selection task. Allthese works however assume single-layer networks and ignore the macro-layerof sub-6GHz BSs that will be required for a continuous connectivity. Only tworeferences are dealing with ML/DL-based beam selection in heterogeneous net-works [3,14]. Authors in [14] have considered the CSI over sub-6GHz and kernel-based ML algorithms to assist handovers for target vehicle discovery problemnd overcome coverage blindness. In [3], authors have proposed the use of sub-6GHz channel and location information for performing the beam-selection andblockage prediction task. However, the solution in [3] is limited to a single BS -single UE communication scenario, where the BS employs co-located sub-6GHzand mmW transceivers. In this paper, we extend the work done in [3] by con-sidering multiple coordinating sub-6GHz and mmW BSs to perform resourceallocation for each UE in the network.Furthermore, location is an important feature that independently can beutilized to perform the task. Most of the previously discussed work on beamselection including [3], considers the availability of the UE location. However, thisinformation may not always be readily available for many cellular devices. Also,location sensors usually has low accuracy and can result in incorrect outputs [8].Hence, we aim to intentionally eliminate the availability of location informationfrom the set of input features and design the proposed DNN based BS and beamselection framework for a heterogeneous mmW network.The main contributions of this paper are listed as follows:1. To guarantee reliable communication and enhanced coverage in mmW com-munication, we consider the heterogeneous architecture and propose DNN-based BS and beam selection by leveraging basic signal components ex-tracted from the sub-6GHz channel as the input features. We consider mul-tiple coordinated sub-6GHz BSs for optimal mmW resource allocation inorder to serve any UE in the network.2. We propose a branched DNN-structure, which divides the problem into twosub-problems of BS selection and beam selection and is well-adapted for thisapplication.3. We eliminate the use of location information from the set of input featuresto perform the considered task. The feature vector considered as input tothe network include: the azimuth and elevation angle of arrival (AoA) fromthe BS, the receive signal power, the signal phase and the propagation delay.The remainder of this paper is organized as follows. Sec. 2 describes thenetwork and transceiver model. The proposed problem is formulated in Sec. 3and then the deep neural network model is discussed in Sec. 4. Sec. 5 presents thesimulation environment and performance evaluation and finally Sec. 6 concludesthe paper. Notations : Throughout this paper, we use bold-faced lowercase letters to denotecolumn vectors and bold-faced uppercase letters to denote matrices. For anymatrix X , X T denotes the transpose operation. We consider a heterogeneous cellular network wherein multiple sub-6GHz BSsand mmW BSs operate together in order to serve UEs in the network as shownin the Fig. 1. We assume that there are B µ sub-6GHz BSs, each equipped with N µ antenna elements. All the sub-6GHz BSs operate in a coordinated manner forig. 1: System model: Heterogeneous network architecture with mmW small cellscoexisting with sub-6GHz macro BSs. Dashed lines represent the connection ofcoordinating sub-6GHz BSs with a central cloud processor whereas solid linesrepresent the connection between any sub-6GHz and mmW BS in a network.their processing such as channel estimation or precoder design, along with DNNcomputations being performed at a central cloud processor unit. We assume thatthere are B m mmW BSs distributed in the network region that are coordinatedwith the sub-6GHz BSs to provide high speed data transfer to the UEs in thenetwork. Each of the mmW BS is assumed to be equipped with N m transceiverantennas. We assume that UEs have a single antenna in both bands .The communication scenario that we study is as follows. A UE is initiallyconnected to a sub-6GHz BS and periodically transmits pilot signals to all macroBSs. Whenever the UE is approaching towards mmW BSs, the coordinated sub-6GHz BSs command the best mmW BS and the best beam that maximizes theachievable rate for this user.Based on this scenario, the signal received by the macro sub-6GHz BSs atthe k -th OFDM sub-carrier, k = 1 , , · · · , K can be given by: y µ [ k ] = h µ [ k ] d s + n µ [ k ] , (1)where d s is the uplink pilot transmitted over the h µ sub-6GHz channel gain ma-trix and n µ is the additive Gaussian noise vector with zero-mean and covariancematrix σ µ I at the sub-6GHz BS antenna arrays. The processing at the sub-6Ghzis performed in the baseband domain as the macro BSs are assumed to employfully-digital architecture.However, due to the high cost and power consumption of mixed signal RFcomponents at mmW frequencies, mmW transceivers are assumed to employeither fully-analog architecture where the transceiving unit is associated withsingle RF chain or it employs hybrid analog-digital architecture with a number UEs may be equipped with several antennas but we don’t address in this paper thebeam alignment problem and we focus on the beam selection at the BS. Once theBS beam is known, the UE may for example perform exhaustive search to select itsown beam. f RF chains less than N m . In this work, mmW BSs adopt fully-analog beam-forming architecture where, at a given time instant, the signal is transmittedthrough a single beam which is selected from a finite set V of M predefinedbeams, where V is the codebook. The total transmit power at the mmW BSis P T . Thus for the downlink transmission, where the mmW BS communicateswith the UE, the signal received at the UE can be given as: y m [ k ] = H m [ k ] v m [ k ] d m + n m [ k ] , (2)where H m is the mmW channel gain matrix, v m is the beamforming vector, d m is the data transmitted by the mmW BS and n m is the additive Gaussian noiseat UE with zero-mean and covariance matrix σ m I .We assume that the mmW channel is modelled as a geometric channel [2]which can be given as: H m [ k ] = L (cid:88) l =1 (cid:114) ρ l K e j ( κ l + πkK Γ l B m ) a ( θ l , φ l ) (3)where (cid:112) ρ l K is the path gain for the l -th channel path in the k -th OFDM sub-carrier and κ l and Γ l represents the path phase and propagation delay for the l -th channel path respectively. L is the total number of channels paths. Thearray response vector at the BS is denoted by a ( θ l , φ l ), where θ l and φ l is theazimuthal and the elevation AoA respectively. The detailed study of the utilizedchannel model can be obtained in [2]. The sub-6GHz channel is modelled in thesame way. Given the uplink channel information at sub-6GHz BSs, we aim at designing anoptimal mmW BS and beam predictor such that it maximizes the achievablesum-rate for each user in the network. Thus the optimal beamforming vector v om can be obtained as: v om = arg max v m ∈V K (cid:88) k =1 log (1 + γ | H m [ k ] T v m | ) (4)where γ = P T /Kσ m . To design this optimal predictor we aim to find a mappingfrom sub-6GHz channel to mmW BS and beam selection. [3] has shown that, un-der the assumption that there is a bijective mapping between sub-6GHz channeland user location, there also exists a bijective mapping between sub-6GHz chan-nel and mmW channel. Motivated by this result, we rely on sub-6GHz channelfeatures to deduce the resources in mmW band. We can thus define two mappingfunctions ζ BS , ζ b as follows: ζ BS : f µ → P BS (5) ζ b : f µ → r b (6)ig. 2: Deep Neural Network (DNN) model for optimal mmW BS and beamselection.where f µ is a feature vector of size n f extracted from the CSI in the sub-6GHzband, P BS is a probability mass function on the set of mmW BSs and r b is avector of achievable rates for every possible beam out of M at a mmW BS. Tofind this mapping, we utilize the DNN-based approach which are well-suited forobtaining the non-linear relationships between different data distributions [10].The ζ BS mapping is formulated as a classification problem, in which each inputfeature is mapped into a finite set of labels; each label representing candidatemmW BSs, while ζ b mapping is obtained by solving this sub-problem as a re-gression task wherein, a real valued achievable rate is obtained for each beamfor the selected BS from ζ BS mapping. The proposed DNN based solution ispresented in details in Sec.4. In this section, we discuss the DL model adopted to learn the mapping from sub-6GHz channel information to mmW-BS identifier and its beam for a given user.In an environment with multiple mmW BSs and large number of beams, it isimportant to have flexibility in the network to incorporate new BSs or beams forfuture requirements. To allow this scalability, the overall beam selection problemcan be divided into two sequential sub-problems: optimal mmW BS selection andthen optimal beam selection. In order to incorporate the two sub-problems ina single neural network, we consider a branched network which takes featurevectors f µ from sub-6GHz CSI as input and predicts both mmW BS and beamsfor that user as shown in Fig 2. .1 Base Network We consider a base network for both the sub-problems to learn the commonfeature vectors. The input of this base network is a matrix of dimension B µ × n f which gathers all the features for every sub-6GHz BS. We consider a convolutionlayer as the first layer of the base network with the kernel of size 1 × n f . This layeracts as a shared weight perceptron layer which is intended to find the correlationwithin the feature vector of each coordinating sub-6GHz BS. The output of thislayer is passed through another convolution layer having kernel size B µ ×
32. Thesecond convolution layer is intended to learn the correlation between the differentmacro BSs. We then flatten the output and pass the learned features through astack of two fully-connected dense layers of size 128 and 256 respectively. All thelayers are with Rectified Linear Unit (ReLU) non-linearity activation functionas in Fig. 2. The output of the final layer of the base network is branched intotwo sub-networks that are designed to solve each of the sub-problem of mmWBS and beam selection as discussed in following subsections.
This sub-network is designed to predict the optimal mmW BS in order to servethe desired UE in the communication area. The input to this network are the fea-tures learned from the base network. This input vector is further passed throughtwo fully-connected dense layers of size 128 and 64 respectively, for the optimalBS selection specific feature learning. These learned features are then projectedonto the B m feature space using a final dense layer. The output of this layer isthen fed to a softmax activation which results in a probability distribution overthe number of mmW BSs. The BS with the highest probability is selected as theoptimal BS. The beam selection sub-network utilizes the learned features from the base net-work in order to predict the best beam. We incorporate two fully-connecteddense layers of size 128 and 64 respectively, each of which is followed by ReLUactivation. Moreover as selection of the best beam also depends or gets im-pacted by the selected BS, we concatenate the feature from the hidden layerof the BS selection network with the output of the previous dense layer fromthis network as depicted in Fig. 2. These concatenated features provide addedinformation and hence result in better performance. The output of this layer isfurther passed through a fully-connected dense layer of size 128, to learn thecorrelation within the concatenated features. Finally, we project these learnedfeatures to M dimensional space and pass it though a ReLU activation layer toget the regression output for the achievable sum-rate at each beam. The indexwith maximum sum-rate value is the selected beam for the selected BS from themmW BS selection network. .4 Discussions The proposed branched neural network architecture has been obtained afterexperimenting several DNN configurations. In this section, we discuss these ex-perimented models and provide reasons for adapted changes in the final DNNmodel. We initially considered a multi-layer sequential DNN with single outputvector. We took a concatenated vector of features from all sub-6GHz BSs asinput and expected a single vector of achievable rates for each beam at eachmmW BS as an output. Though this network architecture is simple and per-forms the task directly, we observed that this network show large variations forsmall changes in the environment. Moreover, when the number of mmW BSsincreases, the number of output nodes increases dramatically and the systemthus requires extensive training to achieve good performance.To overcome this issue, we adopted a branched network, where we separatelyselected the optimal mmW BS and then the optimal beam by solving bothmappings as a classification problem. Branching the complete problem to twosub-problems helped in the learning of the system and also showcased small vari-ations for small changes in the environment. The consideration of BS selectionas a classification problem performed well. It was however much less efficientfor beam selection. The reason lies in the fact that due to the large number ofnarrow beams at mmW BSs, the angular difference between any two adjacentbeams is very small, implying that multiple beams can be selected as best beamfor certain user locations. We observed that this overlapping beam behaviourcould not be solved by classification and the network was unable to converge toa solution.To tackle this issue, we modified our branched network where this time, weconsidered the beam selection as a regression problem. To further improve theperformance of the overall system, we formed a link between the BS selectionbranch and the beam selection branch as both of these operations are not mu-tually independent.We adopt a soft decision for the BS selection process, i.e., we compute forevery BS the selection probability and retain the one with the highest proba-bility. In contrast, a hard decision would have selected a BS with a probabilityhigher than a certain and given threshold. Hard decision has been observed tobe training data centered and can guarantee to provide good solutions for fea-tures within the bounded range of the training data. However, a hard decisionmay fail to give good solutions for feature values outside these bounds. A softdecision however, will still provide a solution. Also, when all the BSs are equallyprobable for selection, a hard decision threshold greater than 1 /B m will not pro-vide any solution, while a soft decision will select any one of them. Furthermore,beam selection task is also modelled with soft decision, where the best beam isselected as the one with highest achievable sum-rate. This allows for multiplebeams selection (by considering the first highest sum-rates), a characteristic wewill use to improve the accuracy of the results, as shown in the next section.ig. 3: Simulation Environment [2]. In this section, we illustrate the performance of the proposed DNN based BSand beam selection in a heterogeneous mmW networks. We first describe thesetting of a simulation environment considered throughout the simulations insubsection 5.1 and then discuss the performance results in subsection 5.2.
We consider the outdoor simulation environment provided with the availableopen source DeepMIMO dataset [2]. From the dataset, we consider two differentray tracing scenarios ’O1 3p5’ and ’O1 28’ operating at 3.5 GHz and 28 GHzfrequencies respectively, in order to construct a heterogeneous simulation envi-ronment. We consider two sub-6GHz coordinated BSs and eight mmW BSs. ThedeepMIMO dataset generates the channel at these frequencies. Given the CSI,we extract the basic components and construct the feature vectors from utilizingonly the sub-6GHz channel, which acts as the input to our proposed DNN model.Essentially, we consider the azimuthal and elevation AoA, signal power, path lossand signal phase as the extracted features from the sub-6GHz CSI. Intentionally,we don’t assume the availability of the UE location, as this information may notbe available at the device. The hyperparameters considered for the generation ofthe dataset for training and testing are given in Table 1. The outdoor simulationenvironment we considered is given in Fig. 3. It is an urban environment with theBSs placed along the side of the road. We considered a subset of BSs and usersfor our experiments, the list of which is given in Table 1. Users are consideredto be present on the road and are densely populated for better data generation.Building of varying height, width and material are placed along the road pro-viding blockages and reflections. For both scenarios, we considered 1024 OFDMsubcarriers with an OFDM sampling factor of one, where sampling factor is theable 1: Dataset parameters for mmW BSs operating at 28 GHz and macrosub-6GHz BSs operating at 3.5 GHz.
Parameters 28GHz Scenario 3.5GHz ScenarioActive BSs 2,3,4,5,6,7,8,17 1,18Active users 1651-2200, 3500-5203 1651-2200, 3500-5203Number of BS Antennas 256 16Antenna spacing ( × wavelength) 0.5 0.5Bandwidth (GHz) 0.5 0.02Number of OFDM subcarriers 1024 1024OFDM sampling factor 1 1OFDM limit 64 64Number of paths 1 1 Training Data Selection Ratio A cc u r acy Total AccuracyBeam Selection AccuracyBS Selection Accuracy
Fig. 4: Performance evaluation of the BS selection, beam selection and totalaccuracy Vs varying ratio of training data with respect to total training dataset.rate at which we can sample the OFDM subcarriers. Furthermore, the OFDMlimit specifies the number of sampled subcarriers to be considered. We set thislimit to 64 for both scenarios, which implies that we calculate the channels onlyat the first 64 sampled subcarriers. Detailed explanation about the simulationenvironment can be referred in [2].
In this subsection, we present the simulation results demonstrating the per-formance of the proposed scheme, while analyzing the effect of the number ofselected beams, the training dataset selection, and the location parameter onaccuracy and latency.In Fig. 4, we evaluate the performance of the proposed DNN-based BS pre-diction, best beam prediction and overall prediction accuracy, where the totalaccuracy is obtained by correct prediction of both BS and beam against the
50 100 150 200 250 300 350 400 450 500
Epochs A cc u r acy With Location, Best 1 beamWithout Location, Best 1 beamWithout Location, Best 3 beam
Fig. 5: Accuracy Vs number of epochs comparing proposed DNN architecturebased beam selection predicting best 1 beam and best 3 beams without consid-ering location information in features and proposed DNN architecture predictingbest beam while considering the location information.varying size of the training dataset. We divide the overall data with a 80:20ratio where 80% of the total data is used for training whereas the remaining20% dataset is used for validation/testing purpose. Out of this total available80% training dataset, we utilise varying training data ratios and observe theperformance in terms of accuracy for the proposed system. The system perfor-mance illustrates that the network is able to achieve high accuracy for both BSand beam selection tasks. The achievable BS selection accuracy is around 97%whereas the beam can be predicted with 88%. The total accuracy of correctlypredicting both the optimal BS and beam is close to 86% when we use completetraining dataset. We observe comparable performance with 50% of training dataas compared to complete training dataset. This means that we can quickly ob-tain good results offline and apply the algorithm online and then improve theperformance by training over the time.We compare the performance of the proposed DNN architecture while nowconsidering the UE location as one of the input features. Fig 5 shows this per-formance as a function of the number of epochs. We compare the performancefor the best beam and the best three beams with and without location. As ex-pected, we observe that the location-aided design performs better. However, theperformance can be improved by selecting the best b beams, b = 1 ...M , hencereducing the performance gap between the architecture with or without locationparameter. In Fig. 6, we demonstrate the beam selection accuracy with respect tonumber of beams predicted for the proposed DNN for varying size of the trainingdataset. As expected, it is observed that the beam selection accuracy increaseswith the increasing number of predicted beams as well as with the increasingsize of the training data. From this figure, we can further observe and analysethe effect of latency for the proposed system. Indeed, an exhaustive search wouldrequire to perform M received power measurements (64 in our case), while with No of beams predicted A cc u r acy Training Data 5%Training Data 30%Training Data 50%Training Data 100%
Fig. 6: Accuracy Vs number of predicted beams by the proposed DNN beamselection with varying training dataset.our solution, we can can achieve 85% accuracy by measuring only the best threebeams selected by the network.
In this paper, we propose a branched DNN model that jointly performs themmW BS prediction and beam selection task in a heterogeneous network ar-chitecture. We consider that multiple mmW BSs coexist with multiple legacysub-6GHz BSs to serve the UE in the network area. The sub-6GHz BSs are as-sumed to function in a coordinated manner and are supported by the centralcloud processor. We formulate the mmW BS prediction as a classification prob-lem whereas the optimal beam selection is mapped into a regression problem.For both the tasks, we utilize the channel components available only at the sub-6GHz BSs as a set of input features. As the location information may not bealways available or it can be inaccurate due to sensor errors, we intentionallyeliminate the use of location as an input feature for the proposed problem. Wecompare the performance of the proposed DNN based design with conventionalexhaustive search and observe the success probability close to 1 for allocatingoptimal mmW BS and beam while using reduced computational resources. Com-parable performance can be achieved with and without user location availableprovided that the three best beams are considered. At last, we show that muchfewer beam power measurements are required compared to exhaustive search,which results in lower latency.
References
1. Ali, A., Gonz´alez-Prelcic, N., Heath, R.W.: Millimeter wave beam-selection usingout-of-band spatial information. IEEE Transactions on Wireless Communications (2), 1038–1052 (Feb 2018). https://doi.org/10.1109/TWC.2017.2773532. Alkhateeb, A.: DeepMIMO: A generic deep learning dataset for millimeter waveand massive MIMO applications. ArXiv abs/1902.06435 (2019)3. Alrabeiah, M., Alkhateeb, A.: Deep learning for mmwave beam and blockage pre-diction using sub-6GHz channels. ArXiv abs/1910.02900 (2019)4. Anton-Haro, C., Mestre, X.: Data-driven beam selection for mmWave communica-tions with machine and deep learning: An angle of arrival-based approach. In: 2019IEEE International Conference on Communications Workshops (ICC Workshops).pp. 1–6 (May 2019). https://doi.org/10.1109/ICCW.2019.87569915. Dias, M., Klautau, A., Gonz´alez-Prelcic, N., Heath, R.W.: Position and LIDAR-aided mmWave beam selection using deep learning. In: 2019 IEEE 20th Inter-national Workshop on Signal Processing Advances in Wireless Communications(SPAWC). pp. 1–5 (July 2019). https://doi.org/10.1109/SPAWC.2019.88155696. Klautau, A., Gonz´alez-Prelcic, N., Heath, R.W.: LIDAR data for deep learning-based mmWave beam-selection. IEEE Wireless Communications Letters (3), 909–912 (June 2019). https://doi.org/10.1109/LWC.2019.28995717. Morocho-Cayamcela, M.E., Lee, H., Lim, W.: Machine learning for 5G/B5G mobileand wireless communications: Potential, limitations, and future directions. IEEEAccess , 137184–137206 (2019). https://doi.org/10.1109/ACCESS.2019.29423908. Odijk, D., Kleijer, F.: Can gps be used for location based servicesat schiphol airport, the netherlands? In: 2008 5th Workshop on Po-sitioning, Navigation and Communication. pp. 143–148 (March 2008).https://doi.org/10.1109/WPNC.2008.45103689. Sakaguchi, K., Tran, G.K., Shimodaira, H., Nanba, S., Sakurai, T., Takinami, K.,Siaud, I., Strinati, E., Capone, A., Karls, I., Arefi, R., Haustein, T.: Millimeter-wave evolution for 5G cellular networks. IEICE Transactions on Communications E98.B (12 2014). https://doi.org/10.1587/transcom.E98.B.38810. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural Networks , 85 – 117 (2015). https://doi.org/https://doi.org/10.1016/j.neunet.2014.09.003,
11. Wang, Y., Klautau, A., Ribero, M., Soong, A.C.K., Heath, R.W.: MmWave ve-hicular beam selection with situational awareness using machine learning. IEEEAccess , 87479–87493 (2019). https://doi.org/10.1109/ACCESS.2019.292206412. Wei, L., Li, Q., Wu, G.: Initial access techniques for 5G NR: Omni/beamSYNC and RACH designs. In: 2018 International Conference on Comput-ing, Networking and Communications (ICNC). pp. 249–253 (March 2018).https://doi.org/10.1109/ICCNC.2018.839032513. Xiao, M., Mumtaz, S., Huang, Y., Dai, L., Li, Y., Matthaiou, M., Karagiannidis,G.K., Bj¨ornson, E., Yang, K., I, C., Ghosh, A.: Millimeter wave communicationsfor future mobile networks. IEEE Journal on Selected Areas in Communications (9), 1909–1935 (Sep 2017). https://doi.org/10.1109/JSAC.2017.271992414. Yan, L., Ding, H., Zhang, L., Liu, J., Fang, X., Fang, Y., Xiao, M., Huang, X.:Machine learning-based handovers for sub-6 GHz and mmWave integrated vehicu-lar networks. IEEE Transactions on Wireless Communications (10), 4873–4885(Oct 2019). https://doi.org/10.1109/TWC.2019.293019315. Yu, H., Qu, W., Fu, Y., Jiang, C., Zhao, Y.: A novel two-stage beam selectionalgorithm in mmWave hybrid beamforming system. IEEE Communications Letters23