Relay Selection for 5G New Radio Via Artificial Neural Networks
RRelay Selection for 5G New Radio Via ArtificialNeural Networks
Saud Aldossari , Kwang-Cheng Chen
Department of Electrical EngineeringUniversity of South Florida Tampa, Florida 33620, USAEmail: [email protected], [email protected]
Abstract —Millimeter-wave supplies an alternative frequencyband of wide bandwidth to better realize pillar technologies ofenhanced mobile broadband (eMBB) and ultra-reliable and low-latency communication (uRLLC) for 5G - new radio (5G-NR).When using mmWave frequency band, relay stations to assistthe coverage of base stations in radio access network (RAN)emerge as an attractive technique. However, relay selection toresult in the strongest link becomes the critical technology tofacilitate RAN using mmWave. An alternative approach towardrelay selection is to take advantage of existing operating dataand apply appropriate artificial neural networks (ANN) anddeep learning algorithms to alleviate severe fading in mmWaveband. In this paper, we apply classification techniques using ANNwith multilayer perception to predict the path loss of multipletransmitted links and base on a certain loss level, and thus executeeffective relay selection, which also recommends the handoverto an appropriate path. ANN with multilayer perception arecompared with other ML algorithms to demonstrate effectivenessfor relay selection in 5G-NR.
Keywords:
Machine Learning, Wireless Communications,MmWave, Neural Network, Multilayer Perceptrons, Classification,Relay Selection, SVM and Logistic Regression, 5G-NR.
I. I
NTRODUCTION
Relay selection [1] to form cooperative communication has be-come a critical technology with the 5G - new radio (5G-NR)and future mobile communications. Relay selection in multi-hopcommunication was shown an adorable technique for mobile com-munication over mmWave frequency bands [2], [3] The sensitivityof mmWave signal to fading remains a fundamental challenge incommunication systems especially for 5G era. Authors of [4] haveproposed a novel adaptive multi-state selection utilizing different ofmmWave frequencies. The fifth-communications generation goals areprioritized base on three pillar which are enhanced mobile broadband(eMBB), ultra-reliable, low latency communications (URLLC), andmassive machine type communications (mMTC). In this work, we aretrying to meet the first two goals base on relay selection with a newmechanism that increases the communication strength and improvethe reliability. Moreover, this work may enhances the communicationbetween massive devices to meet the mMTC. During the commu-nication propagation, the transmitted signal can be affected by thesurrounding environments resulting in the signal diffraction, scatter-ing, and reflection as showing in Figure 1. Having a line-of-sight(LOS) transmission does not mean obtaining a proper transmissionbut relaying on other propagation links may have a better performanceand coverage. Figure 1 demonstrates three propagated signals from(BS) where the first link
P L
LOS is the LOS signal,
P L
Bs,r is thesecond transmitted signal which handover to another station and thethird link is affect due to obstacles OB then penetrate through tothe destination mobile station MS . To allow the destination point toselect best link base on the propagation signal strength to meet the meet the 5G -new radio (NR) requirement (the ultra-reliable and lowlatency communications) in term or reliability with 99.999% [5]. Inthis paper, we demonstrate a new machine learning methodology thatacculturate the link selection from base stations or users equipmentside. Fig 1. Demonstrate the propagation links of a wireless communication wherethe destination user equipment (UE) selects optimum link with lower loss. C i ( x ) = (cid:40) P L < dBm P L ≥ dBm (1)Where C i ( x ) is the link selection classes which depends on the pathloss of the link that can be calculated using models such as Floating-Intercept (FI) model. P L FI ( f, d )[ dB ] = α + 10 βlog ( d ) + X FLσ (2) BL = arg min L s ( P L n ) (3)Where BL is the best link selection and P L n is the path loss of thepropagated links. While n is the number of transmitted links betweenbase station and destination.To meet the 5G pillars technologies particularity the eMBB anduRLLC, we propose a new mechanism to overcome the affectedpropagation link using artificial neural networks. Using relay linkselection, we note that the classification of machine learning issuitable for description and prediction of the optimum link/path.Having a reliable mechanism to meet the uRLLC with trustworthycommunications and eMBB to enhance the coverage and improve thecommunications. The supervised classification algorithms can be per-formed to predict categorical class labels using the training dataset ofthe outdoor urban environment, and can be implemented by artificialneural network (ANN) with deep learning. Furthermore, MultilayerPerceptrons (MLP) methods in ANN, together with different models,are evaluated to predict and select the strongest propagation link.MLP therefore serves the classification in ANN to identify and a r X i v : . [ ee ss . S Y ] M a y haracterize the new link candidates using the path loss parameteror the receive signal strength. This work also will influence to themassive machine type communications (mMTC). The classificationtechnique that was selected is a binary where there are only twoclasses which are strong link and weak link. Once we obtain apropagation results, the new mechanism of this work divides thesignal losses into categories, these levels are classes. In our case wehave a binary classes, where each path loss signal strength is eitherconsidered sufficient or insufficient (no fading). The base stationmakes a decision based on that categorical. Thus, identifying theoptimum link using a classification algorithms to meet the reliabilityand coverage of the 5G NR. The base station learns how to predictthe weakest path loss and select the minimum loss path. The basestation selects the propagated signals base on a certain energy strength(threshold) and once that link energy reaches this threshold the basestation will switch to another link. In this study, the threshold is baseon the path loss which is equal to -120 dBm and below this thresholdis considered a poor propagation. Thus, eMBB and uRLLC will beachieved.[6] suggests using deep learning to identify and classify the mod-ulation nodes, improving the interference alignment and locate theoptimum routing path. Furthermore, applying prediction techniquesusing methods such as classification and clustering etc. to estimatethe channel path loss models that lead to a better performance andprecision. Multilayer Perceptrons is a ML classification techniquewhich is a neural network. The data can be classified based onmaximum probability in Multilayer Perceptrons techniques to predictthe path loss that can be expressed as: ˆ C = arg max C i n (cid:89) i =1 P ( C i /X ) (4) ˆ C is the prediction path loss class and P ( X i /C ) is the conditionalprobability of dataset feature given the class. [7] published an articleshowing how machine learning techniques such as Deep Neural Net-work (DNN) reduce the complexity and increase the performance. Inthis manuscript, Multilayer Perceptrons Neural Network is introducedin II. Followed by section III that shows the dataset. Then, modelsvalidation and results in IV. Lastly, a conclusion is shown in sectionV. II. M
ULTILAYER P ERCEPTRONS N EURAL N ETWORK
Among many DNN structures, Multilayer Perceptrons (MLP) usesa Feed-forward neural networks (FFNNs) and a back-propagationnetwork to compute the loss and adjust the weight [12], which issuitable for deep learning. MLP forms a fully connected networkswhere every single node in a single layer is connected to every nodein the following layers. The subsequent error is usually obtained bythe loss function and optimization methods can be use to minimize theloss such as Adam optimizer. There are multiple of loss functions andcross entropy will be used when relay selection can be initially viewedas a binary classification problem. MLP is actually a multivariatemultiple nonlinear regression and collection of neurons that serves asa classification by building decision decision. Multilayer Perceptronsare usually uncorrelated, and a collection of them make up thenetwork that can be less prone to the notorious overfitting. MLPis mathematically mapping in the form of: (cid:60) n −→ (cid:60) m : ( y , y , ..., y n ) (5) y n = g s (cid:32) w + n (cid:88) i =1 w i y i (cid:33) (6) y = g out (cid:32) w (2) k + M (cid:88) j =1 w (2) k γ (cid:16) w (1) j + n (cid:88) i =1 w (1) ji y i (cid:17)(cid:33) (7) The above structure can proceed with only two layer, where y = 1 as the output of the first layer. g s is the activation function and herewe are using the step function as can be shown as: g ( · ) : R → R (8) g s ( x ) = (cid:40) x < x ≥ (9)To accomplish our purpose, Adam optimization algorithm will beadopted which is different than the traditional stochastic gradientdescent process to update the weights iterative base in the trainingdata [13] with a learning rate or step size α . Artificial intelligence(AI), particularly machine learning (ML), is widely studied to enablea system to learn of intelligence, predict and make an assessmentinstead of the needs of humans [14]. Switching the traditionallink selection such as Adaptive selection scheme [15] to machinelearning link selection still in its early stage. One of the main issuesin current communications is the accuracy of handover, whereasusing machine learning techniques could enhance the prediction andreduce the complexity. The new ML methods predict the best linkusing different mechanisms base on path loss and receive signalstrength. [16] showed how to predict the transmitted signals usingdeep learning techniques [17] used neural networks methods suchas learned decisions-based approximate message passing (LDAMP)network to estimate and learn channel state information (CSI) thensolving the limited number of frequency chains in cellular systemsfrom training data. [18] used a model-based method using Cramer-Rao lower bound (CRLB) to predict the channel state parameters inthe deeper neural network. Moreover, other authors presented in [19]and [10] some communications challenges that reach the complexitylevel such as atmospheric effects, handover, beam direction, MIMOand it is the time for machine learning to get evolved. Machinelearning uses training and testing for letting the machines learn andkeep predicting. In training part, learning from the data while thetesting method, a trained model is used for predicting such as theray selection. Supervised learning techniques require input, target andtraining data to create a model that is used for predicting. If a samplespace consists of X i and output label space y i where i , , .., N thenby using a machine learning algorithms ˚ A , which is a function thatmap the input values to the labels that helps for future predicting. Tomeasure the quality of the mapping, a loss function is used and see[20] for more details. The classification algorithms that will be usedin this journey is Multilayer Perceptrons Neural Network and can bedescribed in the following.Multilayer Perceptrons Neural Network usually can be used toboth classification and regression. When the Multilayer Perceptronsused for the classification, this algorithm works by having binary ormultiple classes. However, with the regression techniques, is usuallyused for continuous outputs while our goal here is to classify the linkstrength to binary classes to predict the optimum link propagation.Multilayer Perceptrons follow the form as shown below. y = Φ( n (cid:88) i =1 w i X i + b ) (10)Where w is the vector of weights of x vector inputs and b is the error. Φ is the non linear activation function. In this work, we proposedsix models of Multilayer Perceptrons with different specification asshown below: • Model 1: One Hidden Layer of 10 Neurons • Model 2: Two Hidden Layers of 50 and 10 Neurons • Model 3: Three Hidden Layers of 10, 50 and 10 Neurons • Model 4: Four Hidden Layers of 10, 50, 50 and 10 Neurons • Model 5: Five Hidden Layers of 10, 50, 100, 50 and 10 Neurons • Model 6: Eight Hidden Layers of 10, 50, 100, 100, 50 and 10Neurons.
Model 7: Logistic Regression Model • Model 8: Dummy Classifier Model • Model 9: Support Vector MachineMultilayer Perceptrons neural network will be employed to predict thethe optimum propagated link in our relay selection. Then, comparewith other machine learning techniques base on precision, recall,F1 Score, accuracy and support. Results will be explored usingsimulated data showing the accuracy of applying deep learninglearning techniques and how this algorithm performs well in relayselection. By considering more from wireless communications, theresult and compassion of these machine learning models to predictand selection of the best relay link will be shown later in Section IV.Since both prediction of link performance and classification toselect an appropriate link, MLP neural networks appears fit ourpurpose due to capability of predicting the link with low path loss,which allows a reliable handover to meet the need of eMBB anduRLLC. While other ANN structures such as convolution neuralnetworks are for images where there exist 2D or 3D inputs and theRNN is for sequential models like time series, machine translation,language generation. Further investigations to check the fitness ofthe MLP models compared to other machine learning models will beconducted later in Section IV. Thus, we proved that deep learningtechnique (MLP) is a capable technique to overcome this wirelesscommunications fading issue using link selection base on MLP tech-nique and performed better than other machine learning techniques.
III. D
ATASET
The dataset of this investigation was generated after some modifi-cation using open source Matlab simulation by New York University[21] [22]. The dataset of the wireless channel is composed of twofragments. The selected model will be trained and validated on thedataset then will be tested using the unseen data. In this work, thetrain part took 75% of the data set and 25% for tasting the model inour training/testing scenario and others can be found in [23]. Classesdata which is zeros and ones are specified base on the path lossstrength and other channel states information (CSI) are used to predictthe path loss. The measurements are specified based on distance from1 m until 40 m. That simulation is suitable for frequencies in a rangeof 500 MHz to 100 GHz, bandwidth up to 800 MHz with differentscenarios and environments. As a summary, simulation parametersare listed in Table I, which exhibits that the channel measurementparameters of that data raw that was used for this paper.
Table IC
HANNEL M EASUREMENT P ARAMETERS . Parameters Values
Distance (m) 1-40Frequency (GHz) 28Bandwidth (MHz) 800TXPower (dBm) 30Scenario UMiPolarization Co-PolTxArrayType ULARxArrayType ULAAntena SISOTx/Rx antenna Azimuth and Elevation (red) 10
The dataset that used in this work are consisting of channelproperties of a communications link such that the information helpsthe base station to execute supervised classification based on a datasetfrom prior measurements or simulations.
IV. M
ODELS V ALIDATION AND R ESULTS
To accomplish a broad exploration, Multilayer Perceptrons NeuralNetwork, Logistic Regression, Dummy Classifier and Support VectorMachine are used to perform the classification techniques. Evaluating the performance of these classification algorithms by confusionmatrix which counts the outcomes of the prediction models comparedto training dataset [24]. Moreover, the precision usually shows howoften a model make a positive prediction and recall shows how themodel is confident of the predicting all positive targets. Accuracy,Precision, Recall, and F1 Score metrics were used to evaluate themachine learning classification algorithms (classifiers). The accuracyis measured by counting the number of true prediction to the totalnumber of predictions. Which is the number of correctly predictedselected links over the total number of links, which tells how theclassifier is able not to misclassify a positive path loss (a sample).Precision is the number of true positives ( T p ) over the number oftrue and false positives ( F p ). Recall stands for the number of truepositives over the number of true positives and false negatives (FN).While F1 Score measures the harmonic mean for both precision andrecall, we obtain the following mathematical expressions:Average Precision = 1 n N (cid:88) i =1 T pT p + F p (11)Total Recal = N (cid:88) i =1 T pT p + F N (12) F Score = 2 × precision × recallprecision + recall (13) Table III
NTERPRETATION OF P ERFORMANCE M EASURES . ANN Models Precision Recall F1 ScoreModel 1
Model 2
Model 3
Model 4
Model 5
Model 6
Logistic Regression
Dummy Classifier
SVM
CCURACY COMPRESSION OF ALL MODELS . Models Accuracy ROC AUC ScoreModel 1
Model 2
Model 3
Model 4
Model 5
Model 6
Logistic Regression
Dummy Classifier
SVM
The interpretations of performance measures that were used tocheck the process of a model via precision, recall, f1 Score, accuracyand support. From the two above tables interpretation of performancemeasures of the Multilayer Perceptrons Neural Network algorithms,we can conclude that all of these techniques did a decent job andmodel 5 has the highest accuracy which consist of five HiddenLayers of 10, 50, 100, 50 and 10 Neurons. Model 5 gained in thebest performance in Precision, Recall, and F1 Score among othermodels. Thus, it has the best performance in classifying the the rayselection, followed by the Model 4 and and worst one is model1 among the ANN model and the dummy classifier compared toll models. The reason for that is some of the features dependson each other such as distance and received power. Electing thenumber of hidden layers and number of neurons are still an openresearch topic where few or more neurons leads to underfitting andoverfitting. An assumption from our trail and error trails, we noticedthat the number of hidden layers should lower than the numberof input by 30%. Model 6, began degrading once the number ofhidden layers have reached 70% of the number of the inputs ascan be seen in figure 2. Figure 2 presents the Receiver Operating
Fig 2. ROC curves of Classification techniques.Fig 3. Precision and recall curve of Classification techniques.
Characteristic (ROC) [25] visualizes a classifier’s performance thatillustrates the wellness of the classification models used in thispaper. The false positive rate is plotted against the true positiverate and it’s very obvious that the model 5 is almost closer to theoptimum with 99%. Other models have been compared our modelssuch support vector machine (SVM) that was presented in [26] thatcould not preform well in overcoming the relay link selection. Theusage of sigmoid kernel while using the linear kernel poly kernelwith degree 4 would improve the performance. Underfitting andoverfitting were examined base on checking the accuracy and theROC score of the training and the testing data which were showingclose to each other and performing good. Figure 3 illustrates therelay selection precision recall curve where model 5 is performingthe best among other models. The precision recall curve shows therelationship between true positive rate and the positive prediction
Fig 4. Loss VS iteration of ANN Classification techniques. value for a varieties of models. While figure 4 shows the historyloss verse neural iterations when the training data will not improvethe performance of the model by at least tolerance (usually assigned1e-4) or having a constant loss for multiple of iteration. From thefigure, we notice the losses of models decreased nicely and smoothlyexcept model one due the adjusted learning rate of this model whichwas 1e-5 while others are 0.05 and increase that rate will affectthe accuracy of the model. Moreover, by looking at model 6, wenotice at iteration number 185, it starts increasing which is a sign tostop the model to avoid issues such as overfitting and decreasing theefficiency of the model. Figure 4 again confirm the best performancegoes to model 5 among other NN models. Simply we concludethis journey from the result section that ANN performs better interm of selecting the optimum link comparing to other machinelearning techniques where the accuracy of selecting the optimumlink is 99% which meet third 5G -new radio (NR) requirement (theultra-reliable and low latency communications). The future workto develop are extraordinarily rich, and powerfully using machinelearning algorithms toward solving wireless communications issuesto enhance the the communications and reduce the complexities forfuture wireless generations. This approach can be further extended tosolve other wireless communications issues such as near-far problembase on more than a binary classification. Base on a certain pathloss, power transmission control can be manipulated to achieve a fullcommunications efficiency.
V. C
ONCLUSION
Wireless communications with the new era of 5G and beyondrequire overcome classical issues in order to meet the 5G pillars thatincludes eMBB, uRLLC and mMTC. One of these critical issuesis relay selection to handover with reliability to the strongest linkto meet the eMBB and uRLLC. This can be solved by applyingmachine learning techniques such as Multilayer Perceptrons NeuralNetwork, Logistic regression, Dummy classifier and support vectormachine to develop alternative techniques to predict the signal pathloss strength that’s usually get affected by the wireless channel.Multilayer Perceptrons is classification technique used in this journeyand compared the result of each model by using the interpretationof performance measures such as accuracy, precision, recall and F1-score and compare with previous studies to end up with a betterperformance. Other techniques influenced a perfect predication thatconfirm the usage of machine learning towards wireless communica-tions.
I. A
CKNOWLEDGEMENT
Saud Aldossari expresses a great appreciation to Prince Sattambin Abdulaziz University for their support of providing scholarship.K.-C. Chen appreciates the support from Cyber Florida. R EFERENCES[1] Liu, K. J. R., Sadek, A. K., Su, W., Kwasinski, A. (2009). Cooperativecommunications and net- working. Cambridge: Cambridge UniversityPress.[2] Qiao, J., Cai, L. X., Shen, X. S., Mark, J. W., and Fellow, L. (2011). En-abling multi-hop concurrent transmissions in 60 GHz. Wireless PersonalArea Networks, 10(11), 38243833.[3] Wu, S., Member, S., Atat, R., and Member, S. (2018). Improving thecoverage and spectral efficiency of millimeter-wave cellular networksusing device-to-device relays. IEEE Transactions on Communications,66(5), 22512265.[4] Attiah, M. L., Isa, A. A. M., Zakaria, Z., Abdullah, N. F., Ismail, M.,and Nordin, R. (2018). Adaptive multi-state millimeter wave cell selectionscheme for 5G communications. International Journal of Electrical andComputer Engineering (IJECE), 8(5), 29672978.[5] M. Amjad, L. Musavian, and M. H. Rehmani, Effective capacityin wireless networks: A comprehensive survey, arXiv preprint arX-iv:1811.03681, 18 Jun 2019.[6] Q. Mao, F. Hu and Q. Hao, ”Deep Learning for Intelligent WirelessNetworks: A Comprehensive Survey,” in IEEE Communications Surveysand Tutorials, vol. 20, no. 4, pp. 2595-2621, Fourthquarter 2018. doi:10.1109/COMST.2018.2846401[7] G. Hackeling, Mastering Machine Learning With scikit-learn. 2014, p. 14[8] C. Wang, J. Bian, J. Sun, W. Zhang and M. Zhang, ”A Survey of5G Channel Measurements and Models,” in IEEE CommunicationsSurveys Tutorials, vol. 20, no. 4, pp. 3142-3168, Fourthquarter 2018.doi: 10.1109/COMST.2018.2862141[9] Piacentini, M. and Rinaldi, F., ”Path loss prediction in urban environmentusing learning machines and dimensionality reduction techniques”, inSpringer Computational Management Science, vol. 8, no. 4, pp. 371–385,Nov. 2011.[10] Aldossari, Saud Mobark and Chen, Kwang-Cheng , ”Machine Learningfor Wireless Communication Channel Modeling: An Overview,” in Wire-less Personal Communications, vol. 106, no. 1, pp. 41-70, March 2019.doi: doi.org/10.1007/s11277-019-06275-4[11] Piacentini, M. and Rinaldi, F., ”Path loss prediction in urban environmentusing learning machines and dimensionality reduction techniques”, inSpringer Computational Management Science, vol. 8, no. 4, pp. 371–385,Nov. 2011.[12] Haykin S (2008)Neural Networks and Learning Machines. Prentice-Hall[13] Diederik Kingma and Jimmy Ba. Adam: A method for stochasticoptimization. arXiv preprint arXiv:1412.6980, 2014.[14] C. Jiang, H. Zhang, Y. Ren, Z. Han, K. Chen and L. Hanzo, ”MachineLearning Paradigms for Next-Generation Wireless Networks,” in IEEEWireless Communications, vol. 24, no. 2, pp. 98-105, April 2017. doi:10.1109/MWC.2016.1500356WC[15] V. K. Shah and A. P. Gharge, A review on relay selection techniquesin cooperative communication, International Journal of Engineering andInnovative Technology, vol. 2, no. 5, pp. 6569, 2012.[16] Ben-Hur, A.; Horn, D.; Siegelmann, H.T.; Vapnik, V. (2001). ”Supportvector clustering”. Journal of Machine Learning Research. 2: 125137.[17] Hengtao He, Chao-Kai Wen, Shi Jin, and Geoffrey Ye Li, DeepLearning-based Channel Estimation for Beamspace mmWave MassiveMIMO System, in arxiv , pp. 25, 2018.[18] Zhiyuan Jiang, Ziyan He, Sheng Chen, Andreas F. Molisch, ShengZhou, and Zhisheng Niu, Inferring Remote Channel State Information:Cramr-Rao Lower Bound and Deep Learning Implementation, IEEEGLOBECOM 2018, accepted.[19] O’Shea, Timothy J., Tugba Erpek, and T. Charles Clancy. ”Deep learningbased MIMO communications.” arXiv preprint arXiv:1707.07980 (2017).[20] M. Mohri, A. Rostamizadeh, and A. Talwalkar. Foundations of MachineLearning.