Learning-based WiFi Traffic Load Estimation in NR-U Systems
Rui Yin, Zhiqun Zou, Celimuge Wu, Jiantao Yuan, Xianfu Chen, Guanding Yu
IIEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x PAPER
Learning-based WiFi Traffic Load Estimation in NR-U Systems ∗ Rui YIN † , Member , Zhiqun ZOU †† , Nonmember , Celimuge WU ††† , Member , Jiantao YUAN †††† ,Xianfu CHEN ††††† , and Guanding YU †††††† , Nonmembers
SUMMARY
The unlicensed spectrum has been utilized to make up theshortage on frequency spectrum in new radio (NR) systems. To fully ex-ploit the advantages brought by the unlicensed bands, one of the key issuesis to guarantee the fair coexistence with WiFi systems. To reach this goal,timely and accurate estimation on the WiFi traffic loads is an importantprerequisite. In this paper, a machine learning (ML) based method is pro-posed to detect the number of WiFi users on the unlicensed bands. An un-supervised
Neural Network (NN) structure is applied to filter the detectedtransmission collision probability on the unlicensed spectrum, which en-ables the NR users to precisely rectify the measurement error and estimatethe number of active WiFi users. Moreover, NN is trained online and therelated parameters and learning rate of NN are jointly optimized to estimatethe number of WiFi users adaptively with high accuracy. Simulation resultsdemonstrate that compared with the conventional Kalman Filter based de-tection mechanism, the proposed approach has lower complexity and canachieve a more stable and accurate estimation. key words:
NR-U, WiFi user numbers, neural network, unsupervised learn-ing
1. Introduction
The fifth generation (5G) of mobile networks worldwide hasalready provide low latency, high transmission rates and fullcoverage services to fulfill the explosive diversified data de-mands, such as virtual reality (VR), vehicles-to-everything (V2X), e-health, and smart grid. To further exploit the per-formance of 5G networks and expand the application sce- † The author is with the School of Information and ElectricalEngineering, Zhejiang University City College, Hangzhou, China.e-mail: [email protected]. †† The author is with the School of Information and ElectricalEngineering, Zhejiang University, Hangzhou, China. †††
The author is with Graduate School of Informatics andEngineering, The University of Electro-Communications, 1-5-1,Chofugaoka, Chofu-shi, Tokyo 182-8585, Japan. ††††
The author is with the Institute of Ocean Sensing and Net-working of the Ocean College, Zhejiang University, Zhoushan,China. †††††
The author is with the VTT Technical Research Centre of Fin-land, Finland. ††††††
The author is with the School of Information and ElectricalEngineering, Zhejiang University, Hangzhou, China. ∗ This work was supported in part by the the National Natu-ral Science Foundation of China under Grant No. 61771429, No.61703368, and in part by Zhejiang University City College Scien-tific Research Foundation under Grant No. JZD18002. ZhejiangLab’s International Talent Fund for Young Professionals and Avia-tion Key Laboratory of Science and Technology on Fault Diagno-sis and Health Management 20183333001 also provided supportfor the project.DOI: 10.1587/trans.E0.??.1 narios of 5G transmission at close region, one solution isto share the unlicensed spectrum with WiFi system on the5GHz unlicensed national information infrastructure (U-NII) band with approximately 600 MHz free bandwidth [1].Even though the unlicensed spectrum resource is freeand juicy, it is for public use and mainly occupied by WiFinetworks nowadays. Therefore, the harmonious coexistencewith WiFi networks must be guaranteed while providing ser-vice on unlicensed bands in the cellular systems. The (3GPP) has studied new Ra-dio on Unlicensed bands (NR-U) in Release 16 [2] after thestandardization of new Radio (NR) in 5G. Unlike the con-ventional centralized medium access control (MAC) proto-col employed in the NR on licensed bands, NR-U has toadopt the contention based MAC to ensure the fair sharingon unlicensed spectrum.Before the start of the study items on NR-U, most ofworks have concentrated on the channel access scheme de-sign under the constraint of fair coexistence with WiFi inthe long term evolution on unlicensed spectrum (LTE-U).The licensed-assisted access (LAA) proposed in Release 13has been considered as the main candidate for the LTE-Unetworks [3]. In [4], an adaptive back-off scheme has beenproposed for the LAA based small cell networks to avoidthe access collision with WiFi network under the assump-tion that the WiFi traffic load is known in advance. In [5], adeep reinforcement learning algorithm based on long short-term memory cells has been applied to learn the cellular traf-fic loads and the utilization pattern of the unlicensed chan-nels by the WiFi networks. Therefore, the LAA small basestation (BS) can decide how to use the unlicensed channelsaccordingly. The application of LAA in NR-U networks hasbeen investigated in [6], where the cooperation among smallcells is leveraged to decrease the collision probability andincrease the system capacity. Moreover, the edge computingtechnologies are applied to integrate both licensed and unli-censed spectrum resources in [7], where a joint aggregation,caching, and decentralization scheme has been proposed toenhance system performance.The obtainable accurate WiFi traffic load at the NR cel-lular systems is a key to guarantee the fair coexistence withWiFi networks while improving the spectrum efficiency (SE)on unlicensed spectrum [8], [9]. However, most of exist-ing literatures configure access parameter to ensure fairnesson the assumption that the WiFi traffic load is known [10]–[14]. As the WiFi traffic load is mainly decided by the ac-Copyright © 200x The Institute of Electronics, Information and Communication Engineers a r X i v : . [ c s . I T ] F e b IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x tive WiFi users, only several works study how to estimatethe real traffic condition. For instance, an extended
Kalmanfilter (KF) scheme has been applied to estimate the numberof active WiFi users based on the measured collision proba-bility on the corresponding unlicensed channels in [15]. Re-sults show that it works well when the WiFi traffic load islow, however, the estimation error can not be ignored whenthe traffic load is heavy. Moreover, its high computationalcomplexity requires more computing resources and limitsits application in practice. Some improved KF based algo-rithms are proposed to predict the number of WiFi terminalsand signal noise ratio (SNR) [16]–[19], where WiFi trans-mission model is further improved. However, the KF basedmethods perform well only on linear, discrete and finite-dimension states and the problem mentioned above still ex-ists. Therefore, a more effective algorithm needs to be de-signed to ensure the accuracy of the estimation in a dynam-ically loaded WiFi environment.In this paper, to avoid high computational complexityand improve the estimation accuracy, an online trained neu-ral network (NN) structure is proposed to learn the WiFitraffic load information. With the excellent extensibility ofNN, it can estimate and track the changes on the number ofWiFi users faster as well as accurately under various WiFitraffic load. With the estimation of WiFi traffic load, theproposed algorithm can maximize the spectrum efficiencyof the unlicensed spectrum while ensuring the best perfor-mance of WiFi system.The main contributions of the paper are summarized asfollows:1. A targeted NN with simple three hidden layers is de-signed to estimate the number of WiFi users. Attributeto its simple structure, the sensitivity of the system isguaranteed and the computational complexity can bereduced significantly.2. The effective threshold mechanism based on cumula-tive summary (CUSUM) is applied to track the changeon the number of WiFi users by the proposed scheme.Moreover, an adaptive learning rate and loss functionadjustment are proposed to accelerate the convergencespeed of the NN as well as ensure the stability of theoutput.3. Numerical results are provided to verify the effective-ness of the NN based learning scheme on the estima-tion accuracy and convergence speed.The rest of this paper is organized as follows. We intro-duce the system model in Section 2. The extended KF basedmethod is introduced in Section 3. Section 4 describes theproposed NN based model. The simulation results are ana-lyzed in Section 5 and conclusions are arranged in Section 6.
2. System Model
In the paper, we assume that there are multiple WiFi accesspoints (APs) with the associated WiFi users transmitting onthe same unlicensed channel and a set of NR users detect
Fig. 1
System model, the NR-U users sense the unlicensed channels be-fore using them. and sense the unlicensed channel before using it, as shown inFig 1. In addition, both the WiFi APs and WiFi users adoptthe channel sensing multiple access with collision avoid-ance (CSMA/CA) based distributed coordination function (DCF) to access unlicensed spectrum. To model the sys-tem mathematically, WiFi APs that use the same unlicensedchannel are represented by set W = { w , w , ..., w N − } ,where N is the number of WiFi APs under the coverage ofNR-U system. All WiFi APs in W employ IEEE 802.11nprotocol. Define a set as N = { n , n , ..., n N − } to demon-strate the number of WiFi users associated with the cor-responding WiFi APs, where n k represents the number ofWiFi users served by WiFi AP w k . Moreover, U is usedto represent the NR cellular users who need to estimate thenumber of WiFi users on the unlicensed channel. For allusers in U , the detection process is independent, so we fo-cus on a single NR user defined as u hereinafter.2.1 Number of Competing WiFi UsersAccording to the IEEE 802.11n protocol, WiFi systemadopts binary slotted exponential back-off scheme to min-imize the collision probability on unlicensed channel. In[20], Markov chain models have been employed to de-rive the closed-form expression on the collision probability,transmission probability and the system throughput. Basedon Markov chain models, the probability that a WiFi usertransmits a data packet at any time slot is given by τ = 2(1 − P )(1 − P )( G + 1) + P G (1 − (2 P ) m ) , (1)where G and m are the size of back-off contention windowand the maximum back-off contention stage, respectively.These two parameters are fixed in the physical layer accord-ing to IEEE 802.11n. P is the transmission collision prob-ability experienced by each WiFi user, which can be calcu-lated as IN et al.: P = 1 − (1 − τ ) n − , (2)where n is the number of WiFi users competing for the sameunlicensed channel, which is equal to n = (cid:80) N − k =0 n k . Bysubstituting (1) into (2), we can write n as a function of P ,which is n = f ( P ) = 1 + log(1 − P )log(1 − − P )(1 − P )( G +1)+ P G (1 − (2 P ) m ) ) . (3)Since G and m are limited and configurable by the specifi-cation in IEEE 802.11n, equation (3) can be utilized at NRuser u to predict the value of n based on the observation ofthe transmission collision probability on the correspondingunlicensed channel. The key factor affecting the accuracyon the prediction of n is the obtainable value of P .2.2 Transmission Collision ProbabilityAccording to (3), if the value of P can be obtained pre-cisely, the estimation on n may achieve high accuracy. Inthis regard, NR user u first needs to measure the value of P by sensing the unlicensed channel during the WiFi trans-mission. At time slot t , u records the number of busy sub-frames of WiFi transmission on the unlicensed channel as K b,t , and collision sub-frames, K c,t . Let K a denote thenumber of all the observed sub-frames during the wholetime slot. Then, the measurement result of P at time slot t is given by ˆ P t = K b,t + K c,t K a . (4)Define ˆ n t as the measurement result of the number of WiFiusers, which can be derived from (3) with ˆ P t , denoted as ˆ n t = f ( ˆ P t ) . Apparently, due to the measurement error on P in (4), Only a few estimates are unable to reflect the realnumber of WiFi users accurately. Moreover, according to(3), a tiny change on P may cause huge fluctuation on theestimation of n . Therefore, the accuracy on the estimationof P is crucial to obtain n precisely. In order to guaran-tee the performance, the conventional method to detect n ismainly based on the extended Kalman Filter (KF), whereKF is utilized to revise output according to the continuousobservation of ˆ P t to obtain accurate estimation of n . How-ever, there are a large number of circulation loop calcula-tions in the KF based algorithm, which leads to high compu-tational complexity. Another weakness is that the extendedKF can only achieve low accuracy when the number of WiFiusers is large. The algorithm and performance of KF basedmethod will be further analyzed in the next section.
3. Extended Kalman Filter based Mechanism
In order to obtain the number of WiFi users based on thetransmission collision probability measured in (4), the ex-tended KF is the most commonly used mechanism. Therein, the number of WiFi users, n = (cid:80) N − k =0 n k , in the system istreated as the state variable of KF and the collision probabil-ity calculated in (4) is the measurement result used to refinethe estimation on n . The state estimation of the frameworkas well as the output of KF at time slot t is expressed as n kt ,which means the estimation of n by the extended KF. Let V t be the estimation error variance of n kt and the updatingformulas of KF are given by n kt = n kt − + K t Z t , (5)and V t = (1 − K t h (cid:48) ( n kt − ))( V t − + Q ) , (6)herein Z t reflects the error between the estimation result andthe measured value of P , which is computed by ˆ Z t = ˆ P t − h ( n kt − ) , (7)where ˆ P t is the measured value that u measures based on(4) at time slot t and h ( · ) is the anti-function of f ( · ) in (3), h ( n kt − ) represents the estimation value of collision proba-bility according to the number of WiFi users predicted at thelast time slot. K t is named as Kalman gain and is calculatedas K t = h (cid:48) ( n kt − )( V t − + Q )( h (cid:48) ( n kt − )) ( V t − + Q ) + R t , (8)where Q represents the system noise estimated by KF andit is set according to the real case scenario. When the valueof Q is large, the extended KF is sensitive to the variationof the measured results. On the other hand, when the valueof Q is small, system is not sensitive to the noise and tendsto output stable values. h (cid:48) ( · ) is the derivative function of h ( · ) , which reflects the sensitivity of actual measurement.Besides, R t denotes the variance of the measurement on P , which is given by R t = (1 − h ( n kt − )) h ( n kt − ) K a . (9)As mentioned above, the value of Q has obvious im-pact on the prediction results of the extended KF. When thenumber of WiFi users is stable, Q requires to be set small,which can provide accurate and robust prediction. On thecontrary, when the number of WiFi users changes, a large Q value should be chosen to improve the sensitivity of thesystem. In order to realize the adaptive adjustment of Q , cumulative summary (CUSUM) based threshold method isutilized, where Q is set to Q + when the threshold is trig-gered and otherwise set to Q − .The simulation results of the estimation performanceby the extended KF based scheme are demonstrated in Fig. 2and Fig. 3. The back-off contention window size G and themaximum back-off contention stage m of WiFi APs are setto and , respectively. Furthermore, Q + is fixed at IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x T h e nu m b e r o f W i F i u s e r s Q = 0.1 Q = 0.01 Q = 0Real number of n Fig. 2
The estimation results by the extended KF scheme when n issmaller than 12. T h e nu m b e r o f W i F i u s e r s Q = 0.1 Q = 0.01 Q = 0Real number of n Fig. 3
The simulation results by the extended KF scheme when n islarger than 20. and performances with different Q − values are compared inthe simulation results. During the simulation, the number ofWiFi users is changed in every 2000 time slots.According to the performance illustrated in Fig. 2, wecan observe that the extended KF based method can achieveaccurate estimation when the number of WiFi users, n , issmall. In addition, it can be found that the smaller Q − valuecan provide more stable output. However, as the numberof active WiFi users increases,the estimation value becomesunreliable and fluctuates significantly, as shown in Fig. 3.The reason is that when the number of WiFi users is large,the small difference on n will cause tiny variation of P , asshown in Fig. 4, and conversely, the noise generated whenmeasuring ˆ P t will be amplified to affect the estimation on n . As a result, the huge measurement error prevents theKalman filter from outputting smoothly and accurately.In addition, as mentioned above, h ( · ) is the anti-function of (3) and h (cid:48) ( · ) is the derivative function of h ( · ) .However, we are unable to derive the close-form expres-sion of both h ( · ) and h (cid:48) ( · ) from (3). The value of h ( n kt − ) and h (cid:48) ( n kt − ) can only be computed by numerical methods. The number of WiFi users T r a n s m i ss i o n c o lli s i o n p r o b a b ili t y G=32, m=3G=32, m=5G=64, m=5
Fig. 4
The relation diagram of P and n in WiFi system. Fig. 5
The structure of NN based network.
Therefore, a large number of circulation loop calculationsneed to be executed in order to achieve the accurate results,which leads to high computational complexity as well asweak real-time ability. The detailed time consumption torun the extended KF scheme will be analyzed in Section 5.In the next section, in order to improve the predictionaccuracy as well as decrease the computational complexity,an online unsupervised training neural network is utilized toestimate the number of active WiFi users. Particularly, torealize more steady and adaptive detection, a specific lossfunction is deployed to train the neural network.
4. Learning based WiFi Users Number Estimation
To reduce the computational complexity and track thechange on the number of WiFi users adaptively, a machinelearning (ML) based approach is developed in this section.Due to the excellent robustness and the low complexity ofNN, an online unsupervised learning NN architecture is de-ployed to reduce the computational complexity as well aspursue high prediction and tracking accuracy in a congestedWiFi system. The structure of the proposed NN is demon-strated in Fig 5.The duration that user u evaluates P based on (4) andoutputs the prediction of n is denoted as a decision time slotin the proposed approach. Then, at time slot t , the output ofthe forward propagation for NN can be expressed as IN et al.: o t = F ([ i t ] , θ t − ) = n dt , (10)where i t = [ n dt − , ˆ n t ] is the input to the NN. n dt is the es-timation of n by NN at t , θ t represents all the weight andbias parameters in NN and is updated based on the gradientdescent algorithm, which is given by θ t = θ t − − l ∂L t ∂θ t − , (11)where l denotes the learning rate, L t is the correspond-ing loss function as well as the objective function of NN.Since conventional supervised learning based approach re-quire large amounts of labeled data for training and has pooradaptability in a variable environment, we propose an adap-tive objective function to train NN unsupervised in a chang-ing system, here the objective function L t is defined as L t = α ( n dt − ˆ n t ) β ( n dt − n dt − ) . (12)The first item in the right-hand-side of (12) represents themeasurement error while the second item denotes the pre-diction error. Parameters α and β are the weight factorsassociated with the measurement error and prediction er-ror, respectively. When α is set to a small value while β is assigned with a large value, the scale of the loss valueis mainly decided by the second item of (12). As a conse-quence, the output of NN tends to converge to the historicalprediction value, which implies that NN is not sensitive tothe change of n , but tends to be stable. On the other hand,when α is set to a large value while β is small, the mea-surement value ˆ n t has a larger contribution on the loss func-tion and the output of NN will be trained towards ˆ n t , whichimplies that NN is sensitive to the change of the real mea-surement result. Therefore, L t can achieve a good tradeoffbetween the estimation in the previous time slot and currentmeasurement on n .Apparently, if the number of active WiFi users, n , doesnot change, the expectation value of ˆ n t and n dt are bothequal to n , i.e. E { ˆ n t } = E { n dt } = n . However, the numberof active WiFi users changes over time, which causes thatthe expectation of n dt deviates from n . To ensure stable esti-mation as well as sensitive adjustment of the NN, objectivefunction should be changed accordingly. If the WiFi traf-fic is stable, the mean value of the loss is equal to , i.e. E { L t } = 0 . Therefore, to achieve a small L t , the valueof L t should be mainly decided by the historical predictionvalue and the influence of the measurement results shouldbe scaled down. On the contrary, when the WiFi traffic loadchanges, the mean value of loss function will not be untilNN converges to a new stable state. Therefore, during theconvergence stage, the value of L t has to concentrate moreon the measurement. In consequence, parameters, α and β ,should change adaptively according to the variation on theWiFi traffic load. Herein, the threshold mechanism based onCUSUM is also employed to estimate the stability conditionof the system and adjust the parameters, where the thresholdparameter, g dt , is defined as Table 1
Simulation parameters
Parameters ValueMeasurement weight α + /α − . / . Experience weight β + /β − . / . Learning rate l + /l − . / . Threshold activation value e d Tolerance parameter q . g dt = (cid:26) max(0 , g dt − + L t − q ) , if g dt − ≤ e d , L t − q, otherwise . (13)Herein, g dt is initialized to when t = 0 and q is a constantvalue which denotes the tolerance to the estimation error, e d is the constant trigger value that determines whether thestate of the system has changed. Based on this mechanism,if g dt > e d , then α will be set to be a large value α + and β is adjusted to a small value β − . Instead, if g dt ≤ e d , α will turn to a small value α − and β will be changed to β + .Moreover, in order to better apply this adaptive mechanismas well as speed up the convergence process, the learningrate, l of the NN, is also adjusted when the change on thenumber of active WiFi users is detected, which is given by l = (cid:26) l − , if g dt ≤ e d ,l + , otherwise . (14)In the above learning rate updating mechanism, when g dt meets the trigger value e d , denoting that u has detected thechange on the number of active WiFi users, the loss valuecomputed by (12) is amplified and the learning rate, l , willincrease accordingly to train the network fast and efficiently.On the contrary, if g dt ≤ e d , the NN regards that the numberof active WiFi users does not change, therefore, both α and l are set to small values. In consequence, the output of NNtends to be stable and fine-tuned when WiFi traffic load isstable.Besides, to further stabilize the NN output when WiFitraffic load is steady, a specific fully-connected structure ofNN is designed, where the number of neurons decreaseswith the number of layers. In such a structure, most weightand bias parameters lie in the lower layers of the NN. Dueto the chain rule of the back propagation training approach,the gradient of the bottom parameters is smaller. Therefore,during the training process, most parameters will be merelytrained and the whole NN can basically keep steady.The complete learning process via the NN is concludedin Algorithm 1, when t = 0 , i is set to [0 , ˆ n ] . T is the en-tire time period for u to estimate n . Here a time slot includesthe time that u detects ˆ P t and the time to execute the algo-rithm. At time slot t , the forward propagation on NN is firstexecuted to get the output o t = n dt . Then L t and g dt are cal-culated according to (12) and (13), respectively. Based on g dt , NN related parameters, α , β , and learning rate l , are setto the corresponding values. Finally, loss value is updatedand the back propagation algorithm is implemented to trainthe NN. IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x
Algorithm 1
NN-based WiFi traffic load estimation algo-rithm Initialize the structure and the relevant parameters of NN; for t ∈ T do The input of NN is i t = [ n dt − , ˆ n t ] ; The forward propagation of neural network is carried out toobtain the output o t = n dt ; Calculate the loss value on the basis of (12); Renew the value of g t based on (13); if g nt > h then l = l + , α = α + , β = β − ; else l = l − , α = α − , β = β + ; end if Calculate the loss value on the basis of (12);
Train the NN based on gradient descent algorithm; end for
Table 2
NN parameters
Parameters ValueNumber of neurons / / / Active function tanh / tanh / tanh /none Optimizer Adam Optimizer
5. Simulation Results
In this section, the performance of the proposed schemeis verified by the numerical simulations. In the simulationsetup, the parameters related to the proposed NN in SectionIII is given in Table 1. To guarantee the high sensitivity andlow complexity, a four-layer fully connected NN structure isemployed and the related parameters are shown in Table 2.The parameters G and m of WiFi APs are also set to and , respectively. The number of active WiFi userschanges in every time slots.The estimation on the WiFi traffic load by Algorithm1 and the extended KF based method is illustrated in Fig. 6and Fig. 7, respectively. In Fig. 6 where the number of ac-tive WiFi users is smaller than . We can observe that theextended KF scheme with Q − = 0 provides the most steadyand accurate output and the performance of NN is very closeto that of the extended KF scheme. On the other hand, whenthe number of WiFi users is larger than , the estimationaccuracy by the proposed approach is much higher than theextended KF methods, as shown in Fig. 7. As analysed inSection 3, when working with a large number of active WiFiusers (over 20 in our system), the variation of n will causeless change on P and the NN in the proposed scheme isable to better detect this subtle change than the extended KFscheme. Although Q − = 0 remains the steady output ofthe extended KF, it could hardly response to the variation intime. As for convergence, based on the simulation results,we can find that the output of the network is basically fine-tuned, which means the loss value of NN is very small inmost of the time. The loss value only becomes larger afterthe error has accumulated to be higher than the thresholdand then the output of NN will converge to the test value. T h e nu m b e r o f W i F i u s e r s Q = 0.1/extended KF Q = 0.01/extended KF Q = 0/extended KFNN predictionReal number Fig. 6
Simulation result when n is smaller than 12. T h e nu m b e r o f W i F i u s e r s Q = 0.1/extended KF Q = 0.01/extended KF Q = 0/extended KFNN predictionReal number Fig. 7
Simulation result when n is larger than 20. Fig. 8 and Fig. 9 are the simulation results of the to-tal data rate on the channel and the data rate of a singleWiFi user during the above detection period respectively.The detailed simulation model can be found in [11], hereNR-U devices access the unlicensed channel based on listenbefore talk (LBT) mechanism. With the estimation resultof NN, NR-U users can adjust the size of back-off windowto use spectrum resources better while NR-U users withoutthe information of traffic load are designed to use a fixedpart of the unlicensed channel. Specifically, the fixed part isequal to half of the spectrum resources when the total datarate is maximum. Simulation results in Fig. 8 show thatwhether the WiFi traffic load is low or high, the proposedalgorithm can maximize the total data rate on the channel.Moreover, as illustrated in Fig. 9, without the known infor-mation, the WiFi users will lose performance significantlywhen the channel becomes busy, while with the NN predic-tion, the data rate of a single WiFi user can be kept basi-cally the same. It means that when the channel is sensedidle, the NR-U users are able to compete for more spec-trum resources on the basis of ensuring the quality of theWiFi transmission and when the WiFi traffic load is high,the NR-U users will reduce own transmission to achieve theharmonious coexistence with the WiFi system.In terms of computational complexity, without a large
IN et al.: The number of time slots D a t a r a t e / bp s Total data rate under low traffic load
With NNWithout NN 1000 2000 3000 4000 5000 6000
The number of time slots D a t a r a t e / bp s Total data rate under high traffic load
With NNWithout NN
Fig. 8
Total data rates in the unlicensed channel.
The number of time slots D a t a r a t e / bp s Data rate of a single WiFi user under low traffic load
With NNWithout NN 1000 2000 3000 4000 5000 6000
The number of time slots D a t a r a t e / bp s Data rate of a single WiFi user under high traffic load
With NNWithout NN
Fig. 9
Data rates of a single WiFi user.
Table 3
Running time comparison
Simulation method NN KFThe duration of an update iteration . . The duration of a single time slot . . number of loop computations in KF based method, the maincomputing task in Algorithm. 1 is to perform a forwardpropagation and a backward propagation on NN. Therefore,the computational complexity of the proposed approach de-creases significantly than the KF based proposal. Table 3shows the time length of an update iteration and the aver-age duration of a whole time slot on two different schemeswhen n is over . As mentioned before, a whole time slotincludes the time when NR-U user listens to the unlicensedspectrum for K a sub-frames and the time to execute the up-date iteration. The sub-frames on the unlicensed bands canbe described as T s = 192 . µs , T c = 45 . µs , T δ = 20 µs and K a = 100 , where T s is the time duration that the unli-censed channel is occupied by a successful transmission, T c denotes the period of a collided sub-frame on the channeland T σ is the duration of an empty sub-frame. Moreover, theupdate iterations of KF and NN are executed on Intel XEONCPU with . GHz and NVIDIA Quadro P2000 GPU withcalculate ability . , respectively. Multiple results of opera-tion show that the convergence speed of NN is much fasterthan the extended KF based method, which verifies the de-crease on the computational complexity of our proposal.
6. Conclusion
In this paper, an online unsupervised NN based approach to estimate the number of competing WiFi users on the unli-censed bands is proposed, where an NN is trained onlineas a filter to trade off the tested values against historicalpredictions. Compared with the conventional Kalman Fil-ter scheme, the NN based method can achieve better accu-racy and stable output when the system is congested. Be-sides, the computational complexity of the proposed methodis much lower and the operation efficiency is improved sig-nificantly. Accurate estimation of the number of WiFi usersin this work is of great significance to estimate WiFi traf-fic load, since in the NR-U system, cellular users can oc-cupy the communication resources on the unlicensed chan-nel adaptively according to the estimated WiFi traffic loadwithout affecting the performance of original WiFi users.
References [1] H. Cui, V. Leung, S. Li, and X. Wang, “LTE in the unlicensed bands:overview, challenges, and opportunities,”
IEEE Wireless Commun. ,vol. 24, no. 4, pp. 99-105, Feb. 2017.[2]
Study on NR-based access to unlicesed spectrum (Release 16) , TR38.889 V1.0.0, 3GPP, Nov. 2018.[3]
Feasibility Sudy on Licensed-Assisted Access to Unlicensed Spec-trum (Release 13) , TR 36.889, v13.0.0, 3GPP, Jul. 2015.[4] R. Yin, G. Yu, A. Maaref, and G. Li, “LBT-based adaptive channelaccess for LTE-U systems,”
IEEE Trans. Wireless Commun. , vol. 15,no. 10, pp. 6585-6597, Oct. 2016.[5] U. Challita, L. Dong, and W. Saad, “Proactive resource managementfor LTE in unlicensed spectrum: A deep learning perspective,”
IEEETrans. Wireless Commun. , vol. 17, no. 7, pp. 4674-4689, Jul. 2018.[6] H. Song, Q. Cui, Y. Gu, G. Stuber, Y. Li, Z. Fei, and C. Guo, “Coop-erative LBT design and effective capacity analysis for 5G NR ultradense networks in unlicensed spectrum,”
IEEE Access , vol. 7, pp.50265-50279, Apr. 2019.[7] C. Wu, X. Chen, T. Yoshinaga, Y. Ji, and Y. Zhang, “Integrat-ing Licensed and Unlicensed Spectrum in Internet-of-Vehicles withMobile Edge Computing,”
IEEE Network , vol.33, no.4, pp.48-53,July/August 2019.[8] H. Ko, J. Lee and S. Pack, “A Fair Listen-Before-Talk Algorithm forCoexistence of LTE-U and WLAN,”
IEEE Trans. Veh. Tech. , vol. 65,no. 12, pp. 10116-10120, Dec. 2016.[9] B. Chen, J. Chen, Y. Gao and J. Zhang, “Coexistence of LTE-LAAand Wi-Fi on 5 GHz With Corresponding Deployment Scenarios:A Survey,”
IEEE COMMUN SURV TUT , vol. 19, no. 1, pp. 7-32,Firstquarter 2017.[10] V. Maglogiannis, D. Naudts, A. Shahid and I. Moerman, “A Q-Learning Scheme for Fair Coexistence Between LTE and Wi-Fi inUnlicensed Spectrum,”
IEEE Access , vol. 6, pp. 27278-27293, Jun,2018.[11] S. Liu, R. Yin, and G. Yu, “Hybrid adaptive channel access for LTE-U systems,”
IEEE Trans. Veh. Tech. , vol. 68, no. 10, pp. 9820-9832,Oct. 2019.[12] H. Hu, Y. Gao, J. Zhang, X. Chu, Q. Chen, and J. Zhang, “Densityanalysis of LTE-LAA networks coexisting with WiFi sharing multi-ple unlicensed channels,”
IEEE Access , vol. 7, pp. 148004-148018,Oct. 2019.[13] J. Yuan, A. Huang, H. Shan, T. Quek, and G. Yu, “Desgin and anal-ysis of random access for standalone LTE-U systems,”
IEEE Trans.Veh. Tech. , vol. 67, no. 10, pp. 9347-9361, Oct. 2018.[14] L. Zhao, G. Han, Z. Li, L. Shu, “Intelligent Digital Twin-based Software-Defined Vehicular Networks”, IEEE Network, DOI:10.1109/MNET.011.1900587, Jun 2020, Early Access.[15] G. Bianchi and I. Tinnirello, “Kalman filter estimation of the num-ber of competing terminals in an IEEE 802.11 network,”
IEEE IN-
IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x
FOCOM , vol. 2, pp. 844-852, Feb. 2003.[16] F. Qin, X. Dai and J. Mitchell, “Effective-SNR estimation for wire-less sensor network using Kalman filter,”
Ad Hoc Networks , vol. 11,no. 3, pp. 944-958, May. 2013.[17] Y. Zheng, “Kalman filter estimation of the number of competingterminals in IEEE802.11 network based on the modified Markovmodel,”
IEEE IIC-BNMT , vol. 18, no. 3, pp. 438-442, Jan. 2010.[18] H. AlSabbagh, “Kalman filter estimation of the number of com-peting terminals in an IEEE 802.11 network utilizing error pron-channel,”
Int. J. Adv. Comp. Techn. , vol. 2, no. 1, pp. 86-92, Jan.2010.[19] Q. Chen, G. Yu, H. Shan, A. Maaref, G. Y. Li and A. Huang, “Cel-lular Meets WiFi: Traffic Offloading or Resource Sharing?,”
IEEETrans. Wireless Commun. , vol. 15, no. 5, pp. 3354-3367, May 2016.[20] G. Bianchi, “Performance analysis of the IEEE 802.11 distributedcoordination function,”
IEEE J. Sel. Areas Commun. , vol. 18, no. 3,pp. 535-547, Mar. 2000.
Rin Yin (M’13-S’19) received the B.S.degree in Computer Engineering from YanbianUniversity in 2001, China, the M.S. degreein Computer Engineering from KwaZulu-NatalUniversity in Durban, South Africa in 2006, andthe Ph.D. degree in Information and ElectronicEngineering from Zhejiang University in 2011,respectively. From March 2011 to June 2013, hewas a research fellow at the Department of In-formation and Electronic Engineering, ZhejiangUniversity, China. He is now a professor in theSchool of Information and Electrical Engineering at Zhejiang UniversityCity College, China, and a joint honorary research fellow in the Schoolof Electrical, Electronic and Computer Engineering at University of Kwa-Zulu Natal, South Africa. His research interests mainly focus on radioresource management in LTE unlicensed, millimeter wave cellular wirelessnetworks, HetNet, cooperative communications, massive MIMO, optimiza-tion theory, game theory, and information theory. Prof. Yin regularly servesas the technical program committee (TPC) boards of prominent IEEE con-ferences such as ICC, GLOBECOM and PIMRC and chairs some of theirtechnical sessions and reviwer for IEEE TWC, IEEE TVT, IEEE Tcom,IEEE Wireless Communications, IEEE Communications Magazine, IEEENetwork, and IEEE TSP journals.
Zhiqun Zou received the B.E. from Zhe-jiang University, Hangzhou, China, in 2018. Heis currently pursuing the M.E. degree at the Col-lege of Information Science and Electronic En-gineering (ISEE), Zhejiang University. His re-search interests include D2D communications,Wireless communications and Machine learningapplications.
Celimuge Wu (Senior Member, IEEE) re-ceived the M.E. degree from the Beijing Insti-tute of Technology, China, in 2006, and thePh.D. degree from The University of Electro-Communications, Japan, in 2010. He is cur-rently an Associate Professor with the Grad-uate School of Informatics and Engineering,The University of Electro-Communications. Hiscurrent research interests include vehicular net- works, sensor networks, intelligent transportsystems, the IoT, and mobile cloud computing.He is/has been the TPC Co-Chair of Wireless Days 2019, ICT-DM 2019,and ICT-DM 2018. He is also the Chair of IEEE TCBD Special InterestGroup on Big Data with Computational Intelligence. He is/hasbeen servingas an Associate Editor for IEEE ACCESS, the IEICE Transactions on Com-munications, the International Journal of Distributed Sensor Networks,and Sensors (MDPI), and a Guest Editor of the IEEE TRANSACTIONSON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE, theIEEE Computational Intelligence Magazine, and ACM/Springer MONET.
Jiantao Yuan received the B.E. de-gree in electronic and information engineer-ing from Dalian University, Dalian, China, in2009; the M.S. degree in signal and informa-tion processing from The First Research Insti-tute of Telecommunications Technology, Shang-hai, China, in 2012; and the Ph.D degree inthe college of information science and electricalengineer from Zhejiang University, Hangzhou,China. He was with Datang mobile communica-tion equipment co. LTD, Shanghai, China, from2012 to 2013, where he was involved in LTE network planningand opti-mization.He currently holds a post-doctoral position with the Institute of OceanSensing and Networking of the Ocean College, Zhejiang University,Zhoushan, China. His research interests include cross-layer protocol de-sign, 5G new-radio based access to unlicensed spectrum (NR-U) and ultra-reliable low latency communications.
Xianfu Chen received his Ph.D. degreewith honors in Signal and Information Process-ing, from the Department of Information Sci-ence and Electronic Engineering (ISEE) at Zhe-jiang University, Hangzhou, China, in March2012. Since April 2012, Dr. Chen has beenwith the VTT Technical Research Centre of Fin-land, Oulu, Finland, where he is currently a Se-nior Scientist. His research interests cover vari-ous aspects of wireless communications and net-working, with emphasis on human-level and ar-tificial intelligence for resource awareness in next-generation communica-tion networks. Dr. Chen is serving and served as a Track Co-Chair anda TPC member for a number of IEEE ComSoc flagship conferences. Heis a Vice Chair of IEEE Special Interest Group on Big Data with Compu-tational Intelligence, the members of which come from over 15 countriesworldwide. He is an IEEE member.
Guanding Yu (S’05-M’07-SM’13) re-ceived the B.E. and Ph.D. degrees in commu-nication engineering from Zhejiang University,Hangzhou, China, in 2001 and 2006, respec-tively. He joined Zhejiang University in 2006,and is now a Full Professor with the College ofInformation and Electronic Engineering. From2013 to 2015, he was also a Visiting Professorat the School of Electrical and Computer En-gineering, Georgia Institute of Technology, At-lanta, GA, USA. His research interests include5G communications and networks, mobile edge computing, and machinelearning for wireless networks.
IN et al.:
IEEE Communications Magazine special issue on Full-Duplex Communications, an editor of
IEEE Journalon Selected Areas in Communications
Series on Green Communicationsand Networking, and a lead guest editor of
IEEE Wireless CommunicationsMagazine special issue on LTE in Unlicensed Spectrum, and an Editor ofIEEE Access. He is now serving as an editor of
IEEE Transactions onGreen Communications and Networking and an editor of