Received April 14, 2018, accepted June 4, 2018, date of publication June 22, 2018, date of current version August 7, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2847609

Unsupervised Learning Algorithm forIntelligent Coverage Planning andPerformance Optimization of MultitierHeterogeneous NetworkJURAJ GAZDA1, EUGEN ŠLAPAK1, GABRIEL BUGÁR1, DENIS HORVÁTH2,TARAS MAKSYMYUK3, (Member, IEEE), AND MINHO JO 4, (Senior Member, IEEE)1Department of Computers and Informatics, Technical University of Košice, 040 01 Kosice, Slovakia2Center of Interdisciplinary Biosciences, Technology and Innovation Park, P. J. Šafárik University, 040 01 Košice, Slovakia3Department of Telecommunications, Lviv Polytechnic National University, 79000 Lviv, Ukraine4Department of Computer Convergence Software, Korea University, Seoul 02841, South Korea

Corresponding author: Minho Jo (minhojo@korea.ac.kr)

This work was supported in part by the Slovak Research and Development Agency under Project APVV-15-0055 and in part by theEuropean Intergovernmental Framework COST Action CA15140: Improving Applicability of Nature-Inspired Optimisation by JoiningTheory and Practice. The work of T. Maksymyuk was supported by the Ukrainian Government Project (Designing the methods of adaptiveradio resource management in LTE-U mobile networks for 4G/5G development in Ukraine) under Grant 0117U007177. The work ofM. Jo was supported by the National Research Foundation of Korea Grant through the Korean Government underGrant 2016R1D1A1B03932149.

ABSTRACT The densification of mobile network infrastructure has been widely used to increase the overallcapacity and improve user experience. Additional tiers of small cells provide a tremendous increase in thespectrum reuse factor, which allows the allocation of more bandwidth per user equipment (UE). However,the effective utilization of this tremendous capacity is a challenging task due to numerous problems, includingco-channel interference, nonuniform traffic demandwithin the coverage area, and energy efficiency. Existingsolutions for these problems, such as stochastic geometry, cause excessive sensitivity to the pattern of the UEtraffic demand. In this paper, we propose an intelligent solution for both coverage planning and performanceoptimization using unsupervised self-organizingmap (SOM) learning.We use a combination of two differentmobility patterns based on Bézier curves and Lévy flights for more natural UE mobility patterns comparedwith a conventional random point process. The proposed approach provides the advantage of adjustingthe positions of the small cells based on an SOM, which maximizes the key performance indicators, suchas average throughput, fairness, and coverage probability, in an unsupervised manner. Simulation resultsconfirm that the proposed unsupervised SOMalgorithm outperforms the conventional binomial point processfor all simulated scenarios by up to 30% in average throughput and fairness and has an up to 6-dB greatersignal-to-interference-plus-noise ratio perceived by the UEs.

INDEX TERMS Coverage planning, heterogeneous network, self-organizing map, unsupervised learning.

I. INTRODUCTIONMobile networks are currently facing a rapid increase intraffic demand, forcing operators to enhance the existingradio access network (RAN) infrastructure. Conventionalsolutions such as cell splitting and sectoring have alreadyreached their limits in terms of the effective trade-off betweennetwork performance and total cost of ownership [1]. There-fore, a heterogeneous network (HetNet) architecture hasbeen proposed by the 3rd Generation Partnership Project(3GPP) as an acceptable alternative for improvement of the

overall system throughput by introducing multitier coverage.These additional tiers are created by small cells in areaswith particularly high traffic demand [2]. The importantbenefit of the small cells is their convenient deployment inany indoor or outdoor area (e.g., lampposts, metro stations,restaurants, and shopping malls). Moreover, indoor smallcells are frequently purchased by the location owners,who also cover the electricity cost to power these basestations (BSs). This, in turn, results in a considerablereduction of capital expenditures (CAPEX) and operating

VOLUME 6, 20182169-3536 2018 IEEE. Translations and content mining are permitted for academic research only.

Personal use is also permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

39807

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

expenses (OPEX) for the mobile operators, while providinga significant improvement in overall network capacity andcoverage quality [3].

The overall capacity of multitier HetNets is greater thanthat of conventional single-tier RANs, owing to the manyadditional cells, which overlay the coverage of the macro-cell. Because these cells are considerably smaller thanmacrocells, the operator can reuse spectrum among themmore effectively, providing more bandwidth per user and,theoretically, unlimited area spectral efficiency. Moreover,the small cell itself has higher spectral efficiency owing to theimproved quality of shorter wireless channels. This advantageis especially noticeable in the macrocell edge areas, wheremacrocell user equipment (UE) suffers from a low signal-to-interference-plus-noise ratio (SINR) owing to the highdistance from the serving BS. OffloadingUEs frommacrocelledges to small cells can substantially improve user experi-ence and overall network performance [4]. Small cells can beinstalled in any arbitrary location that provides an advantagein flexibility of coverage deployment. However, the locationsof the small cells cannot be completely random because theiruncoordinated deployment could result in severe interferenceamong cells, especially in dense scenarios. It is importantto consider the area traffic demand during coverage plan-ning to ensure the highest gain from any additional smallcells. Therefore, optimized planning of HetNet coveragemust be applied by the operator to avoid semi-blindplacement of small cells and ensure the best networkperformance [5].

Typically, the deployment of small cells follows localdesign guidelines, which means that there are no strictrequirements for electromagnetic compatibility or exposurelimits to electromagnetic radiation because of low trans-mission power. Thus, installation of a small cell is similarto the installation of an additional Wi-Fi access point andshould not be considered as a substantial CAPEX, suchas that of a macrocell. However, technical and regula-tory requirements remain, which must be considered beforeroll-out of a HetNet [6]. First, the network capacity must beplanned considering the existence of the shared and neutralhost infrastructure and the availability of a reliable back-haul connection, covered by novel flexible deployment stan-dards and regulations. Nevertheless, solving the optimizationproblem of determining the best network configuration interms of throughput and coverage can be computationallyintensive and not feasible for dynamic systems such asHetNets [7]. The performance of a HetNet is extremely sensi-tive to parameters such as the UE location, mobility, andtraffic demands. For example, a small cell does not provideany significant advantage in terms of capacity if currenttraffic demands are less than the capacity of the macrocell.Therefore, it is important to consider the geographic distri-bution of the traffic demands around the target coveragearea to determine where to install the small cells. Thecomplexity of optimal HetNet deployment is in the randommovement of the UEs, which makes it difficult to predict

the areas with the highest average traffic demand and toidentify the most profitable allocation for the small cells.In this paper, we address this complex problem and proposea self-learning algorithm for optimization of HetNet deploy-ment using self-organizing maps (SOMs). The proposedalgorithm employs two mobility patterns that reflect thefeatures of pedestrian and vehicular UE. Based on thesimulated behavior of the UEs, the proposed algorithmattempts to deploy small cells in areas with higher loads,while considering the presence of co-channel interferenceamong the small cells. In general, an SOM represents anunsupervised learning technique based on neural networks,which was introduced by Kohonen [8]. Further extensionsof SOMs are presented in [9] and [10]. SOM networks havedemonstrated excellent performance in various applications,such as document collection [11], image processing [12],and security-threat detection [13]. Recently, SOMs havebeen applied as the basic algorithm for the event-sensordeployment problem with random distribution of events [14].In [15], Salem et al. apply the SOM algorithm to effectivelyminimize the probability of sensing errors in dynamic spec-trum access networks. A detailed and comprehensive surveyof different SOM applications in wireless communications isprovided by Zhang et al. [16].

In this paper, we follow the conclusions regarding theefficient deployment of emerging HetNets based on artificialintelligence as presented in [17]. We utilize the unsuper-vised SOM algorithm for solving both HetNet coverage plan-ning and performance optimization. Unlike previous studies,where random point processes (e.g., Poisson and binomial)are used to simulate the mobility of the UEs, we proposea combination of two different mobility patterns based onBézier curves and Lévy flights. The former represents aperiodic semi-dynamic pattern, which reflects the movementin public transport, and is interchangeable with the periods offixed user positions, such as at home or at the office. The lattermodel employs a Lévy flight pattern with truncated power-law flight lengths [18]. The Lévy flight is a series of shortmovements with random longer movements. Such patternshave been observed in human behavior in Song et al. [19].Thus, the proposedmobilitymodel is noisier andmore naturalthan a conventional random point process. The proposedapproach targets explicitly topologically correct mappings ofthe spatial probability density distributions of high demandareas.

The major contributions of this paper can be summa-rized as:• The new intelligent algorithm for HetNet coverage plan-ning and performance optimization is proposed based onunsupervised SOM algorithm.

• The new realistic UE mobility pattern has been devel-oped based on the combination of Bézier curves andLévy flights random movement.

• Comprehensive simulations have been conducted fordifferent number of small cells within the targetcoverage area.

39808 VOLUME 6, 2018

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

• Practical recommendations for HetNet coverage plan-ning have been derived by finding the best trade-offbetween average throughput per UE, coverage proba-bility and fairness of resource allocation.

The remainder of this paper is organized as follows. Section IIprovides an overview of the recent related work on multitierHetNet coverage design. Section III presents a detailed expla-nation of the investigated systemmodel andmobility patterns.Section IV describes the proposed SOM-based algorithm foroptimization of the deployment of small cells. Simulationresults and performance analyses are presented in Section V.Section VI concludes the paper.

II. RELATED WORKThe deployment of a two-tier HetNet consisting ofmacrocellsand small cells has attracted significant attention recently formany reasons, including improved user throughput, reducedlatency, and overall energy savings. In this paper, we focusmainly on the optimization of small cell deployment withinthe existing coverage of macrocells. Therefore, this sectionaims to advance our research in the context of existingworks in this area. Although there are many research papersregarding the advantageous characteristics of small celldeployment, specific research targeting small cell optimiza-tion for outdoor scenarios remains limited.

We must acknowledge that the simulation of two-tierHetNets, where the investigation of different positionsand transmission power of a small cell within the cellcoverage of the macrocell, was conducted in [20]. Theresults rely mainly on the brute-force-based investigationof the various parameters, with a general recommendationregarding the small cell settings (large macro-to-small celldistances, relatively high transmit power, and high deploy-ment density). Shin and Zain [21] investigate the coverage ofthe geo-clustering algorithm in HetNets, where small cellsare added to complement the coverage holes of the macrocellcoverage. This approach performs static clustering of networkcoverage without consideration of UE mobility. A sampling-based optimization method for small cell deployments withthe objective of maximizing UE throughput is introducedin [22]. More general objectives such as maximization ofthe traffic offloading to small cells or minimization of thecost of service delivery are targeted in [23]. However, withinthe scope of the above-mentioned research, no specific UEdynamics have been considered that limit their application ingeneral.

Applications that exploit the self-organizing principle incombination with the dynamic behavior of UEs are investi-gated in [24] and [25]. In [24], Park et al. propose an algo-rithm for outdoor small cell deployment that considersmutualinterference and load variation among different coveragezones. The algorithm is based on the weighted sum ofobjectives, which belongs to the class of multi-objectiveoptimization methods. Another study is provided in [25],where the authors analyze the impact of the imperfectsmall cell positioning concerning dynamic hotspots (HSs).

Grid search based onmetric calculations for several candidatelocations of small cells is proposed and evaluated in [26].Nomadic spatially quasi-static UE distribution is consideredthroughout the simulation scope. Zhou et al. [27] investigatethe optimum location of small cells given a specific UE trafficpattern. They assume predefined probable locations for thesmall cells, and thus they are not obliged to search through theentire parametric search space, which reduces the complexityof the algorithm.

However, the existing (primarily static) model solutionsremain significantly distant from realistic dynamic situationsinvolving continuous changes in the physical locations ofthe UEs. Therefore, the dynamics we study herein considernot only the specific deterministic motion modulated byperiodic cycles, but also the noisy, unpredictable influenceof UEs following Lévy flight motion. We assume that thetime-varying and spatially dependent requirements of theUEs can be achieved, for example, using appropriate heuris-tics. Considering the results obtained, the expansion of theportfolio of the existing heuristic approaches to the problemof optimizing small cells in HetNets is also expected.

III. PRELIMINARIES AND SYSTEM MODELIn this section, we outline the system model considered inthis paper, the investigated metrics characterizing the impactof HetNet deployment optimization on theUE experience, thetrafficmobility patterns considered throughout the simulationscope and finally, the model simplifications.

A. SYSTEM MODELWe consider the downlink transmission of a two-tier HetNet.The tiers operate at different frequencies, i.e no cross-tierinterference is observed. Nevertheless, interference amongBSs of the same type is present (i.e., co-channel inter-ference). Whereas the traditional macrocell layer broad-casts traffic on a dedicated 2100 MHz carrier, small cellsutilize the 3350–3650 MHz band. The reason for this derivesfrom the recent FCC release document Notice of ProposedRulemaking, where the standard development bodies issued100 MHz of available spectrum to be used by small cells inthe 3350–3650 MHz band [28]. This bandwidth is character-ized in general by high penetration losses, and thus is notfeasible for long-distance transmission. However, small cellscan exploit the available spectrum for providing services andto offload traffic from the macrocells.

In our simulation setup, the BSs of both layers (i.e., macro-cells and small cells) use the round-robin resource schedulingalgorithm with full allocation, which shares the availableresources among all UEs. Each UE is associated with aBS based on the calculated SINR, which provides improvedresults compared with conventional measurements of refer-ence signal receiving power (RSRP).

We consider a model where the two-tier HetNet consistsof NM macrocells with dedicated positions and NS smallcells randomly distributed in the macrocell coverage area.Each macrocell has access to NRB resource blocks (RB)

VOLUME 6, 2018 39809

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

(NRB = 100); for the small cells, NRB = 20 as proposed by3GPP in [29]. Denoting the number of all BSs (macrocellsand small cells) to be deployed as NBS, the total number ofBSs is given by NBS = NM + NS. Depending on where theUE is associated, we classify the UE as a macro-UE (theUE is associated with a macrocell) or small-UE (the UE isassociated with a small cell).

Let us denote the positions of the small cells and macro-cells by yi ∈ R2, where i = 1, 2, . . . ,NM + NS. TheUEs, following a specific mobility model, are dispersed inthe investigated region and their locations are denoted byxj ∈ R2, where j = 1, 2, . . . ,Nu. Here Nu represents thenumber of UEs in the region. Assuming that the channelon each RB experiences independent and identical Rayleighfading, the channel power gain of the link between the j-th UEand associated i−th BS in a RB is expressed as

gij(t) = Aij gf,ij(t) gpl,ij, (1)

where Aij denotes antenna gain; gf,ij(t) is the exponentiallydistributed fading gain with unit mean representing the small-scale channel fading gain at time t; and gpl,ij is the pathloss. Here, we consider the ECC-33 channel propagationmodel, as it demonstrates the closest agreement with themeasurement results in the frequency range 2000–3500 MHzand is common for urban environments. The path-loss for theECC-33 model can be defined as in [30]:

gpl,ij = Afs + Abm − Gb − Gr[dB], (2)

where Afs, Abm, Gb, and Gr are the free space attenuation,the basic median path loss, the BS height gain factor, and UEheight gain factor, respectively. These elements can be furthercharacterized as

Afs = 92.4+ 20 log10(d)+ 20 log10(f ),

Abm = 20.41+ 9.83 log10(d)+ 7.894 log10(f )+

+ 9.56[log10(f )]2,

Gb = log10(hb/200){13.958+ 5.8[log10(d)]2},

Gr = [42.57+ 13.7 log10(f )]

× [log10(hr)− 0.585 ], (3)

where f denotes the frequency in GHz; hb is the antennaheight; hr is the height of the UE; and d is the Euclideandistance between the associated cell-UE pair. The instanta-neous received SINR of the dedicated RB of the j-th macro-UE located at xj, associated with the macrocell at location yiat time t can be expressed as

γij(t) =Pigij(t)

NM∑k=1,k 6=i

Pkgkj(t)+ σ 2

=Pigij(t)Iij + σ 2 , (4)

where Pi is the transmit power of the i-th macrocell; σ 2 isthe additive white Gaussian noise variance; and Iij is theinterference power received by macro-UE j from macrocellsother than the i-th macrocell.

For a small-UE, we have the similar description:

γij(t) =Pigij(t)

NS∑k=1,k 6=i

Pkgkj(t)+ σ 2

=Pigij(t)Iij + σ 2 , (5)

where Pi denotes the transmition power of the i-th small cell;and Iij is the interference power received by the small-UEj from neighboring small cells.

The throughput of the jth macro-UE when associated withthe i-th macrocell at time t (the aggregated value accumulatedthrough all assigned RBs) can be defined as

0ij(t) = WNRB∑k=1

log2

(1+

Pigkij(t)bkij

I kij + σ2

), (6)

where bkij = 1 if the k-th RB of the associated i-th macrocellis served to the j-th UE (otherwise bkij = 0); W denotes thebandwidth per RB; gkij(t) and I

kij denote the channel power

gain of the link between the j-th UE and the associated i-th BSin the k-th RB and interference received in the k-th RB,respectively. The variableNRB

ij determines the number of RBsassigned to the j-th macro-UE in the i-th macrocell as follows:

NRBij =

NRB∑k=1

bkij =⌊NRB

Nc

⌋, (7)

where b. . .c denotes the floor function, and Nc determinesthe number of macro-UE associated with the i-th macrocell(Nc � Nu). This concept ensures the proportional fairnessdistribution of the resources among the UEs of macro cells.associated with the i-th macrocell (round-robin scheduling).A similar assumption is derived for the UEs of small cells.

Throughout this paper, the network model is assumed tooperate in a saturated mode where all BSs (both macro-cells and small cells) transmit at their maximum powersetting. This represents a worst-case scenario, and is atypical methodology used for network capacity dimen-sioning, as suggested in [31]. By applying the full-load condi-tion in the network, the various deployment solutions arepushed to their ultimate limits in a systematicmanner. Finally,we assume that UEs are capable of forwarding the trafficonly over one cell at a time (either macrocell or small cell,depending on the conditions), which conforms with the LTEstandard without the advanced carrier aggregation feature.

B. METRICSWith the objective of analyzing the impact of HetNet deploy-ment optimization, we now introduce various metrics thatallow us to numerically quantify our solutions. In additionto the most intuitive metric, the throughput 0ij, we havealso chosen to consider fairness and coverage probabilitiesto achieve more general conclusions.

In wireless communication networks, all UEs expect tohave uniform quality of experience (QoE) across the wholenetwork coverage. If UEs in the network do not experiencea similar QoE (expressed using, for example, the throughput

39810 VOLUME 6, 2018

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

FIGURE 1. Schematic diagram of a HetNet model with a set of macrocelland small cell stations, and mobility patterns of HSs and UEs.

metric), the system is less fair and thus, we must introduce afunction that quantifies the fairness. Here, we use Jain’s indexdefined as

F({0}) =

(∑Nuj=1 0j

)2Nu∑Nu

j=1 02j

. (8)

For readability, we drop the index i from the definition of0ij as we are interested in particular UE contributions to thefairness regardless of the associated BS. Jain’s index has thefollowing characteristics [32]:• The index provides values in the interval (0, 1], where‘‘1’’ means a completely fair allocation (all UEs havethe same SINR value) and value close to ‘‘0’’ meanstotally unfair allocation (single UE has highest SINRvalue, while other UEs are out of coverage), respectively.

• The fairness is independent of the number of UEs activein the system.

• The fairness is a continuous function. Any change in theperceived {0} is reflected in the fairness.

The coverage probability pc determines the mean fraction ofUEs that, at any given time, achieve an SINR greater than thegiven threshold β, i.e.,

pc =1Nu

Nu∑j=1

1γj≥β , (9)

where 1(...) represents the indicator function.

C. TRAFFIC MOBILITY MODELSIn recent years, a number of techniques have been developedto model different transport mobility situations. The mainidea behind the application of mobility models is that theoutputs (i.e., trajectories) depict synthetic data that repre-sents the realities of transport. One of the key points of ourpresent work is the extension and application of mobilitymodeling principles [33]. This is mainly the formal decom-position of motion into its regular (trending or periodic) and

random components. Furthermore, we focus particularly onthe assemblies, including periodic and stochastic trajectories.It is clear that long-term stationarity is essential for the qualityof SOM learning required to provide the (static) small celllayouts. In our applications, we consider the regular transportof UE clusters (semi-dynamic HSs) having a typical transportperiod of half an hour. Regular trajectories are parameterizedby the function composition consisting of the Bézier curves,and logistic and harmonic functions. Independently generatedLévy flights are used tomimic the unpredictable and intermit-tent two-dimensional (2D) moves. An additional assumptionis that the simulated entities do not change their manner ofmovement from random to regular, or vice versa.

1) REGULAR MOBILITY, PERIODIC TRAJECTORIESAS MODULATED BÉZIER CURVESWe describe the periodic regular movements of thesemi-dynamic HSs. HSs can be used to calculate smoothpathways between obstacles using a Bézier curve character-ized by control points. The application of Bézier trajectoriesis advantageous in describing the movement pattern withcollision-free and obstacle constraints [34]. The basic ideais to create an assembly of smooth trajectories consisting ofa system of regular visits to predetermined starting, target,and intermediate points. The intermediate points (only forthe Bézier curves of order higher than two) are used tocontrol trajectory curvature. Assume that M semi-dynamicHSs enumerated as j = 1, 2, . . . ,M are dispersed in 2D.The simplified assumption is considered here, so that eachHS consists of a constant number of UEs. We introduce 1tas the time step used to generate the mobility data. For agiven simulation time interval [ 0,Tsim1t] that is greater thanthe period of motion T (i.e., Tsim1t � T ), a system ofnumerical trajectories can be generated that is parametrizedby the modulated quadratic Bézier curves {xj(t) ∈ R2;t ∈ [0,Tsim1t]; j = 1, 2, . . . ,M } with the particulartrajectory

xj(t) = (1−8j(t))2 xinit,j+ 2(1−8j(t))8j(t) xinterm,j+82

j (t) xtar,j. (10)

Its structure depends on the temporal modulation medi-ated by the periodic activation functions {8j(t) ∈ [0, 1];t ∈ [0,Tsim]; j = 1, 2, . . . ,M }, where

8j(t) =1

1+ exp[−a1 − a2 sin

(2π (t+tj)

T

)] . (11)

The persistent (time-invariant) properties of the HSs can becharacterized by a system of four-tuples{⟨

xinit,j, xinterm,j, xtar,j, tj⟩; j = 1, 2, . . . ,M

}, (12)

where xinit,j ∈ R2 determines the starting point of the j-th HStrajectory and xtar,j ∈ R2 is the vector directed to thetarget destination. An additional aspect is the control point,

VOLUME 6, 2018 39811

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

FIGURE 2. Time evolution of activation function 8(t).

xinterm,j ∈ R2, which can be indirectly used to alter the curva-ture of xj(t). Each 8j(t) is determined by the HS-specificphase shift tj; all trajectories are characterized by the posi-tive real-valued parameters T , a1, and a2. The parametricpair (a1, a2) determines the extent to which the logisticform tunes the periodic activity. Some important elementaryfacts must be considered before parametrization is suggested.As observed from (11), the choice a2 � a1 pushes 8jto the vicinity of asymptotic (boundary) values zero andone, which means that the limit points xj|8j→0 = xinit,jand xj|8j→1 = xtar,j are never attainable. For some typicalintermediate situation 8j = 1/2, the weighted averagexj|8→1/2 = (xinit,j + 2xinterm,j + xtar,j)/4 is obtained, whichunderlines the known fact that xinterm,j does not necessarilylie on the trajectory xj(t). In Fig. 2, we depict an illustrativeexample of 8j(t), important for the parametric choices andfurther simulations.

2) LÉVY FLIGHT MOBILITYAnother variant of motion is provided by the Lévy model,where, in general, human mobility has been empiricallyobserved to exhibit Lévy flight [19]. This combined motionis based on the following interpretation. If a UE travels toa certain relatively distant destination (in the city), the bestdirection of movement appears to be a straight line. Paradox-ically, moving closer to a target point causes the informationregarding the exact location of the UE to become vague;hence, another local search (though random) is required.In this precarious situation, a random strategy for the nextroute choice becomes the most efficient. The Lévy flightprocess arises as a repeated application of these two move-ment strategies. As UE moves are considerably subjec-tive, and therefore unpredictable and interrupted by variabletargets, we represent them through 2D Lévy flights.

In this segment, we present essential simulation details thatillustrate this situation. Under the assumption that angulardecisions are distributed uniformly and randomly, at eachsimulation step, the random angle of the j-th UE is gener-ated by θLFj (t) = 2πrang,j(t), where rang,j(t) is drawnuniformly from [0, 1]. The instantaneous length of each iter-ative step is obtained as the product of the typical spatial steplength `u and the random multiplier r−1/αLFstp,j , where αLF isthe stability parameter and rstp,j(t) ∈ [0, 1] is independently

and uniformly generated. The position xLFj (t) ∈ R2 (LF is thedistinguishing label for a UEwith this type of mobility) of thej-th UE is updated iteratively according to

θLFj (t) ← 2πrang,j(t),

xLFj (t +1t) = xLFj (t)+

+ `u

(cos(θLFj (t)), sin(θLFj (t)

)2D

[rstp,j(t)]1/αLF, (13)

where (., .)2D is used to represent the vector in 2D.An additional simulation detail is that a Lévy type of UE is

not permitted to move a step that overcomes the rectangularboundaries of the system (i.e., no wrapping is assumed).Of course, the accumulation of limited movements couldcause a marginal increase in the concentration of UEs in aborder region. These subtle consequences are left for furtherstudy.

D. SYSTEM MODEL SIMPLIFICATIONSAlthough the system model is designed to reflect the realHetNet as precisely as possible, there are still some simplifiedassumptions, which were introduced to decrease the compu-tational complexity of the proposed algorithm.• The spectrum bands of macrocells and small cells arecompletely separated to simplify the task of interferenceaware resource allocation for small cells. Such simplifi-cation is common for the real network deployment, espe-cially with the advent of LTE-Unlicensed technology,which assumes utilization of unlicensed spectrum for thesmall cells.

• Round-robin resource scheduling is assumed to ensurethat all simulation results will solely depend on thepositions of small cells, rather than on other randomparameters. Such simplification allows to ensure thatSOM algorithm will find the HetNet topology with themost uniform QoE across coverage area.

• In real network scenario, UEs accept the QoE, whichis not lower than some minimum threshold. In thispaper, we assume that QoE requirements are the highestpossible for each UE. This allows us to ensure that eachUE will tend to maximize his throughput as much aspossible. Such assumption is important for SOM algo-rithm to find the most effective network topology.

• In this paper, we limit our investigation to the two-tierHetNet architecture with fixed transmission powers ofall macrocells and small cells. This allows us to reducethe algorithm complexity only to small cells positioning,which is the main goal of this paper. In our futureresearch we will extend this problem to more tiers ofcoverage and flexible power adjustment of BSs, in orderto reflect more realistic network scenarios.

• ECC-33 wireless channel propagation model is assumedas the best suitable for the chosen frequency range andurban environment. In real scenarios, the channel prop-agation models are different across coverage area due to

39812 VOLUME 6, 2018

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

different type of landscape, height of buildings and linkdistances. However, this is very complex to simulate,so most of similar studies usually focus on the singlechannel propagation model.

• TheUEs’ mobility pattern is assumed to be the combina-tion of Bézier curves and Lévy flights randommovementto reflect more realistic behavior of UEs. Nevertheless,this pattern is still far from the real UEs’ mobility,because it does not reflect the typical UEs’ mobility inthe big city. In our further research, we will implementmore realistic patterns for UEs’ mobility based on themachine learning algorithms.

IV. SELF-ORGANIZING MAP ALGORITHM:PROPOSED ALGORITHM FOR HETNETDEPLOYMENT OPTIMIZATIONIn this section, we discuss the details of the application ofa known SOM network to the problem of learning, wherethe parameters of the neurons define the coordinates of thecells of a small cell network.1 If we know the trajectoriesof the UEs, then the SOM-mediated reduction projects theseinput data to the significantly reduced, yet representative,locations that are associated with the small cell positions inour application. Because the learning process is unsupervised(self-organized), it can reveal a hidden data structure, thatis, it can perform categorization without a priori knowledgeof the appropriate solutions. The ability of SOM networksto achieve self-organized states is equivalent to the ability toform clusters of input data. In our case, the data represent themoving coordinates of the UEs, and the persistent positionsof the small cells belong to the locations of the neurons(or centers of clusters) forming the SOM content. To harmo-nize the concepts of the proposed model with the standardneural network dictionary, we can say that the neuron (whichencodes the location of the small cell) whose position isclosest to the data entry (for example, a UE position) is thebest matching unit (BMU neuron). After each SOM iteration,the BMU vectors are modified; they are trained linearly tobe closer to the appropriate data inputs. What is consideredto be ‘‘closest’’ in the context of the problem being studied,of course, can be interpreted in different manners and is theprimary idea and specificity of the proposed SOM applica-tion. In many cases, the proximity of the vector objects cansimply be quantified by their distance defined for the metricspace. Euclidean distance is frequently used as the stan-dard in this area. However, the distance based on the purelygeometric properties acquired by the conventional Euclideanprism is not suitable in cases where (i) heterogeneousproperties (e.g., those associated with signal propagation)become important; (ii) other parties influence the processof indirect distance evaluation, e.g., based on the signaltransmission between entities; (iii) if there is a manifestly

1In this paper, we use the terms neuron and small cell interchangeably,as they represent the same element within the algorithm framework.

functional asymmetry between a pair of entities (such asa transmitter-receiver) that violates symmetry, and thusmakes the applicability of the distance concept difficult.Therefore, we propose a modified SOM application wherethe immediate SINR in (4) is employed instead of the tradi-tional Euclidean distance. Based on the previous discussion,we assume that the SOM application in the HetNet networkcan be useful not only to improve HetNet coverage but alsoto explore other investigated metrics. In the next segment,we will briefly review the SOM preliminary studies andprovide details of the SOM application in the HetNet deploy-ment optimization process.

In the proposed SOM version and other implementa-tions, the network consists of a set of neurons arrangedon a 2D regular square mesh. To obtain an optimized(self-organized) 2D structure of the neurons (small cell equiv-alents), an iterative learning process is required. Syntheticdata inputs presented to the SOM network are the trajectorysamples of the UEs, which are partially random (xLFj (t)) ordeterministic (xj(t)) in nature. In the following text, only xj(t)symbols are used to represent the synthesized patterns forSOM learning. It is also worth mentioning that the standardtheory of learning recognizes two types of training/learningpractices, called sequential and batch algorithms. In thiswork, we use a sequential learning algorithm because itretains the time sequencing of the samples, which can beimportant for the analysis of periodic UEmovements. Finally,we do not propose changing the macrocell topology structure,and thus the macro-UEs (the UEs that are associated at partic-ular learning time steps with the macrocell) do not have anyinstantaneous impact on the SOM learning process.

The parallel dynamics of the moving UEs and the learningdynamics of the SOM are linked by specific time variablest ′ and t , where t ′ is a discrete-time variable describingthe stages of SOM learning; the positions of the UEsacquired within the mobility models (mobility curves) areparametrized by the continuous (real) time variable t . Themutual relation between t and t ′ is given by t ′ = bt/1tc,where 1t is the time step of the mobility model sampling.

In the following segment, we discuss the compu-tational details of our specific SOM implementation.An SOM network is a recurrently evolving system ofvectors–neurons S(t ′) ≡ {mi(t ′) ∈ R2; i = 1, . . . ,NS}.The mobility models provide instantaneous data samplesincluding {x1(t), x2(t), . . . , xNu (t)} consisting of Nu UE posi-tions at a given time. The single-learning step for the occur-rence and processing of data item xj(t) includes:(i) the introduction of particular input xj(t) to the SOM. Herewe emphasize the neuronal dependence not only on t ′ but alsoon the pair (j, t ′), which stems from the fact that for serialupdate, the learning depends not only on t ′ but also on theorder j of the respective data item xj(t) (i.e., data items arevisited in a typewriter manner).(ii) the identification of the BMU vector mc(j,t ′) ∈ S(t ′)specified by its index c(j, t ′) ∈ {1, 2, . . . ,NS};

VOLUME 6, 2018 39813

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

(iii) the adjustment of mc(j,t ′) and neighbors mi(t ′) ∈ S(t ′)with the learning rate modified by the neighborhood functionhi,c(j,t ′) ∈ R+ (the form is specified below). The processesof BMU identification and the calculation of hi,c(j,t ′) dependon the distance (or SINR-based alternative) between thedata samples and neurons. Given ‖ . . . ‖2D is the Euclideandistance in 2D, then c(j, t ′), the index representing the BMU,is identified using

c(j, t ′) ≡ arg mini=1,...,NS

‖ xj(t)−mi(t ′) ‖2D. (14)

At this stage, neurons begin adapting their components inresponse to changes in their own network topology andchanges in the data environment

mi(t ′ + 1) = mi(t ′)

+α(t ′)hi,c(j,t ′)(t′)[xj(t)−mi(t ′)

], (15)

where α(t ′) is the learning rate and i = 1, . . . ,NS. However,not all neurons necessarily require modification; rather, localchanges are typical for the SOM architecture. The time-varying limitations for the S(t ′) update are defined at the levelof the neighborhood function

hi,c(j,t ′)(t′) ≡

{1, if ‖mi(t ′)−mc(j,t ′)‖2D ≤ σ̃ (t ′)0, else.

(16)

The convergence toward a static limit S∗ is achieved by thedecreasing learning rate (varying in a simulated annealingmanner)

α(t ′) = αstart k t′

1 , k1 =(αtar

αstart

)1/Tsim, (17)

where k1 ∈ [0, 1] is a constant and Tsim � T is a typicaltime scale, which at the same time characterizes not only thesimulation of mobility but also the slowing of the learning(simulation length). The learning rate is delimited by the posi-tive constants αstart and αtar. The complementary heuristicσ̃ (t ′) = k2α(t ′) is used for the process of the ‘‘narrowing’’of the network focus. Because k1 and k2 have a significantimpact on the convergence and results achieved, several meta-optimization tests have been applied to perform the selectionof these constants.

If no algorithmic optimization (or pre-processing of inputdata) is performed for the respective SOMvariant, the compu-tational complexity of each iteration of the network can begiven as O(NSN 2

u ) [10]. The above discussed literature (andreferences therein) reports that the main options to improvethe SOM performance may be available in situations wherethe data environment is relatively sparse. We can mentionthat the computational load due to N 2

u contribution can besimply reduced by representing the spatially dispersed UEsas the centroid using appropriate cluster mechanism. In lieuof detailed valuation of all UEs, we assume that some pre-treatment procedure from the class of coarse-graining proce-dures can also be applied.

We refer to the procedure presented above as the conven-tional application of SOM to the HetNet deployment opti-mization problem. However, as long as the Euclidean distancein (14) is the decisive factor for selection of the BMU, theEuclidean scenario finds somewhat limited application in theHetNet deployment problem. The problem of the optimaluser-cell association can be extremely pronounced in thespecific situations where the optimality UE association isachieved for a remote (in the Euclidean sense) BS. The reasonfor this seemingly paradoxical choice is the ability to gain ahigher SINR of the receiver. In general, a remote BS can havea more favorable position in terms of ubiquitous interference,and can thus provide higher throughput to the associated UE.

Because this phenomenon cannot be captured using theconventional Euclidean metric, we choose to use the SINR‘‘quasi-metric’’ formula. Our presented heuristic modifica-tion of the SOM is limited to one calculation step. To identifyc(j, t ′), we use the following modified variant of (14):

c(j, t ′) ≡ arg maxi=1,2,...,NS

fSINR(xj(t),mi(t ′)), (18)

where fSINR(., .) represents the SINR function (see (5))between the j-th UE and the i-th neuron (encoding the posi-tion of the small cell). Although this may not be optimal,the next steps of the SOM algorithm remain consistent withthe conventional application. Regarding the modified form,the most serious doubts arise in connection with the relaxedmetric properties that are irrevocably lost when the Euclideandescription is changed to fSINR(., .). Although the modifiedapproach does not provide sufficiently precise arguments,this deficiency can be compensated by the empirical-computational, intuitive, and comparative approach (wheree.g., a comparison of the conventional and modified SOMsis performed in the specific cases). A mathematical intuitioncan be based on a return to the metric concept. The inverseproportionality 1/‖xi − xj‖SINR2D ∼ 1/fSINR(xj,mi) is anelementary idea in this direction.

V. SIMULATION RESULTSTo evaluate the performance of the proposed approaches(based on the Euclidean metric and SINR metric), wedistributed NS small cells following the binomial pointprocess (BPP) in an investigation region having an areaof 1 km2. A simple 2D BPP is typically used to modelthe positions of small cells of single tier or K-tier cellularnetworks. Unlike other stochastic geometry tools, the use ofa BPP is becoming prevalent in coverage-limited scenariosto model a finite number of small cells distributed withinthe coverage area. As HetNets are commonly deployedin an increasingly irregular and random manner, modelingthe locations of the small cells using tools from stochasticgeometry would seem to be a fair solution to model theHetNet network performance. In general, the application ofthe BPP modeling approach in our discussion is twofold.First, the BPP modeling approach serves as a referencepoint for a performance comparison with the proposed

39814 VOLUME 6, 2018

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

SOM application. Furthermore, a BPP also provides theinitial HetNet topology for the SOM deployment optimiza-tion process.

The numerical tests were conducted using the 3GPPLTE-A specification and small cell parameters as describedin [23] and [29] using Monte Carlo simulation. Throughoutthe simulation, we followed scenario #2a presented in [35],where the outdoor small cell deployment was consid-ered jointly with the overlaid macrocell network. Separatefrequency deployment of the macrocells and small cells wasassumed. The detailed simulation parameters are introducedin Table 1.

TABLE 1. Simulation parameters.

We first discuss in detail the specific manner in which thepositions of the neurons were changed via sequential moves.Here, we illustrate the use-case with a Euclidean-based SOM;however, similar results were achieved for the SINR case.Even a visual inspection can indicate that the dynamicsgoverning the radius of learning can have an important rolein the efficiency of the convergence process. We initiated theSOM learning procedure with a neuron distribution followingthe BPP (Fig. 3a). Owing to the fast learning process andinitial large radius, the neurons were grouped together and

jointly traversed the region (see supplementary material)).As we proceeded with the review of the simulation snap-shots, we observed that there were clearly identifiable earlystages of the fast learning process where the instantaneousarrangement fluctuated irregularly with relatively low globalthroughput efficiency. Under this regime, the arrangements ofthe neurons reflected the combined effect of instantaneouslylarge radius of learning (low competition, high plasticity),high learning rate, and a local cumulative effect of the mobileUEs, which can be localized in an extremely narrow regionof the traffic funnel that acts as an intense temporary attractorfor all neurons. Because the effect of stochasticity is ubiqui-tous, the analogy between our variant of SOM and the simu-lated annealing processes is clear. In simulated annealing,the temperature determines the probability of the transition tohigher energy from some instantaneous configuration. In thespecific search for the optimal configurations of neurons,the complex density deviations cause unpredictable quasi-thermal dynamics. The Lévy flight randomness, the regularperpetual motion of the traffic stream, and the funnel togetherwith the gradual non-equilibrium reduction of the learningrate are the dynamical components that form the independenttime scales shaping the environment of the neurons. From anintuitive viewpoint, the timing structure mixing the globaland local trends resembles the standard annealing sched-ules. As can be observed, the overheated neurons wanderedinitially within a broad region of the search space (see video,supplementary material and Fig. 3b). When the learning ratebecame sufficiently slow (the movements of the neuronswere less pronounced), a relatively sharp threshold transi-tion occurred that was accompanied by a qualitative shifttoward high learning (optimization) efficiency and consid-erable structural changes in the arrangement of the neuraldegrees of freedom (Fig. 3c). Structural tendencies supportingaxial symmetry are readily observable. It can be seen thatsmaller, yet significant, changes due to slow learning haveled to the disappearance of locally unstable neuro-spatialdeviations. Gradually, almost axially symmetrical arrange-ments are formed that reflect the overall symmetry of theenvironment occupied by the moving UEs. Finally, whenthe learning rate approached zero, neural movements becamelimited (virtually static), and thus was considered adapted,and a fixed arrangement was obtained that integrated all theenvironmental movements. As a by-product, axial symmetryof the neural arrangements was developed, which was exter-nally dictated by the axis of the transport funnel and themacrocell axis (Fig. 3d). The principal positive simulationfinding is that final neuronal alignment is reflected by theenhanced global throughput value despite the fact that theglobal fitness function is not explicitly defined for the SOMnetwork. By observing small, residual, deviations from theregularity of the agreements, we conclude that barriers tooptimization appear to be an obstacle to the relaxation andfurther improvement of the learning (optimization) process.

Figure 4 explains how the SOM algorithm learns overtime to improve the performance of deploying small cells.

VOLUME 6, 2018 39815

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

FIGURE 3. Visual illustration of the SOM learning framework. Four plots indicate the small cell topology design at different stages of theSOM learning process. The subplot notation a), b), c), and d) correspond to the different SOM learning stages depicted in Fig. 4.

Starting from a random initial BPP placing (point a), thealgorithm attempts to determine improved small cell place-ment. It is noticed that initially the algorithmic process is‘‘confused’’ regarding how to place the small cells effectively.Such behavior is normal considering the unsupervised natureof an SOM. Figuratively, its cognitive abilities are at the levelof a baby taking its first steps. Thus, the average throughputper user is significantly less compared with BPP during thislearning period. After several iterations, the SOM algorithmdetermines where small cells should be placed to improvethe experience of the UEs. Then, in a short time, the averagethroughput per user increases more than twice (from point bto point c), reaching values unattainable by BPP or otherexisting solutions. Point c corresponds to the local optimalHetNet coverage. From point c to point d, we do not observeany significant changes in the network performance becausethe SOM algorithm performs only minor shifting of the small

cells to determine the best configuration. The difference inthroughput between points c and d can be neglected due tothe additional noise in the system caused by the UE mobility.

To compare the performance of the proposed SOMalgorithm with existing stochastic geometry algorithms,we conducted simulations for three key parameters of HetNet:average throughput per user, fairness of resource allocation,and coverage probability. Figure 5 displays the results forthe average throughput per user. It is clear that both SOMimplementations outperformed the conventional BPP deploy-ment, especially when the number of small cells increased.However, the difference between the two SOM metrics(SINR and Euclidean distance) was minimal and diminishedfor a large number of small cells. This phenomenon canbe explained by the features of the simulation scenario.Increasing the number of small cells per area naturallyresults in a smaller cell size. If the cell size is sufficiently

39816 VOLUME 6, 2018

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

FIGURE 4. Average throughput per user vs. SOM training time. Initially,the throughput dramatically decreases owing to the large radius and fastlearning rate. However, the throughput gradually improves over thelearning time duration. Notations a), b), c), and d) indicate differentstages of the SOM learning process.

FIGURE 5. Average throughput per UE vs. number of small cells. Thestudied algorithms provide enhanced results in terms of the throughputcompared to BPP. However, the differences between the proposedmetrics diminish for NS = 18.

FIGURE 6. Fairness vs. number of small cells for the studied algorithms.Clearly, the fairness decreases with an increased number of small cells.The higher the number of small cells, the greater the degree of freedomwith user-cell association present, resulting in a decrease of fairness.The proposed algorithms provide superior results in terms of fairnesscompared with BPP.

small, we can assume that the SINR variation over the cellis negligible; SINR depends only on the received interferencepower. Received interference power increases proportion-ally to the Euclidean distance from the serving BS. Thus,Euclidean and SINR metrics are closely related and indicatesimilar performance for all simulated parameters.

FIGURE 7. Coverage probability for NS = 6. SOM with proposed SINRmetric performs the best. SOM with Euclidean metric performs betterthan BPP, yet significantly worse than using the SINR metric.

FIGURE 8. Coverage probability for NS = 12. SOM with both proposedmetrics significantly outperforms BPP topology design. However,the differences between the metrics do diminish to a virtually negligiblelevel.

Although Fig. 5 displays the advantage of SOMs, thisdoes not necessarily mean that the most effective networkconfiguration is achieved. It is common in HetNet to observea situation where some UEs experience significantly higherthroughput than average, whereas the throughput for otherUEs is significantly below average. Thus, it is importantto determine what configuration of small cells providesa more uniform throughput distribution among the UEs.Figure 6 depicts fairness versus the number of small cells forthe studied algorithms. As can be observed from the results,the fairness decreases with an increasing number of smallcells. This result was expected because UEs tend to stayless time in cells with a smaller radius, resulting in a higherprobability of small cell overload or underutilization. Never-theless, the proposed SOM algorithm demonstrates superiorfairness compared to the BPP, regardless of the metric used.This advantage can be explained by the self-optimizationcapability of SOM. Unlike the BPP, which tends to fill thearea without coverage holes, the SOM algorithm determinesthe positions of the small cells according to the probabilityof traffic demand in each area. Thus, SOM avoids small cellplacement in areas where traffic demand does not exceedthe capabilities of the macrocell. Therefore, considerablyless inequality is observed in HetNet if the small cell layoutfollows the SOM topology.

VOLUME 6, 2018 39817

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

Summarizing the results in Figs. 5 and 6, we can concludethat the effective number of small cells, which satisfy bothcriteria, is from 6 to 12. If the number of small cells is lessthan 6, the average throughput is overly low, even thoughthe fairness distribution is sufficiently fair. Conversely,increasing the number of small cells to more than 12 canimprove average throughput per user. However, the fairnessin this case degrades significantly and the majority of UEswill experience unacceptable throughput.

For the two chosen cases, we compare the coverage prob-ability versus target value of SINR in Figs. 7 and 8. Thedifference between the two simulated scenarios indicates thateach additional small cell decreases the SINR; this is becausenumber of interfering transmitters is increased. These resultscorrelate well with the results presented in Figs. 5 and 6 andhelp to highlight the advantages of the proposed SOM overthe conventional BPP deployment of small cells. Coverageprobability is an important parameter; it provides insightregarding the percentage of UEs who experience acceptableSINR values. Given the direct relation between SINR andthroughput, it is clear that SOM deployment provides animprovement of overall HetNet performance compared withconventional BPP deployment regardless of the simulatedscenario. Considering the fact that quasi-optimal networkconfiguration depends solely on the statistics of trafficdemand in a particular coverage area, the proposed SOMalgorithm is a promising tool for the task of HetNet topologydesign.

VI. CONCLUSIONA new approach to HetNet topology design based on unsu-pervised learning was presented in this paper. The proposedapproach aims to address the challenge of optimal smallcell placement within a target coverage area consideringthe statistics of traffic load HSs and co-channel interfer-ence among the UEs of different cells. The HetNet topologyis designed based on the unsupervised SOM algorithm.The proposed SOM-based algorithm can adjust the posi-tion of small cells using one of two criteria: average UEdistance from the serving BS or average SINR perceived byeach UE. The important advantage of this approach is thatit can be trained for any type of UE mobility pattern andprovides unlimited opportunities for HetNet design for anytype of urban environment. Simulations were conducted tocompare the performance of the proposed SOM algorithmwith the stochastic BPP algorithm, which has been widelyconsidered as the most feasible solution for two-tier HetNettopology design. We compared both algorithms based on theaverage throughput per UE, fairness of throughput distri-bution among UEs, and coverage probability. The resultsconfirm that the proposed approach outperforms the BPPfor all simulated scenarios. In particular, the SOM gainin average throughput was noticeable with an increase inthe number of small cells. However, for a large number ofsmall cells, the probability of unfair throughput distribution

increased, due to the reduced coverage area of the smallcells. The results of coverage probability indicate that inthe case of 6 small cells, SOM provides SINR gains of upto 2 dB for the Euclidean distance metric and up to 4 dBfor the SINR metric. In the case of 12 small cells, SOMprovides an SINR gain of up to 6 dB for both studied metrics,meaning that for small cells, the difference between SINRand the Euclidean metric is negligible. In further research,we will provide additional insight into the HetNet coverageplanning for different scenarios using unsupervised learningalgorithms with different performance criteria.

REFERENCES[1] P. Bhat et al., ‘‘LTE-advanced: An operator perspective,’’ IEEE Commun.

Mag., vol. 50, no. 2, pp. 104–114, Feb. 2012.[2] B. Soret, H. Wang, K. I. Pedersen, and C. Rosa, ‘‘Multicell coopera-

tion for LTE-advanced heterogeneous network scenarios,’’ IEEE WirelessCommun., vol. 20, no. 1, pp. 27–34, Feb. 2013.

[3] Y.-W. Chen andW.-H. Lu, ‘‘Small cells placement scheme in metropolitanarea environment,’’ in Proc. 18th Asia–Pacific Netw. Oper. Manage.Symp. (APNOMS), Oct. 2016, pp. 1–6.

[4] J. Hoadley and P. Maveddat, ‘‘Enabling small cell deployment withHetNet,’’ IEEE Wireless Commun., vol. 19, no. 2, pp. 4–5, Apr. 2012.

[5] W. Guo and S. Wang, ‘‘Interference-aware self-deploying femto-cell,’’IEEE Wireless Commun. Lett., vol. 1, no. 6, pp. 609–612, Dec. 2012.

[6] ‘‘Small cell siting challenges,’’ Small Cell Forum, Dursley, U.K.,Tech. Rep. 192.09.01, 2017.

[7] C. Kim, R. Ford, and S. Rangan, ‘‘Joint interference and user associationoptimization in cellular wireless networks,’’ in Proc. 48th Asilomar Conf.Signals, Syst. Comput., Nov. 2014, pp. 511–515.

[8] T. Kohonen, ‘‘The self-organizing map,’’ Neurocomputing, vol. 21,nos. 1–3, pp. 1–6, 1998.

[9] S. Kaski, T. Honkela, K. Lagus, and T. Kohonen, ‘‘WEBSOM—Self-organizing maps of document collections,’’ Neurocomputing, vol. 21,nos. 1–3, pp. 101–117, 1998.

[10] T. Kohonen, ‘‘Essentials of the self-organizing map,’’ Neural Netw.,vol. 37, pp. 52–65, Jan. 2013.

[11] T. Kohonen et al., ‘‘Self organization of a massive document collection,’’IEEE Trans. Neural Netw., vol. 11, no. 3, pp. 574–585, May 2000.

[12] C. Amerijckx, M. Verleysen, P. Thissen, and J. Legat, ‘‘Image compressionby self-organized Kohonen map,’’ IEEE Trans. Neural Netw., vol. 9, no. 3,pp. 503–507, May 1998.

[13] O. Depren, M. Topallar, E. Anarim, and M. K. Ciliz, ‘‘An intelligentintrusion detection system (IDS) for anomaly and misuse detection incomputer networks,’’ Expert Syst. Appl., vol. 29, no. 4, pp. 713–722, 2005.

[14] C. Koutsougeras, Y. Liu, and R. Zheng, ‘‘Event-driven sensor deploy-ment using self-organizing maps,’’ Int. J. Sensor Netw., vol. 3, no. 3,pp. 142–151, 2008.

[15] T. M. Salem, S. Abdel-Mageid, S. M. Abdel-Kader, and M. Zaki,‘‘ICSSSS: An intelligent channel selection scheme for cognitive radio adhoc networks using a self organized map followed by simple segregation,’’Pervasive Mobile Comput., vol. 39, pp. 195–213, Aug. 2017.

[16] Z. Zhang, K. Long, and J. Wang, ‘‘Self-organization paradigms and opti-mization approaches for cognitive radio technologies: A survey,’’ IEEEWireless Commun., vol. 20, no. 2, pp. 36–42, Apr. 2013.

[17] X. X. W. Li and V. C. M. Leung, ‘‘Artificial intelligence-based techniquesfor emerging heterogeneous network: State of the arts, opportunities, andchallenges,’’ IEEE Access, vol. 3, pp. 1379–1391, 2015.

[18] I. Rhee, M. Shin, S. Hong, K. Lee, S. J. Kim, and S. Chong, ‘‘On the Levy-walk nature of human mobility,’’ IEEE/ACM Trans. Netw., vol. 19, no. 3,pp. 630–643, Jun. 2011.

[19] C. Song, T. Koren, P. Wang, and A.-L. Barabási, ‘‘Modelling the scalingproperties of human mobility,’’ Nature Phys., vol. 6, no. 10, pp. 818–823,2010.

[20] P. Tian, H. Tian, L. Gao, J. Wang, X. She, and L. Chen, ‘‘Deploy-ment analysis and optimization of macro-pico heterogeneous networks inLTE-A system,’’ in Proc. 15th Int. Symp. Wireless Pers. MultimediaCommun., Sep. 2012, pp. 246–250.

39818 VOLUME 6, 2018

J. Gazda et al.: Unsupervised Learning Algorithm for Intelligent Coverage Planning and Performance Optimization

[21] S. Y. Shin and I. F. M. Zain, ‘‘Cellular network planning for heteroge-neous network using geo-clustering algorithm,’’ in Proc. Int. Conf. ICTConverg. (ICTC), Oct. 2013, pp. 448–449.

[22] X. Li, X. Tang, C.-C.Wang, and X. Lin, ‘‘Gibbs-sampling-based optimiza-tion for the deployment of small cells in 3G heterogeneous networks,’’in Proc. 11th Int. Symp. Workshops Modeling Optim. Mobile, Ad HocWireless Netw. (WiOpt), May 2013, pp. 444–451.

[23] M. H. Qutqut, H. Abou-zeid, H. S. Hassanein, A. M. Rashwan, andF. M. Al-Turjman, ‘‘Dynamic small cell placement strategies for LTEheterogeneous networks,’’ inProc. IEEE Symp. Comput. Commun. (ISCC),Jun. 2014, pp. 1–6.

[24] Y. Park et al., ‘‘Effective small cell deployment with interference andtraffic consideration,’’ in Proc. IEEE 80th Veh. Technol. Conf. (VTC-Fall),Sep. 2014, pp. 1–5.

[25] A. Jaziri, R. Nasri, and T. Chahed, ‘‘System-level analysis of heteroge-neous networks under imperfect traffic hotspot localization,’’ IEEE Trans.Veh. Technol., vol. 65, no. 12, pp. 9862–9872, Dec. 2016.

[26] I. Bahçeci, ‘‘On femto-cell deployment strategies for randomly distributedhotspots in cellular networks,’’ in Proc. IEEE Wireless Commun. Netw.Conf. (WCNC), Apr. 2014, pp. 2330–2335.

[27] L. Zhou et al., ‘‘Green cell planning and deployment for small cellnetworks in smart cities,’’ Ad Hoc Netw., vol. 43, pp. 30–42, Jun. 2016.

[28] T. Nakamura et al., ‘‘Trends in small cell enhancements in LTE advanced,’’IEEE Commun. Mag., vol. 51, no. 2, pp. 98–105, Feb. 2013.

[29] H. Holma and A. Toskala, LTE for UMTS-OFDMA and SC-FDMA BasedRadio Access. Hoboken, NJ, USA: Wiley, 2009.

[30] V. S. Abhayawardhana, I. J. Wassell, D. Crosby, M. P. Sellars, andM. G. Brown, ‘‘Comparison of empirical propagation path loss modelsfor fixed wireless access systems,’’ in Proc. IEEE 61st Veh. Technol.Conf. (VTC-Spring), vol. 1. May/Jun. 2005, pp. 73–77.

[31] S. F. Yunas, W. H. Ansari, and M. Valkama, ‘‘Technoeconomicalanalysis of macrocell and femtocell based HetNet under differentdeployment constraints,’’ Mobile Inf. Syst., vol. 2016, May 2016,Art. no. 6927678.

[32] D.-M. Chiu and R. Jain, ‘‘Analysis of the increase and decrease algorithmsfor congestion avoidance in computer networks,’’ Comput. Netw. ISDNSyst., vol. 17, no. 1, pp. 1–14, 1989.

[33] C. Guang G. Jian, and D. Wei, ‘‘A time-series decomposed model ofnetwork traffic,’’ in Advances in Natural Computation (ICNC) (LectureNotes in Computer Science), vol. 3611, L. Wang, K. Chen, and Y. S. Ong,Eds. Berlin, Germany: Springer, 2005, pp. 338–345.

[34] W.Wang, J.Wang,M.Wang, B.Wang, andW. Zhang, ‘‘A realistic mobilitymodel with irregular obstacle constraints for mobile ad hoc networks,’’Wireless Netw., to be published.

[35] Small Cell Enhancements for E-UTRA and E-UTRAN—Physical LayerAspects, document TR 36.872, 3GPP, 2013.

JURAJ GAZDA was a Guest Researcher withRamon Llull University, Barcelona, Spain, andalso with the Technical University of Hamburg-Harburg, Germany, under the supervision ofProf. H. Rohling. He was also involved inthe development activities of Nokia SiemensNetworks. In 2017, he was recognized as the BestYoung Scientist of the Technical University ofKošice. He is currently an Associate Professorwith the Faculty of Electrical Engineering, Tech-

nical University of Košice, Slovakia. His research interests include spectrumpricing, techno-economic aspects of cognitive networks, co-existence ofheterogeneous networks, and application of machine learning principlesin 5G networks.

EUGEN ŠLAPAK received the B.A. and M.S.degrees in informatics from the Technical Univer-sity of Košice, Košice, Slovakia, in 2016 and 2018,respectively. His master’s thesis focused on theuse of self-organizing maps for HetNet topologydesign, under the supervision of Prof. J. Gazda.He works part-time with the research team at theIntelligent Information Systems Laboratory, Tech-nical University of Košice.

GABRIEL BUGÁR is currently a Research Assis-tant with the Faculty of Electrical Engineeringand Informatics, Technical University of Košice,Slovakia. His research interests include cognitivenetworks, image processing, and e-learning.

DENIS HORVÁTH was a Junior Scientist withthe Department of Theoretical Physics, SlovakAcademy of Sciences, and then with the Depart-ment of Theoretical Physics and Astrophysics,P. J. Šafárik University, Košice, Slovakia. He iscurrently a member of the Research Group, Centerof Interdisciplinary Biosciences, P. J. ŠafárikUniversity. He is also an experienced researcher,having worked in the interdisciplinary team foralmost 25 years. His present mission is to develop

and implement theoretical models of cognitive networks, artificial markets,and cancer progression. He has extensive experience in data analysis andnumerical modeling, theory and simulation of complex and biologicalsystems, and solutions of optimization problems.

TARAS MAKSYMYUK received the B.A. degreein telecommunications, the M.S. degree ininformation communication networks, and thePh.D. degree in telecommunication systemsand networks from Lviv Polytechnic NationalUniversity, Lviv, Ukraine, in 2010, 2011, and2015, respectively. He received the Post-DoctoralFellowship from the Internet of Things andCognitive Networks Lab, Korea University, undersupervision of Prof. M. Jo. He is currently

an Assistant Professor with the Telecommunications Department, LvivPolytechnic National University. His research interests include cognitiveradio networks, LTE in unlicensed spectrum, mobile cloud computing,massive MIMO, 5G heterogeneous networks, software-defined radio accessnetworks, Internet of Things, big data, and artificial intelligence. He wasrecognized as the Best Young Scientist of Lviv Polytechnic National Univer-sity in 2015, received the Lviv State Administration Prize for outstandingscientific achievements and contribution in 2016, and Lviv MetropolitanPrize for Best Scientists in 2017. He is currently an Associate Editor of theIEEE Communications Magazine, and an Editor of the International Journalof Internet of Things and Big Data.

MINHO JO received the B.A. degree from theDepartment of Industrial Engineering, ChosunUniversity, South Korea, in 1984, and the Ph.D.degree from the Department of Industrial andSystems Engineering, Lehigh University, USA,in 1994. He is currently a Professor with theDepartment of Computer Convergence Software,Korea University, Sejong, South Korea. He is oneof the founders of the Samsung Electronics LCDDivision. He has published over 100 publications

in reputed journals/magazines and international conferences. His currentresearch interests include IoT, LITE-Unlicensed, artificial intelligence andbig data in the IoT, blockchain and security, HetNets in 5G, cloud computing,wireless energy harvesting, 5/6G wireless communications, autonomouscars, and optimization and probability in networks. He received the HeadongOutstanding Scholar Prize in 2011. He was the Vice President of the Instituteof Electronics of the Korea Information Processing Society. He is currentlythe Vice President of the Korean Society for Internet Information. He iscurrently an Editor of the IEEE WIRELESS COMMUNICATIONS, an AssociateEditor of the IEEE ACCESS, the IEEE INTERNET OF THINGS JOURNAL, and Secu-rity and Communication Networks (Wiley), and Wireless Communicationsand Mobile Computing (Hindawi). He is the Founder and Editor-in-Chiefof KSII Transactions on Internet and Information Systems (SCI and Scopusindexed).

VOLUME 6, 2018 39819