5G D2D Transmission Mode Selection Performance & Cluster Limits Evaluation of Distributed Artificial Intelligence and Machine Learning Techniques
Iacovos Ioannou, Christophoros Christophorou, Vasos Vassiliou, Andreas Pitsillides
55G D2D Transmission Mode Selection Performance& Cluster Limits Evaluation of Distributed ArtificialIntelligence and Machine Learning Techniques
Iacovos Ioannou ∗† , Christophoros Christophorou ∗‡ , Vasos Vassiliou ∗† , and Andreas Pitsillides ∗∗ Department of Computer Science, University of Cyprus † CYENS Center of Excellence, Nicosia, Cyprus ‡ CITARD Services LTD, Nicosia Cyprus
Abstract —5G D2D Communication promises improvements inenergy and spectral efficiency, overall system capacity, and higherdata rates. However, to achieve optimum results it is importantto select wisely the Transmission mode of the D2D Device inorder to form clusters in the most fruitful positions in terms ofSum Rate and Power Consumption. Towards this end, this paperinvestigates the use of Distributed Artificial Intelligence (DAI)and innovative to D2D, Machine Learning (ML) approaches (i.e.,DAIS, FuzzyART, DBSCAN and MEC) to achieve satisfactoryresults in terms of Spectral Efficiency (SE), Power Consumption(PC) and execution time, with the creation of clusters andbackhauling D2D network under existing Base Station/Small Cell.Additionally, one of the major factors that affect the creation ofhigh quality clusters (e.g. higher Sum Rate) under a D2D networkis the number of the Devices.Therefore, this paper focuses ona small number of Devices (i.e., < =200), with the purpose toidentify the limits of each approach in terms of number of devices.Specifically, to identify where it is beneficial to form a cluster,investigate the critical point that gains increases rapidly and atthe end examine the applicability of 5G requirements. Addition-ally, prior work presented a Distributed Artificial Intelligence(DAI) Solution/Framework in D2D and a DAIS TransmissionMode Selection (TMS) plan was proposed. In this paper DAISis further examined, improved in terms of thresholds evaluation(i.e., Weighted Data Rate (WDR), Battery Power Level (BPL)),evaluated, and compared with other approaches (AI/ML). Theresults obtained demonstrate the exceptional performance ofDAIS, compared to all other related approaches in terms ofSE, PC, execution time and cluster formation efficiency. Also,results show that the investigated AI/ML approaches are alsobeneficial for Transmission Mode Selection (TMS) in 5G D2Dcommunication, even with a smaller number ( > =5) of devices asa lower limit. Keywords -5G, D2D, Transmission Mode selection, DistributedArtificial Intelligence, Unsupervised Learning, Clustering
I. I
NTRODUCTION
Device-to-Device (D2D) Communication is expected to bea contributing factor in achieving the demanding requirementsof 5G Mobile Communication Networks [1], [2]. The mainreasons are that D2D communication is not constrained bythe licensed frequency bands and that it is transparent to the
This research is part of a project that has received funding from theEuropean Union’s Horizon 2020 research and innovation programme undergrant agreement Nº739578 and the government of the Republic of Cyprusthrough the Directorate General for European Programmes, Coordination andDevelopment. cellular network. Also, it permits adjacent User Equipment(UE) to bypass the Base Station (BS) and establish direct linksbetween them, either by sharing their connection bandwidthand operate as relay stations, or by directly communicatingand exchanging information. For the aforesaid reasons, D2Dcan improve spectral efficiency, data rates, throughput, energyefficiency, delay, interference and fairness [2]–[5].However, in order to achieve optimum results, it is impor-tant, among others, to select wisely the Transmission Mode ofthe D2D Device in order to form clusters in the most fruitfulpositions in terms of Sum Rate and Power Consumption. Themain reason is that the Transmission Mode selection for adevice can affect the creation of the clusters, the way data willbe communicated between the D2D Devices, and it can alsooptimize backhauling links between disconnected/disjointedclusters by forming better paths.Additionally, for higher Sum Rate (Total Spectral Effi-ciency) and reduced total Power Consumption there are factorsthat affect the quality of Cluster forming in D2D. The majorcontributing factors in the successful realization of a D2Dcluster under a network are the following: i) number ofdevices; ii) backhauling Data Rate achieved by a link; iii)position of Cluster Head (CH); iv) Data Rate of CH; and v)QoS & QoE. In this paper the factor of number of devicesis examined in terms of limits evaluation, in the directionof the small number of devices network. Also all factors areexamined for the investigation approaches.Towards this end, our previous work [2] proposed: i) a BDIx(BDI extended) agents based Distributed Artificial Intelligence(DAI) Framework that can achieve D2D communication in anefficient and flexible manner by focusing on the local environ-ment rather the global environment. A BDIx agent is an agentthat has Believes (i.e., knowledge about the environment),Desires (i.e., it has some objectives to achieve) and Intentions(i.e., objectives that are currently executed through selectedplans). Note that the Desires of a BDIx agent, and thus itsintentions, can change with the raising of an event (i.e., a newD2D Device entering the Mobile Network). More specifically,an event may update believes, trigger plans or modify goals(believes) and intentions. With the examined approach the BDIagents concept is extended, by utilizing AI techniques (e.g.,Fuzzy Logic, Deep Learning Neural Networks etc) to form the a r X i v : . [ c s . N I] F e b gent Believes; ii) an autonomous and intelligent TransmissionMode selection approach, called ”DAIS”, to be executed asa plan of DAI Framework towards the Intention (realizedfrom Desire) of selecting the transmission mode of the D2DDevice (in the event of ”entering the Mobile Network”), in adistributed, flexible and efficient manner.In this paper, the efficiency of DAIS is further exam-ined, evaluated, and compared with other related approaches,like Distributed Random, Sum Rate Approach, Centralizednon-D2D-UE (shown in [2]) and other, currently introducedto D2D and Transmission Mode Selection Artificial Intelli-gence/Machine Learning (AI/ML) techniques (i.e., FuzzyART[6], [7], DBSCAN [8], [9] and MEC [10], [11]) in a 5GD2D communication network with a reduced number ofdevices ( < = 200 UEs/D2D candidates). Note that FuzzyART,DBSCAN and MEC are centralized unsupervised learningclustering techniques that, for the purposes of this research,we utilized for D2D communication. These approaches donot require a learning process in order to be used in theD2D communications and they provide good clustering results.The underlying reasons for selecting unsupervised learningclustering techniques are the following: i) the TransmissionMode Selection is directly associated with the selection of bestCluster Head, therefore the clustering techniques must be used;and ii) due to the dynamic nature of mobile communicationnetwork the training part of supervised learning can notconclude to the best results because of the devices movementand due to the fact that in D2D communication the best dataare the current data.For assessing the efficiency of the DAIS approach, thresholdvalues affecting spectral efficiency and power usage of thenetwork, like the Weighted Data Rate (WDR) and the BatteryPower Level (see Section III) of the D2D Device, have beenemployed. In addition, those achieving strong performancehave been determined. The effect of the Transmission Power(TP) variation of each Device on the investigated approaches,in terms of total Spectral Efficiency (SE), Power Consumption(PC) and Execution Time (ET) was also examined. This inves-tigation focuses on D2D communication network with a smallnumber of devices for the following reasons: i) applicabilityof 5G requirements; ii) investigate the critical point that gainsincreases rapidly; iii) coverage expansion; and iv) find thelimits of the approaches.The results obtained demonstrate that with the right tuningof the thresholds, DAIS could provide significant improvementin the network. Furthermore, from the results obtained fromthe comparison of the investigated approaches it was observedthat DAIS outperforms all other approaches, except Sum RateApproach, in terms of total SE and total PC. The reasonthat Sum Rate Approach achieved better results than DAISis because Sum Rate Approach has a global knowledge ofthe network and thus can select the best transmission mode.Even so, DAIS approaches the performance of the Sum RateApproach, acting on only local information. In addition, it wasobserved that Transmission Power (TP) alteration of the D2DDevices with a small number of UEs ( < =200) can affect SE and PC for all investigated approaches.The rest of the paper is structured as follows. Section IIprovides some background information and related work as-sociated with transmission mode selection approaches. SectionIII presents the problem that this paper tackles and providessome specifics about the investigated approaches. Specifically,the implementations, assumptions, constraints, thresholds andmetrics utilized are provided. The efficiency of the investigatedapproaches, is examined, evaluated and compared in SectionIV. Finally, Section V contains our Conclusions and FutureWork.II. B ACKGROUND K NOWLEDGE AND R ELATED W ORK
A. Background Knowledge
This section provides background knowledge regarding themain characteristics of D2D communications. More specif-ically, the types of control that can be exploited for theestablishment of D2D communication links, as well as thetypes of transmission modes that a D2D Device can operate,are outlined in this section.
1) Types of Control in D2D Communication:
The types ofcontrol that can be used for the establishment of D2D Com-munication links can be categorized as follows: i) Centralized:The Base Station (BS) completely oversees the UE nodeseven when they are communicating directly; ii) Distributed:The procedure of D2D node management does not obligeto a central entity, but it is performed autonomously by theUEs themselves; iii) Distributed Artificial Intelligence (DAI):All control processes run in parallel and begin at the sametime through collaboration in an intelligent manner; and iv)Semi distributed/hybrid: A mix of centralized and distributedschemes.
2) Types of Transmission Modes in D2D Communication:
The different transmission modes in D2D Communication arethe following: i) D2D Direct: Two UEs connect to each otherby utilizing licensed or unlicensed spectrum; ii) D2D Single-hop Relaying: Contribution of bandwidth between a UE andother UEs [12]. One of the D2D UEs is connected to a BS orAccess Point and provides access to an additional D2D UE;iii) D2D Multihop Relay: The single-hop mode is extendedby empowering the connection of more D2D UEs in chain.This chain can be one to one relationship or one to more[13]; iv) D2D Cluster [14]: D2D Cluster is a group of UEs(D2D Devices acting as clients) connected to a D2D relaynode performing as a Cluster Head (CH) [15]; and v) D2DClient: D2D Client is the selection of UE to participate in aD2D Cluster and act as client.
B. Related Work
This section provides a brief description of the DAI Solu-tion/Framework along with its Desire Plan DAIS together withSum Rate and Distributed Random algorithms that performTransmission Mode Selection as shown in [2]. Additionally,this section provides information regarding FuzzyART, DB-SCAN, and MEC unsupervised learning Machine Learning(ML) clustering techniques, and other related approachesrom open literature on Transmission Mode selection in D2DCommunication. It is important to highlight here that theaforesaid AI/ML techniques were not designed for applicationin D2D communication but they are utilized and applied toD2D communication by us, for the purposes of this research,due to their scalability, metric used, parameters and way ofcalculation of labels of clusters.
1) Distributed Artificial Intelligent Solution/Framework:
Inthis section, the paper explains in brief the DAI Frameworkthat as concept it was introduced in the [2]. The main objectiveof the DAI framework is to implement 5G D2D communi-cation with the purpose to achieve the D2D challenges (asshown in [2]). By enabling D2D UEs through BDIx agents thatinstantiate through BDIx framework, the investigation aimsfor the devices to act independently, autonomously and as aself-organizing network. More precisely, in order to achievethe aforementioned characteristics, the framework it utilizessoftware agents and especially BDI (Belief-Desire-Intention)agents with extended Artificial Intelligence/Machine Learningcapabilities (ex. Neural networks, Fuzzy logic) named as BDIxAgent. The framework acts as a glue in the employment ofmore than one of successful, optimized intelligent technologies(e.g. Neural Networks, Fuzzy Logic). Therefore, the BDIxframework will be modular and the believes and desires canbe substituted, added by any proposed approach that will haveas target to achieve the D2D communication in 5G, as long thestability of the agent is achieved. Additionally, such agents inthe framework can be implemented at the UEs as a softwareand there is no need to change how BSs operate or to changethe hardware at BSs or UEs.In this paragraph we will show the inner workings of DAIFramework, how it achieves D2D communication in 5G. Morespecifically, the DAI Framework utilizes the networks events(i.e. Device entering in a D2D network) and thresholds (DataRate is acceptable by the user) that are derived from theDesires and an agent must monitor in order to achieve the tasksof implementing 5G D2D communication. More precisely,the events and thresholds can trigger the Desires to becomeIntentions with the use of priority values (from 0% to 100%).For the aforementioned task the Fuzzy Logic (FL) is used asthe Plan library in order to assign priority values to Desires(the framework can let 10 concurrent intentions to run atthe same time). Additionally, the DAI Framework flowchartof execution of the BDIx agents, supports queue of runningIntentions that are realised from Desires with priority value of100% (as shown in the Figure 1). Also, at the Intentions thereare assigned Plans that act as algorithms for the purpose toachieve the selected Desires.The Believes represent the understanding of the agent orthe environment around. The events are actively affecting theBelieves and then Desires are converted to Intentions andsatisfied based on the affected Believes. The events can be pre-specified with the declaration of thresholds. These thresholds,if exceeded, can raise events at the event system Plan Library(FL). The set of Believes in terms of D2D communicationthat a BDIx agent can use, that are derived from the D2D Fig. 1: Flowchart of BDIx Agent Operation [2]requirements are: • Frequency Band connected to BS • Battery Power Level • Used Metric Value (e.g. Weighted Data Rate (WDR) asshown in paper [], ICQ, interference) • Transmission Mode Selected • Frequency Band used • Best reused Frequency Band to be used with less inter-ference. • Percentage of Bandwidth utilization • Data Rate • Lat/Long (Coordinates) • Next Hop that D2D Device connects to (D2D Relay/ D2DMulti Hop Relay as D2D Relay/BS) • Distance from the Next Hop that the D2D Device (UE)connect to • Coordinates of the Next Hop that the D2D Device con-nects to • Percentage signal quality to where I connect to (D2DRelay/ D2D Multi Hop Relay as D2D Relay/BS) • Percentage WDR change to where I connect to (D2DRelay/ D2D Multi Hop Relay as D2D Relay/BS) • Speed (D2D Device moving speed) • Number of users that the D2D Device serves (if trans-mission mode is D2D Relay) • IPs/MSISDN of Users that the D2D Device serves (iftransmission mode is D2D Relay) • Sharing subnet (if transmission mode is D2D Relay) • IP v4 • IP v6 • List of surrounding D2D Relays with coordinates, Fre-quency Band, number of D2D Clients Serve, Frequenciesshared to D2D clients (inband or outband) and metricused (e.g. WDR) • List of surrounding D2D Multi Hop Relays with co-ordinates, Frequency Band, Frequencies shared to D2Dclients (inband or outband) and metric used (e.g. WDR)
Round time of packet to access gateway. • Number of Users that the D2D Device serves as D2DRRelay • Security Breach • Counters of Packets For each D2D Client (for securityreason) • Fuzzy Logic (IF-THEN rules) assigning priority valueson the Desires based on events and BelievesIn DAI Framework the desires are directly related to D2Dchallenges and Network events.In the DAI Framework thereare Desires that have direct relation between each other.Desires are the purpose of existence of the BDIx agent.Intentions are selected, as a subset of Desires. The Intentionsare chosen Desires to be executed at the current moment. TheDesires in our framework are chosen to be executed with theuse of priority. The purpose of existence of priority is becausesome Desires must run before other Desires. Therefore, someDesires must have higher priority, become Intentions, andconclude before other Desires become Intentions and start tobe executed. The set of Desires is the following: • Preferred network is D2D network, always with 100%priority. • Hardware Health is acceptable. • Identify D2D Relays and D2D Multi Hop Relays around. • Find the best reused Frequency with the least Interfer-ence. • Signal quality is acceptable. • Data Rate is acceptable. • WDR is acceptable. • Achieve Maximum Sum Rate • Distance of D2D Client Device with D2D Relay/D2Dmulti Hop Relay is acceptable. • Number of Users that the D2D Device serves as D2DRRelay is acceptable. • Bandwidth consumed by Users that the D2D Deviceserves as D2DR Relay is acceptable • Achieve QoS specified by 5G requirements, always with100% priority. • Achieve QoE specified by 5G requirements, always with100% priority. • The latency (round time/ultra-reliable Low Latency com-munication) of accessing gateway or any other D2DDevice is acceptable, always with 100% priority. • Battery Power Level reservation at D2D Device, alwayswith 100% priority. • Security Monitoring at D2D Device, always with 100%priority.Thus, the DAI Framework can achieve D2D communicationby focusing on the local environment rather than the globalenvironment with the use of LTE Proximity Services (LTEProSe). The plan that this research investigates is DAIS asshown in the [2] and is executed in the network event of ”D2DDevice entering in D2D communication network”.
2) DAIS, Sum Rate Approach and Distributed Random [2]:
DAIS is a distributed, autonomous and intelligent Transmis- sion Mode Selection approach, implemented in a BDIx agentbased DAI Framework, that selects the transmission mode ofa D2D Device in a distributed artificial intelligence manner.More specifically, the DAIS approach exploits software agentsand especially Believe-Desire-Intention (BDI) agents withextended Artificial Intelligence/Machine Learning (AI/ML)capabilities (BDIx), to select the transmission mode that willbe used by a a new D2D Device. For the Transmission modeselection, the WDR (Weighted Data Rate), a new metric thatwe introduced in [2], is considered. Sum Rate Approach, isa distributed intelligent approach which uses the sum rateof the network as a metric for the UE Device to select thebest Transmission mode. Note that in the Sum Rate Approachthe D2D Device selects the most appropriate TransmissionMode by having all the knowledge of the network (i.e., D2DRelays, D2D Multi Hop Relays, D2D Clients, connectionlinks). On the other hand, the Distributed Random approachis a distributed approach which performs Transmission modeselection in a random manner (e.g. the algorithm for Trans-mission Selection selects randomly a mode of the enteringdevice).
3) FuzzyART [6], [7]:
FuzzyART is an unsupervised learn-ing algorithm that uses structure calculus based on fuzzylogic and Adaptive Resonance Theory (ART), for the purposeof pattern recognition and to enhance generalization. TheFuzzyART consists of a comparison field and a recognitionfield composed of neurons, a vigilance parameter (thresholdof recognition), and a reset module. The comparison field takesan input and transfers it to its best match to a single neuronwhose set of weights most closely matches the input vector inthe recognition field. Each recognition field neuron outputs anegative signal to each of the other recognition field neurons.Additionally, in FuzzyART the computation of choice functionvalue consists of fuzzy ”AND” operator. The aforementionedprocedure allows each neuron in it to represent a category towhich input vectors are classified. After classification, the resetmodule compares the strength of the recognition match to thevigilance parameter, if it has greater strength it adjusts weights,elsewhere the search procedure is carried out. The vigilanceparameter has considerable influence on the system (e.g., morecategories). So, FuzzyART provides a unified architecture forbinary and continuous value inputs. The consequential numberof clusters depends on the distances between the investigatedelements that we want to cluster (this also depends on themetric chosen for the approach, i.e., Gaussian distance) amidall input patterns, introduced in the direction of the network forthe period of training cycles. For FuzzyART the algorithmiccomplexity is of order O( N )+O(MN), N being the numberof categories, and M the input dimension. Because it can havemaximum N x N recursive iterations and form clusters basedon the period of training cycles (MN).
4) DBSCAN [8], [9]:
The DBSCAN algorithm dependson a density-based concept of clusters, which is outlined todetermine clusters of unacquainted shape. In DBSCAN, foreach point of a cluster, the neighborhood of a prearrangedradius has to enclose at least a minimum number of pointsMinPts in DBSCAN). DBSCAN starts with an arbitrary start-ing point that has not been visited. Afterwards, the surroundingpoints, called neighborhood, are retrieved. If the examinedpoint contains a sufficient number of points around it thena cluster is initialized and the identified neighborhood pointsare added in the cluster. Otherwise, the investigated pointis labeled as noise, note that this point might be a part ofanother future examined cluster. This process continues untilthe cluster is completely found or unvisited points are retrievedand processed. The algorithmic complexity is mostly governedby the number of area Query requests. DBSCAN executesone area query for each point, in the case of utilizationof indexing structure executing a neighborhood query, theresulting algorithmic complexity achieved to be O(N), whereN is the maximum number of points that can be involved inthe neighboring query. However, by taking under considerationall the cases an overall algorithmic complexity of O( N ) isachieved.
5) Minimum Entropy Clustering (MEC) [10], [11]:
TheMEC algorithm proficiently minimizes the conditional entropyof clusters. By analyzing given samples consequently, at theend it concludes with the clusters. In MEC, the clusteringcriterion is based on the conditional entropy H ( C | x ) , whereC is the cluster label and x is an observation. MEC withFano’s inequality, C can be estimated with a low probabilityof error only if the conditional entropy H ( C | x ) is small.This algorithm utilizes mathematical facts, such as Havrda-Charvat’s structural. The replacement of Shannon’s entropywith Havrda-Charvat’s structural α -entropy is selected for thepurpose of achievement of the generalization of the clusteringcriterion, α -entropy indicates if the probability error is equal tothe nearest neighbor method when α =2. Additionally, Fano’sinequality and Bayes probability of error is utilized with theParzen density estimation, a non-parametric approach. Themethod performs very well even when the correct number ofclusters is unknown, with the utilization of maximum distanceas input. It can also accurately reveal the structure of dataand efficiently identify outliers simultaneously. However, thisapproach is an iterative algorithm initialized with a parti-tion set by any other clustering approaches (e.g., K-Means)and random initialization should not be used. The resultingalgorithmic complexity achieved is O( N ), where N is thenumber of all points that can be involved in the neighboringquery, in the formula the calculation of the entropy is included.However, by taking under consideration all the cases an overallalgorithmic complexity of O( N ) is achieved. C. Related work on Transmission Mode Selection in D2DCommunication
Approaches related to the Transmission mode selectioninvestigated in this paper, are provided in a plethora of articles[2], [16]–[21]. The metrics considered for selecting the trans-mission mode to be adopted are: power, interference, resourceblocks (RB), SINR, distance, power, frequencies and WDR. Inthe literature one can find approaches with a focus on: i) D2DDevice Selection [16], [17], [22]; ii) Relay selection only [18], [19], [23]; and iii) D2D multi-hop relay forming by selectingas modes the D2D or D2D Multihop [20], [21]. In our workwe are examining all of the possible transmission modes thatcan be assign to a UE, by itself (e.g. BDIx Agent) or by otherentities (e.g. BS).A classification on the related approaches based on thetype of control (see Section II-A1) is: i) Centralized [16]–[18], [20], [21], where the decision is taken by the BS;ii) Semi-distributed approaches [22], where the decision istaken by both the BS and the D2D Devices in collaboration;iii) Distributed [19], where the decision is taken by theD2D Devices; however in this case the D2D Devices needsome information from the BS; and iv) Distributed ArtificialIntelligent (DAI) [2], where the decision is taken by each D2DDevice independently; however, in this case they may shareinformation with other D2D Devices.It is evident from the above preliminary survey that mostworks use the Centralized approach and only a few useSemi or Fully Distributed algorithms. Additionally, we couldnot identify any other approach in the open literature thattackles the problem of having a D2D Device utilizing alltransmission modes (D2D Relay, D2D Multi-Hop Relay andD2D Cluster) in a distributed AI manner. Furthermore, to thebest of our knowledge, there is not any other D2D transmissionmode selection approach in the literature that is utilizingunsupervised learning AI/ML clustering techniques. Therefore,the usage of unsupervised learning AI/ML approaches for theTransmission mode selection in D2D communication, is alsoa contribution of this paper.III. P
ROBLEM F ORMULATION
In this paper we aim to use DAI and ML in order for aD2D Device to select a Transmission Mode and create a D2Dcommunication network with for the purpose to reduce thedistance to the Access Point, reduced the latency, increaseSE and reduced PC in a small ( < =200) number of DevicesD2D Network. The number of UEs examined is small, dueto one of the major contribution of the paper, because weaim to calculate the investigated approaches lower limits inan environment of small number of devices in order to toshow where is fruitfully to achieve cluster with drones, otherrelay devices and an operator should consider not change thetopology of the network.Additionally, please note that similarproblems,even with the same number of devices, are resolvedwith the use of small cells [24], [25]. Therefore, the problemthat this paper tries to tackle is threefold: • It tries to maximize the total SE (i.e, sum rate) andreduce the total PC of the DAIS algorithm as wellas the other investigated unsupervised learning AI/MLclustering techniques, in the case of a small number N ofdevices ( < =200 UEs) under a BS. Therefore, this paperhave the following constrains about the physical link: – The D2D network consists of N devices under the BaseStation (BS) – Our approach focuses on the mobile and wireless net-works with a single-antenna and point-to-point scenario
Our approach uses the Free Space Model and FreeSpace Path Loss – Our approach uses the Additive White Gaussian Noise(AWGN) as the basic noise model – The Transmission Power (TP) is known – The Spectral Efficiency is calculated per linkMore specifically, the following paragraph will show theequations used in order to do the problem formulation, theparameters description is shown in the Table I. Startingfrom Shannon–Hartley theorem, the spectral efficiency isshown in Equation 1, measured in (bits/s/Hz). SE = CB = log2 Ç SN å (1) Therefore, with the use of the aforementioned modelthe spectral efficiency calculated from channel capac-ity is used with the power-limited and bandwidth-limited scheme and is indicated below in Equation 2(SE/ SE Link ), measured in (bits/s/Hz). SE = C AWGN W = log2 (1 + SNR ) SNR = ¯ PN W (2) Also, the average received power (in W) is calculatedas ¯ P , Transmission Power (TP) is known to the channel(TP), Power Consumption is shown in Equation 3, SNRis the received signal-to-noise ratio (SNR) and lastly thenoise is N (W/Hz). PC = TP − ¯ P (3) Therefore, the problem is based on the Equations 4 and5 which is the maximization of Total SE with as resultthe reduction of the Total PC. This is a NP-hard problemto solve (e.g. see [24]–[26]), this is the reason that aheuristic algorithm is implemented for the utilized MLalgorithms.
TotalSE = max N (cid:88) i =1 SELink (4)
Link ∈ { D D Relay, D D Multi Hop, D D Client } TotalPC = min N (cid:88) j =1 PC (5) • It examines the problem of forming Back-hauling links,with the selection of D2D Multi Hop Relay Transmissionmode, form DAIS and Sum Rate in small number ofdevices network. • It examines the problem of identifying the best clusterheads in a D2D communication network with the use ofTransmission Mode Selection and AI/ML techniques forthe Unsupervised Learning Clustering techniques. • It examines if unsupervised learning techniques can beutilized in order to achieve equal or better results as DAISand Sum Rate Approach in terms of Transmission modeselection (as shown at [2]). • It examines the cluster formation in terms of number ofclusters and number of devices not enter any cluster. • It examines the number of messages exchanged for com-pletion of the algorithm. • It examines the time that each approach used for struc-turing the D2D communication network.Overall in our approach we consider as the worst case scenariothe Random approach and the best approach as the Sum Rateapproach that knows all the D2D Devices and the links in
Parameter Parameters DescriptionC capacity (in bits per second b/s)B bandwidth (in Hertz Hz)S signal power (in mini Watts mW)N noise power (in decibel dB) C AWGN capacity with the use of the AWGN noise modelW bandwidth (in bits per second bps)SNR received signal-to-noise ratio (SNR) N noise (in Watts per Herz W/Hz) ¯ P average received power (in mini Waatts mW)calculated using aFree Space Model and a Free Space Path LossTP Transmission Power known to the channel(from the UE and Base Station specifications) TABLE I: Parameters Descriptionthe D2D network with the opportunity to do a brute forcecalculation, with the target to calculate the maximum possibleSE that results to reduced PC.With the implementation of DAIS and the use of BDIxagents, there are some assumptions, constraints, thresholds,and a new metric that are introduced. However, in order toshow how the BDIx Agents framework can be optimizedin terms of threshold investigation, only the ”Weighted DataRate” (WDR) metric has been analyzed and utilized. Basi-cally, the aim of the DAIS approach is to maximize the WDR(i.e., WDR = max(min(Link Rate))) for each path. In this paperan investigation of the DAIS thresholds is executed with thepurpose to increase the Total SE and Total PC.Additionally, a heuristic algorithm (see section IV) hasbeen developed that utilizes the clustering results extractedby FuzzyART, DBSCAN and MEC approaches to select thebest D2D Device in the identified cluster to be set as a D2DRelay node. Note that the metric used to perform the selectionis the Data Rate (as described in Algorithm 1). Likewise, thefeature set used for all the unsupervised learning clusteringapproaches is the same and it is the set composed withlatitude and longitude (coordinate). Additionally, note that theaforementioned approaches does not form backhauling morethan one hop and the selection of D2D Multi Hop Relay isnot provided as selection option of Transmission Mode in theapproaches.It is worth mentioning that in order to apply the Fuzz-yART, DBSCAN and MEC approaches to the needs of D2DCommunication, we utilized these approaches and set theconstraints/settings set out below: • For all approaches, we set the maximum radius distanceto form a cluster to 200 meters (WiFi Direct). • For FuzzyART we do not limit the maximum number ofclusters allowed (maxClusterCount=-1). • For DBSCAN we set the minimum points (minPts) of thecluster to 2. • For MEC we set the number of clusters (k) to 100 (notethat the final number of clusters may be less). The WDR metric is defined at each node in D2D communication as theminimum data rate in the path that the UE has selected, either this is directlyconnected to the BS or through another D2D Device. ote that except from the aforesaid constraints/settings set,all other default settings and constraints provided by the“SMILE” framework are the same [27].
Algorithm 1
Heuristic Algorithm to Calculate Cluster andCluster Head of FuzzyART/DBSCAN & MEC i: radius of Cluster Head T: a set containing clusters procedure C LUSTER H EAD D ETECTION ( T th , i ) T ui ← list of ClustersfromT th for each cluster c in T ui do Nodec i ← maximumDataRateinclusterc Nodesc i ← list of Nodes from c for each node n in Nodesc i do W E HAVE TWO DIMENSIONS OF EACH COORDINATE ( LATI - TUDE , LONGITUDE ) FOR EUCLIDEAN DISTANCE d ( n, Nodec i ) = » (cid:80) j =1 ( n j − Nodec ij ) IF d ( n, Nodec i ) < = r THEN n ← Cluster HEAD Nodec i END IF
END FOR
END FOR
END PROCEDURE
IV. P
ERFORMANCE E VALUATION
This section examines, evaluates, and compares the effi-ciency of DAIS with the other investigated approaches, undera D2D communications network with a small number of UEs.
A. Methodology
First, the performance of DAIS for a scenario with asmall number of D2D Devices ( < = 200), as compared tothe number of D2D Devices in [2] which rose up to 1000,is investigated, while varying the device Battery Power Leveland the WDR thresholds. For this, a “brute force” investigationof the aforesaid thresholds was executed with values from 0%to 100% using a step of 5%. • The Battery Power Level threshold determines the mini-mum value (in %) that a D2D Device must have in theremaining battery, in order to become D2D Relay or D2DMulti-Hop Relay and accept connections from other UEs.More specifically, a D2D Relay or D2D Multi-Hop RelayDevice will admit connections from new D2D Devicesentering the Network only when their battery power levelis greater than or equal to the battery threshold. Thereason for utilizing the battery power level threshold is:i) fairness; ii) network stability and iii) longevity. • The WDR threshold determines: i) the minimum WDRthat an existing D2D Device operating as D2D Relay/D2D Multi-Hop Relay must have in order for a newD2D Device entering the network to connect to it; orii) the maximum WDR that a new D2D Device enteringthe D2D Network must have in order to replace a D2DDevice operating as D2D Relay/ D2D Multi-Hop Relayand take its role. The WDR threshold is used by thealgorithm for four purposes. More specifically, throughthe WDR threshold, new D2D Device entering in theNetwork: – Can perform a quality check of the D2D Relay, in orderto connect to it as a D2D Client. – Can perform a quality check of the D2D Multi-HopRelay, in order to connect to it either as a D2D Clientor a D2D Relay. – Can perform a replacement of a D2D Relay/ D2DMulti-Hop Relay device and take its role, if the newD2D device’s WDR is greater than the WDR of theexisting D2D Relay/D2D Multi-Hop Relay device. – Can connect to a D2D Relay/ D2D Multi-Hop RelayDevice in its proximity, and act as a D2D Relay.In addition, the effect the Transmission Power (TP) hason the investigated approaches, in terms of overall total PCand total SE achieved, is also investigated and demonstrated.For the communication power a “brute force” investigationwas executed with values from 160 mW to 60 mW using adecreasing step of 10 mW.The FuzzyART, DBSCAN and MEC AI/ML unsupervisedlearning clustering techniques are compared with the DAISalgorithm, the Random clustering approach and the Sum RateApproach (shown in [2]) in a D2D communication network.The case where D2D communication is not used is alsocompared (we refer to this as non-D2D-UE approach). TheFuzzyART, DBSCAN and MEC AI/ML are unsupervisedlearning clustering techniques that separates UEs into clusters(hence implement ultra-dense networks) under the BS, by uti-lizing distances, like the Euclidean Distance, as a metric. Then,the heuristic algorithm, that we developed (and presented inAlgorithm 1), utilizes the clustering results extracted by theseapproaches, and selects the D2D Device in the identifiedclusters with the best Data Rate to be set as D2D Relay nodeand made D2D Relay Cluster Head (CH). Once the D2D RelayCH is selected, the algorithm assigns the UEs within a radiusof 200m (WIFI Direct) from the D2D Relay CH, to becomeD2D Clients of the cluster and connect to it. Also, UEs notwithin the radius will stay connected to the BS (non-D2D-UEs).The Sum Rate Approach is utilizing distributed control.With this approach, each node adds the data rate of theconnections (that is the Sum Rate) that each D2D Devicehas in the D2D communication network. Then it decides thebest transmission mode, best link and best path to the BS orother Gateway, in order to achieve the maximum Sum Rateof the whole network. The Random approach is a simpleapproach that selects the Transmission mode of each nodein a random manner. The non D2D UE approach describesthe current approach used in Mobile Networks. This approachkeeps all the UEs connected directly to the BS and a constantpredefined transmission power, that is specified for the UEsthat are directly connected to the BS, is used.
B. Simulation Environment
In order to investigate how to achieve the best results in anetwork with a low number of D2D devices, a range of 1 to200 D2D Devices were used. The devices are placed in a cellrange of 1000 meter radius from the BS using a Poisson Pointig. 2: Total SE & Total PC vs Battery Power Level (BPL)ThresholdProcess distribution model. In our simulation environment wekeep the same comparison measurements of performance andthese are the Total SE (Sum rate), Total PC and ExecutionTime as in [2]. Also, the Channel State Information (CSI)used in the investigation is the Statistical CSI. In addition wekeep the same formulas for D2D UEs battery power levelestimation and WDR and the same simulation constraints andsimulation parameters. However, we introduce new constraintsand parameters as in section III. The simulation environmentis implemented in Java (i.e. Java 11.0 with Apache Netbeans11.6 IDE) using the JADE Framework [28], LTE/5G Toolboxlibraries from Matlab (2020a) and also the SMILE library thatis used for AI/ML implementation. The hardware used for thesimulation is the following: i) an Intel(R) Core(TM) i7-8750HCPU @ 2.20GHz; ii) 24 GB DDR4; iii) 1TB SSD hard disk;and iv) NVIDIA GeForce GTX 1050 Ti graphics card with4GB DDRS5 memory.
C. Results1) Evaluation of DAIS Approach:
The results related to theperformance of DAIS are illustrated in Fig. 2 and Fig. 3. Notethat for the results provided, a “brute force” investigation wasexecuted, by varying the Device Battery Power Level (in %)and the Weighted Data Rate (WDR) Thresholds with valuesfrom 0% to 100% using a step of 5%. During this investigationthe optimum thresholds were also selected. As observed fromthe results (see Fig. 2), varying the Device Battery power levelthreshold does not cause noticeable changes on the total PCnor the sum rate (i.e., total SE).On the other hand, by varying the WDR Threshold, weobserve that the results are considerably affected, in terms ofSE and PC. More specifically, as shown in Fig. 3, with adifferent number of D2D Devices and different values for theWDR threshold there are major changes in the resulting totalPC and total SE. However, in order to achieve these results atleast a number of 75 D2D Devices must exist under the BS.Furthermore, as depicted in Fig. 3, the WDR threshold valueachieving optimized results is 20% (see section IV-A for anexplanation on the use of this threshold). Fig. 3: Total SE & Total PC vs WDR Threshold
2) Effect of Transmission Power Alteration on the Inves-tigated Approaches:
The effect that the transmission powerhas on the investigated approaches, in terms of total PC andtotal SE (sum rate) achieved, are illustrated in Fig. 4 andFig. 5. As observed, by altering the transmission power ofthe communication and the number of UEs (D2D Devices)gains are provided on the total PC with a small trade off onthe SE.More specifically, by altering (decreasing) the transmissionpower, the following observations are made: i) for the scenar-ios with low number of UEs (i.e., up to 100 UEs), there isnoticeable improvement on the total network PC (i.e., up to64.10% for DAIS; Fig. 5), with a small decrease on the SE(i.e., a maximum of 20% decrease for DBSCAN shown in Fig.4); ii) for the scenarios with more than 100 UEs, significantgains are also observed on the total PC (i.e., up to 66.10 %decrease for MEC; Fig. 5) but with minor decrease on the SE(i.e., a maximum of 13% decrease of Random; Fig. 4).In addition, as shown in Fig. 5, for all approaches compared(except the non-D2D-UE), the values of total PC changerapidly from 0 UEs to 200 UEs, but they do not have a largescale of difference in each approach. On the other hand, forthe non-D2D-UE approach the total PC used compared to allother approaches is significant. The reason is that with thisapproach, all the UEs have direct connections with the BS,which are power consuming.
3) Performance Comparison of the Investigated Ap-proaches:
In this section, the performance of the approachesis compared in terms of total SE (Sum Rate) and total PCachieved. For this comparison a predefined transmission powerof 160 mW is used for all approaches as shown in Fig. 6.As depicted in Fig. 6, in terms of total power needed (i.e.,power consumption), the best results are provided by the SumRate Approach, while the worst performance is observed forthe non-D2D-UE approach. In addition, all approaches arerelatively close, in terms of total SE from a range of UEs of0 to 50. Beyond 50 UEs, the DAIS and Sum Rate Approach,approaches start to show increased SE and they conclude tohave better SE than other centralized AI approaches as shownin Fig. 6.ig. 4: Total SE vs Transmission Power (TP) of DifferentApproaches 5/50/100/200 UEsFig. 5: Total PC vs Transmission Power (TP) of DifferentApproaches 5/50/100/200 UEsFig. 6: Total Spectral Efficiency & Total Power Needed vsNumber of devices of Different Approaches TABLE II: Number of Devices, Devices Under Cluster, Re-sulting non D2D UEs, Number of Messages and Number ofClusters per Approach
DAIS non-D2D UE
65 13
121 19
230 26
Sum Rate Approach Random
17 34 294 7 25 75 151 1
34 43 595 7 49 150 300 1MEC
17 42 213 5
38 92 414 4
In terms of SE, DAIS seems to under-perform comparedto the other approaches for a network with a small numberof devices (i.e., 10 UEs as shown in Fig. 4 at 160 mW).However, from 50 UEs and above, DAIS is better than theDBSCAN, Random and non-D2D-UE approaches as shownin Fig. 4 and in Fig. 6. Finally, at 200 UEs (maximumnumber of UEs examined) DAIS really shows its benefitsby reaching the results provided by the Sum Rate Approachshown in Figure 6. Continuing our examination on total PC,DAIS outperforms the non-D2D-UE approach for all numberof devices examined. Furthermore, at 200 devices the DAISis better than DBSCAN, MEC, non-D2D-UE approach andRandom, but it has the same total PC with FuzzyART; Fig. 6.The non-D2D-UE approach has the worst performance interms of total PC, compared to all other related approaches(the change percentage in total PC for non-D2D-UE approachis 12.50% for 5 devices and 4% from 5 devices to 200devices), as shown in Fig. 6. In terms of SE, it provides betterperformance than other approaches only when the numberof UEs in the Network is 10 or less (as shown in Fig. 4).However, for more UEs it provides the worst results in termsof SE. Additionally, below 50 UEs, the non-D2D-UE approachhas better SE than DAIS. However, in the examined range ofnumbers of UEs (0 until 200) DAIS has better total powerusage for communications rather than non-D2D-UE approach;this is shown in figure 6.Random approach is always the worst than all other ap-proaches in terms of SE (as shown in Fig. 6). However,Random provides better performance in terms of total PCcompared to the non-D2D-UE approach (as shown in Fig. 6).Additionally, in our examination we investigated some extracharacteristics of each algorithm and compared the perfor-mance of the different approaches in terms of number ofmessages exchanged, number of resulting non-D2D UEs,number of clusters formed and total number of devices undercluster. The results are provided in Table II. Messages is an important factor in the delay of the execution of the controlmechanism of an algorithm. The number of messages in addition to thepackets/frames overhead that must be included in each message (TCP/IP)makes this factor of significant importance in the selection of the mostappropriate approach for D2D communication Transmission Selection. egarding the number of messages that each approach needsto exchange in order to conclude on the Transmission modeselection for all runs , from the worst to best performanceis provided by Sum Rate Approach, FuzzyART, MEC, DB-SCAN, DAIS, non-D2D UE and Random.Additionally, for all runs, with the only approaches thatall UEs finally conclude to become D2D Devices are DAIS,Sum Rate Approach and Random approach. For the rest ofthe approaches, FuzzyART has the least number of resultingnon-D2D UEs followed by MEC and DBSCAN.In terms of the created clusters, the total number of usersthat are served by cluster (D2D Relay/D2D Multi Hop Relaythat are directly connected to BS are not included) and numberof clusters created per approach are investigated. The benefitsof having a large number of D2D Devices under a cluster aresignificant for the SE and PC. More specifically, by havinga large number of D2D Devices under a cluster the total SEis increased, total PC is reduced and the number of directlinks to BS are decreased. On the other hand, in the case ofa large number of Clusters the links to BS are reduced butSE may not be affected effectively. Moreover, balancing ofboth metrics can be achieved with maximum SE, minimum PCand reduced number of links to BS for large towards mediumnumber of clusters with equal assigned D2D Client Devices.Therefore, by investigating the clusters density and numberof clusters the following results are provided: i) for 50 UEsthe maximum number of devices that can be included in acluster is provided by DBSCAN (10) and then MEC (9) withthose establishing 1 and 5 clusters respectively. The MEC isthe second in order, but DBSCAN is in the last approaches interms of Total SE/PC; ii) for 100 UEs the maximum number ofdevices that can be included in a cluster is provided by DAIS(97) with 19 clusters established and then by DBSCAN (25)with 1 cluster. The DAIS is the third in order, but DBSCANis in the last approaches in terms of Total SE/PC; and iii) for200 UEs the maximum number devices that can be includedin a cluster is provided by DAIS (146) with 26 clusters andthen DBSCAN (49) with 1 cluster. The DAIS is the secondin order, but DBSCAN is in the last approaches in terms ofTotal SE/PC.In our analysis, we examine the mean time of execution ofeach approach (centralized, distributed, semi distributed andDAI) in terms of the duration of the calculation of Trans-mission Mode Selection of a D2D Device. More precisely,we calculate for each approach the mean time when thealgorithm started to compute the transmission mode until theconclusion of the algorithm in each run (for different numbersof UEs). For example with 50 Devices in centralized andsemi-distributed mode, the procedure computes the sum ofexecution time from 1..50 UEs of each iteration when theapproach examines 1,2,3,4..,50 Devices and then it dividesthe result with the number of devices (50). However, for thedistributed mode the time is calculated in each D2D Device Run is the execution of the algorithm with a different number of UEs ineach instance of the scenario
TABLE III: Time of execution of each approach (1 = 100 ms)
Number of Devices DAIS non-D2D UE Sum Rate Approach DR FuzzyART DBSCAN GMEANS MEC
50 0 0 1 0 0 1 0 15100 1 1 12 0 2 7 1 36200 1 1 12 0 2 7 1 36 and at the end the sum of the calculated times divided withthe number of devices is the resulting execution time. Notethat in Sum Rate Approach there is a need to investigate foreach D2D Device all transmission modes and links in order toachieve the best sum rate (this is the reason it is slow). On theother hand for centralized approaches the duration depends onthe calculation of the transmission mode selection of the wholenetwork. Overall, the faster approach is the DAIS (DAI) with100 ms with any UE (from 1..200 UEs), the second faster isthe DR with the non-D2D UEs, the most slowest approachesare MEC, DBSCAN (centralized) and Sum Rate Approach(distributed) as shown in the Table III.
4) Overall Remarks:
In the performance comparison pro-vided above the different investigated approaches are evaluatedin terms of SE and PC. The results illustrated that the worstperformance is provided by the Random approach, whilethe best performance is provided by Sum Rate Approach,FuzzyART and DAIS. On the other hand, in terms of total PC,the worst performance is provided by non-D2D-UE approach,while best is provided again by the Sum Rate Approach, DAISand FuzzyART.Additionally, the paper shows that unsupervised learningapproaches such as FuzzyART can be used for transmissionmode selection in D2D Communication. In addition, by con-sidering Table II, we observe that Sum Rate Approach needsto exchange a lot of messages before a decision is established,this is the reason that is taking a lot of time to conclude.Also, another observation that is made in this investigationis that, compared to all other investigated approaches, DAIScreates the greatest amount of clusters with the greatest amountof D2D Clients in each cluster, however without alwaysproviding the best performance in terms of SE and PC (e.g.,for 50 UEs, Sum Rate Approach provides the best performancewith 6 D2D Clients and 12 clusters in contrast to DAIS with6 D2D Clients and 13 Clusters). Also, it is observed thateven if DBSCAN creates only one cluster , it achieves betterresults than the non-D2D UE approach. In addition, it is shownin Table II and in Figure 6 that all investigated approachesexcept Random and non-D2D-UE approaches create clustersin the most accurate positions (increased SE/reduced totalPC) with the use of WDR (i.e. DAIS) and sum rate (i.e.Random, FuzzyART, MEC, DBSCAN, Sum Rate Approach)measurements under the mobile network in the D2D network.Therefore, the approaches are good alternatives to be usedfor Transmission mode selection in the D2D communication.In addition, the following findings extracted from Fig. 6 andTable II: i) Some of the 5G requirements are achievablethrough Transmission Mode Selection (i.e. High Data Rates, Because in our investigation: i) parameters are restricted and pre specifiedwith the use of WiFi Direct (i.e. 255 UEs per D2DR and a radius of 200 m);and ii) is restricted with the use of a small number of UEs. ow Power Consumption); ii) The critical point that SE, PCgains increases rapidly is 100 UEs for all approaches; iii)coverage expansion is achieved; and iv) the lower limit ofall approaches is 5 UEs.V. C
ONCLUSIONS AND F UTURE W ORK
The research objective on this paper is threefold. Firstly,it examines the performance of the DAIS algorithm with theproposed changes in threshold (i.e., WDR Threshold), in termsof SE and PC, considering scenarios with a small number ofDevices (i.e., < = 200). During this examination, the WDRand the BPL DAIS’ thresholds, affecting the SE and PC ofthe network, have been examined and values achieving bestperformance have been determined. Secondly, it introduces theuse of unsupervised learning AI/ML approaches in Transmis-sion mode selection in D2D Communication and compares theperformance of DAIS with FuzzyART, DBSCAN and MECas well as other related approaches (i.e., Distributed Random,Distributed Sum Rate Approach, Centralized non-D2D-UE).Last, it examines the effect the transmission power has onthe investigated approaches, in terms of PC and SE achieved.The results obtained demonstrated that DAIS, compared to allother related approaches, with the right tuning of WDR andBPL threshold values, can provide significant gains in terms ofSE, PC, and cluster formation efficiency. Precisely, the resultsshowed that DAIS and Sum Rate Approach outperformedall other approaches in terms of SE. FuzzyART, DAIS andSum Rate Approach outperformed all other related approachesin terms of PC. Additionally, our findings showed that, byreducing the transmission power of communication, the SEand PC of the network is significantly affected (SE in anegative way and PC in a positive way) when the amountof UEs is less than 100. On the other hand, from 100 to 200UEs the effect on SE becomes smoother while on PC the gainsremain the same. Also, results showed that the investigatedAI/ML approaches are also beneficial for Transmission modeselection in D2D communication, even with a small number ofDevices. As future work we will investigate the performanceof the same AI/ML approaches in scenarios with large numberof UEs (i.e., up to 1000 UEs under the same BS) consideringnon ideal CSI in D2D communication network.R EFERENCES[1] I. F. Akyildiz, S. Nie, S. C. Lin, and M. Chandrasekaran, “5G roadmap:10 key enabling technologies,”
Computer Networks , vol. 106, pp. 17–48,2016.[2] I. Ioannou, V. Vassiliou, C. Christophorou, and A. Pitsillides, “Dis-tributed artificial intelligence solution for d2d communication in 5gnetworks,”
IEEE Systems Journal , vol. 14, pp. 4232–4241, Sep. 2020.[3] G. Fodor, E. Dahlman, G. Mildh, S. Parkvall, N. Reider, G. Mikl´os,and Z. Tur´anyi, “Design aspects of network assisted device-to-devicecommunications,”
IEEE Communications Magazine , vol. 50, pp. 170–177, mar 2012.[4] P. Gandotra and R. Jha, “Device-to-Device Communication in CellularNetworks: A Survey,”
Journal of Network and Computer Applications ,vol. 71, no. 4, pp. 1801–1819, 2016.[5] M. Ahmad, M. Azam, M. Naeem, M. Iqbal, A. Anpalagan, and M. Ha-neef, “Resource management in D2D communication: An optimizationperspective,”
Journal of Network and Computer Applications , vol. 93,no. December 2016, pp. 51–75, 2017. [6] G. A. Carpenter, S. Grossberg, and D. B. Rosen, “Fuzzy ART: Fast stablelearning and categorization of analog patterns by an adaptive resonancesystem,”
Neural Networks , vol. 4, pp. 759–771, jan 1991.[7] S. G. Akojwar and R. M. Patrikar, “Real time classifier for industrialwireless sensor network using neural networks with wavelet preproces-sors,” in
Proceedings of the IEEE International Conference on IndustrialTechnology , pp. 512–517, 2006.[8] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based algorithmfor discovering clusters in large spatial databases with noise,” pp. 226–231, AAAI Press, 1996.[9] M. Li, D. Meng, S. Gu, and S. Liu, “Research and improvementof DBSCAN cluster algorithm,”
Proceedings - 2015 7th InternationalConference on Information Technology in Medicine and Education,ITME 2015 , pp. 537–540, 2016.[10] F. Golchin and K. K. Paliwal, “Minimum-entropy clustering and itsapplication to lossless image coding,”
IEEE International Conferenceon Image Processing , vol. 2, pp. 262–265, 1997.[11] Haifeng Li, Keshu Zhang, and Tao Jiang, “Minimum entropy clusteringand applications to gene expression analysis,” pp. 136–145, 2004.[12] J. Deng, A. A. Dowhuszko, R. Freij, and O. Tirkkonen, “Relay se-lection and resource allocation for D2D-relaying under uplink cellularpower control,” , pp. 1–6, 2015.[13] G. Steri, G. Baldini, I. N. Fovino, R. Neisse, and L. Goratti, “A novelmulti-hop secure LTE-D2D communication protocol for IoT scenarios,” ,2016.[14] L. Song, D. Niyato, Z. Han, and E. Hossain, “Game-theoretic resourceallocation methods for device-to-device communication,”
IEEE WirelessCommunications , vol. 21, no. 3, pp. 136–144, 2014.[15] B. Peng, T. Peng, Z. Liu, Y. Yang, and C. Hu, “Cluster-based multicasttransmission for device-to-device (D2D) communication,”
IEEE Vehic-ular Technology Conference , 2013.[16] K. Doppler, C. H. Yu, C. B. Ribeiro, and P. J¨anis, “Mode selectionfor device-to-device communication underlaying an LTE-advanced net-work,”
IEEE Wireless Communications and Networking Conference,WCNC , 2010.[17] M. Jung, K. Hwang, and S. Choi, “Joint mode selection and powerallocation scheme for power-efficient device-to-device (D2D) commu-nication,”
IEEE Vehicular Technology Conference , pp. 1–5, 2012.[18] C. Xu, J. Feng, B. Huang, Z. Zhou, S. Mumtaz, and J. Rodriguez,“Joint relay selection and resource allocation for energy-efficient D2Dcooperative communications using matching theory,”
Applied Sciences(Switzerland) , vol. 7, no. 5, pp. 1–24, 2017.[19] X. Ma, R. Yin, G. Yu, and Z. Zhang, “A distributed relay selectionmethod for relay assisted Device-to-Device communication system,”
IEEE International Symposium on Personal, Indoor and Mobile RadioCommunications, PIMRC , pp. 1020–1024, 2012.[20] G. Rigazzi, F. Chiti, R. Fantacci, and C. Carlini, “Multi-hop D2Dnetworking and resource management scheme for M2M communicationsover LTE-A systems,”
IWCMC 2014 - 10th International Wireless Com-munications and Mobile Computing Conference , pp. 973–978, 2014.[21] J. Gui and J. Deng, “Multi-Hop Relay-Aided Underlay D2D Communi-cations for Improving Cellular Coverage Quality,”
IEEE Access , vol. 6,pp. 14318–14338, 2018.[22] Y. Liu, “Optimal Mode Selection in D2D-Enabled Multibase StationSystems,”
IEEE Communications Letters , vol. 20, no. 3, pp. 470–473,2016.[23] S. Xiang, T. Peng, Z. Liu, and W. Wang, “A distance-dependent modeselection algorithm in heterogeneous D2D and IMT-Advanced network,” , pp. 416–420, 2012.[24] A. Mesodiakaki, F. Adelantado, L. Alonso, and C. Verikoukis, “Energy-efficient context-aware user association for outdoor small cell heteroge-neous networks,” in , pp. 1614–1619, 2014.[25] E. Pateromichelakis, M. Shariat, A. Quddus, M. Dianati, and R. Tafa-zolli, “Dynamic clustering framework for multi-cell scheduling in densesmall cell networks,”
IEEE Communications Letters , vol. 17, no. 9,pp. 1802–1805, 2013.[26] U. Siddique, H. Tabassum, and E. Hossain, “Spectrum allocation forwireless backhauling of 5g small cells,” in2016 IEEE InternationalConference on Communications Workshops (ICC)