clustering algorithms for cognitive radio networks: a survey

17
Review Clustering algorithms for Cognitive Radio networks: A survey Kok-Lim Alvin Yau a,n , Nordin Ramli b , Wahidah Hashim b , Hazal Mohamad b a Faculty of Science and Technology, Sunway University, No. 5 Jalan Universiti, Bandar Sunway, 46150 Petaling Jaya, Selangor, Malaysia b Wireless Network and Protocol Research Laboratory, MIMOS Berhad, Technology Park Malaysia, 57000 Kuala Lumpur, Malaysia article info Article history: Received 16 August 2013 Received in revised form 17 June 2014 Accepted 21 July 2014 Available online 1 August 2014 Keywords: Cognitive Radio Software dened radio Topology management Clustering Routing abstract Cognitive Radio (CR) networks enable unlicensed or Secondary Users (SUs) to sense for and operate in the underutilized spectrum (or white spaces) owned by licensed or Primary Users (PUs) without causing unacceptable interference to the PUs' activities. Clustering, which is a topology management mechan- ism, organizes nodes into logical groups in order to provide network-wide performance enhancement. Clustering aims to achieve network scalability and stability, as well as to support cooperative tasks, such as channel sensing and channel access, which are essential to CR operations. While clustering has been well investigated in traditional networks such as mobile ad hoc networks, similar investigations in CR networks remain in the infancy stage. New clustering algorithms must be designed to address new challenges associated with the intrinsic characteristics of CR, namely the dynamicity of channel availability that changes with time and location. This article reviews clustering algorithms, and they are characterized by clustering objectives, metrics and the number of hops in each cluster. We also present complexity analysis, performance enhancements achieved by the clustering algorithms, as well as open issues, in order to establish a foundation for further research and to spark new research interests in this area. & 2014 Elsevier Ltd. All rights reserved. 1. Introduction Cognitive Radio (CR) (Akyildiz et al., 2006) enables unlicensed users (or Secondary Users, SUs) to sense for licensed users' (or Primary Users, PUs) underutilized channels (called white spaces), and subsequently uses the channels in an opportunistic manner conditional on the interference to the PUs being below an acceptable level. IEEE 802.22 is a standard for CR networks that uses white spaces in the television frequency bands; and it is designed for centralized networks comprised of base stations and wireless hosts (or Customer Premise Equipment (CPE)) (Fan and Rocky, 2009). A distributed Cognitive Radio network is comprised of a number of SUs who communicate with each other in the absence of xed network infrastructure such as an access point or a base station. This article focuses on clustering in distributed CR networks. Clustering organizes nodes into clusters in order to provide network-wide performance enhancement. Generally speaking, there are three main advantages brought about by clustering to CR net- works, namely scalability, stability, and supporting cooperative tasks, such as channel sensing and channel access, which are essential to CR operations, and these advantages have led to the use of clustering in CR networks. An intrinsic characteristic of CR networks that warrants further investigation on clustering is dynamicity of channel availability in which the channel availability (or white spaces) of each SU is different, and it changes with the level of PU activities, time and location. Popular traditional clustering algorithms, such as lowest ID (Ephremides et al., 1987) and maximum node degree (Jeng and Jan, 2007), may not be suitable, and so various new clustering algorithms have been proposed. The lowest ID clustering algorithm selects a node with the lowest ID as the leader of a cluster (or clusterhead); while the maximum node degree clustering algorithm selects a node with the highest number of neighbor nodes as the clusterhead. There are three reasons why traditional clustering algorithms applied to wireless ad hoc and sensor networks (Peiravi et al., 2013; Lee and Lee, 2013; Khan et al., 2011) are not suitable for CR networks. Firstly, the algorithms do not adapt to the channel dynamics in CR networks. Secondly, the algorithms may not achieve objectives speci c to CR networks. For instance, a common channel may not be established due to the lack of white spaces, so some nodes in a cluster may not be able to communicate with the clusterhead. Thirdly, the algorithms may not enhance network performance pertinent to CR networks such as achieving a higher number of common channels in a cluster, and application-specic network performance, such as achieving lower error probability in channel sensing outcomes. Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/jnca Journal of Network and Computer Applications http://dx.doi.org/10.1016/j.jnca.2014.07.020 1084-8045/& 2014 Elsevier Ltd. All rights reserved. n Correspondence author. Tel.: þ60 3 7491 8622x3216. E-mail addresses: [email protected] (K.-L. Yau), [email protected] (N. Ramli), [email protected] (W. Hashim), ha[email protected] (H. Mohamad). Journal of Network and Computer Applications 45 (2014) 7995

Upload: hafizal

Post on 03-Feb-2017

217 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Clustering algorithms for Cognitive Radio networks: A survey

Review

Clustering algorithms for Cognitive Radio networks: A survey

Kok-Lim Alvin Yau a,n, Nordin Ramli b, Wahidah Hashim b, Hafizal Mohamad b

a Faculty of Science and Technology, Sunway University, No. 5 Jalan Universiti, Bandar Sunway, 46150 Petaling Jaya, Selangor, Malaysiab Wireless Network and Protocol Research Laboratory, MIMOS Berhad, Technology Park Malaysia, 57000 Kuala Lumpur, Malaysia

a r t i c l e i n f o

Article history:Received 16 August 2013Received in revised form17 June 2014Accepted 21 July 2014Available online 1 August 2014

Keywords:Cognitive RadioSoftware defined radioTopology managementClusteringRouting

a b s t r a c t

Cognitive Radio (CR) networks enable unlicensed or Secondary Users (SUs) to sense for and operate inthe underutilized spectrum (or white spaces) owned by licensed or Primary Users (PUs) without causingunacceptable interference to the PUs' activities. Clustering, which is a topology management mechan-ism, organizes nodes into logical groups in order to provide network-wide performance enhancement.Clustering aims to achieve network scalability and stability, as well as to support cooperative tasks, suchas channel sensing and channel access, which are essential to CR operations. While clustering has beenwell investigated in traditional networks such as mobile ad hoc networks, similar investigations in CRnetworks remain in the infancy stage. New clustering algorithms must be designed to address newchallenges associated with the intrinsic characteristics of CR, namely the dynamicity of channelavailability that changes with time and location. This article reviews clustering algorithms, and theyare characterized by clustering objectives, metrics and the number of hops in each cluster. We alsopresent complexity analysis, performance enhancements achieved by the clustering algorithms, as wellas open issues, in order to establish a foundation for further research and to spark new research interestsin this area.

& 2014 Elsevier Ltd. All rights reserved.

1. Introduction

Cognitive Radio (CR) (Akyildiz et al., 2006) enables unlicensedusers (or Secondary Users, SUs) to sense for licensed users' (orPrimary Users, PUs) underutilized channels (called white spaces),and subsequently uses the channels in an opportunistic mannerconditional on the interference to the PUs being below anacceptable level. IEEE 802.22 is a standard for CR networks thatuses white spaces in the television frequency bands; and it isdesigned for centralized networks comprised of base stations andwireless hosts (or Customer Premise Equipment (CPE)) (Fan andRocky, 2009). A distributed Cognitive Radio network is comprisedof a number of SUs who communicate with each other in theabsence of fixed network infrastructure such as an access point ora base station. This article focuses on clustering in distributed CRnetworks.

Clustering organizes nodes into clusters in order to providenetwork-wide performance enhancement. Generally speaking, thereare three main advantages brought about by clustering to CR net-works, namely scalability, stability, and supporting cooperative tasks,

such as channel sensing and channel access, which are essential to CRoperations, and these advantages have led to the use of clustering inCR networks. An intrinsic characteristic of CR networks that warrantsfurther investigation on clustering is dynamicity of channel availabilityin which the channel availability (or white spaces) of each SU isdifferent, and it changes with the level of PU activities, time andlocation. Popular traditional clustering algorithms, such as lowest ID(Ephremides et al., 1987) and maximum node degree (Jeng and Jan,2007), may not be suitable, and so various new clustering algorithmshave been proposed. The lowest ID clustering algorithm selects a nodewith the lowest ID as the leader of a cluster (or clusterhead); while themaximum node degree clustering algorithm selects a node with thehighest number of neighbor nodes as the clusterhead. There are threereasons why traditional clustering algorithms applied to wireless adhoc and sensor networks (Peiravi et al., 2013; Lee and Lee, 2013; Khanet al., 2011) are not suitable for CR networks. Firstly, the algorithms donot adapt to the channel dynamics in CR networks. Secondly, thealgorithms may not achieve objectives specific to CR networks. Forinstance, a common channel may not be established due to the lack ofwhite spaces, so some nodes in a cluster may not be able tocommunicate with the clusterhead. Thirdly, the algorithms may notenhance network performance pertinent to CR networks such asachieving a higher number of common channels in a cluster, andapplication-specific network performance, such as achieving lowererror probability in channel sensing outcomes.

Contents lists available at ScienceDirect

journal homepage: www.elsevier.com/locate/jnca

Journal of Network and Computer Applications

http://dx.doi.org/10.1016/j.jnca.2014.07.0201084-8045/& 2014 Elsevier Ltd. All rights reserved.

n Correspondence author. Tel.: þ60 3 7491 8622x3216.E-mail addresses: [email protected] (K.-L. Yau),

[email protected] (N. Ramli), [email protected] (W. Hashim),[email protected] (H. Mohamad).

Journal of Network and Computer Applications 45 (2014) 79–95

Page 2: Clustering algorithms for Cognitive Radio networks: A survey

The main contribution of this article is to present an extensivereview on the various aspects of clustering algorithms in distrib-uted CR networks, including clustering objectives, clusteringcharacteristics (i.e. metrics and intra-cluster distance), perfor-mance enhancements, complexity analysis (i.e. time and messagecomplexities), as well as open issues. It focuses on clusteringalgorithms, particularly clustering metrics and how these metricshave been applied to form clusters in CR networks, rather thanapplication schemes that apply the cluster structure such ascluster-based routing that focuses on the enhancement of routingrather than clustering (Talar and Altilar, 2011), and it does notfocus on the traditional coordination mechanisms for clusteringalgorithms because there has been a considerable amount ofliterature being published in the context of mobile and static adhoc (or mesh) networks. Since clustering algorithms have beentraditionally incorporated into the network layer to form clusters,which help to limit the flooding of routing overheads (e.g. RouteRequest and Route Reply) throughout the entire network(Ephremides et al., 1987; Jeng and Jan, 2007), the other layerssuch as the application and physical layers are not discussed in thisarticle. Additionally, the contributions are to The purposes of thisarticle are to establish a foundation and to spark new interests inthis emerging research area. Note that, for simplicity, the termsSUs and nodes are used interchangeably throughout the entirearticle. The organization of this article is as follows. The rest ofSection 1 presents an overview of CR networks, as well as anoverview, advantages and challenges of clustering in CR networks.Section 2 presents taxonomy of the attributes of clustering algo-rithms in CR networks. Section 3 presents various clusteringalgorithms in CR networks, as well as to relate them to theattributes of clustering algorithms presented in Section 2.Section 4 presents performance enhancements achieved by clus-tering algorithms. This section also presents complexity analysis ofthe clustering algorithms. Section 5 presents open issues. Section 6presents conclusions.

1.1. Cognitive radio: an overview

The traditional spectrum allocation policy has been partitioningradio spectrum into smaller ranges of licensed and unlicensedfrequency bands (also called channels). The licensed channels,which are auctioned off by the government, provide exclusivechannel access to PUs. SUs, such as the popular wireless commu-nication systems IEEE 802.11, are forbidden to access any of thelicensed channels. Instead, they access unlicensed channels with-out incurring any monetary cost.

Cognitive Radio enables SUs to sense radio spectrum and usewhite spaces whilst minimizing interference to PUs. The purposeis to improve the availability of bandwidth at each SU, and henceimproving the overall utilization of radio spectrum, which is one ofthe scarcest resources in wireless communications. The maindifference between CR and the traditional wireless networks isthe presence of PU activities in CR networks; hence the mainchallenge of CR is to establish a “friendly” environment, in whichthe PUs and SUs coexist without causing interference with eachother as shown in Fig. 1. In Fig. 1, a SU switches its operatingchannel across various channels from time to time in order toutilize white spaces in the licensed channels. Note that, each SUmay observe different white spaces, which are time and locationdependent. For a successful communication, a particular whitespace must be available at both SUs in a communication node pair.

Opportunistic spectrum access, which can be realized using CR,has been approved in 2004 in US through Notice of Proposed RuleMaking (NPRM) by Federal Communications Commission (FCC)(2006), and in 2007 in UK through Digital Dividend Review (DDR)by Office of Communications (Ofcom) (2007). Upon completion of

the digital television switchover, which replaces the analogterrestrial TV services with the digital terrestrial TV services calledDigital Video Broadcasting – Terrestrial (DVB-T), there will bevacant spectrum called digital dividend. In UK, opportunisticspectrum access has been proposed in the award of digitaldividend, and no digital dividend will be awarded exclusively tolicensed users (Office of Communications (Ofcom), 2007), whichhas strengthened the essential role of CR in the near future. Thereare approximately 128 MHz and 90 MHz of digital dividend in UK(Rayment et al., 2009) and Korea (Kim et al., 2010), respectively. CRhas become more and more prevalent since the digital televisionswitchover has completed in year 2009 in US and 2012 in UK(Nekovee, 2008) with some CR products (e.g. xG technology(2013)) emerging in the market.

Recently, there has been growing research interest in CognitiveRadio Sensor Network (CRSN). CRSN incorporates CR capabilityinto the traditional Wireless Sensor Networks (WSNs) so that eachsensor node can sense radio spectrum and use white spaces (Ozgerand Akan, 2013) while detecting and monitoring physical andenvironmental events. Generally speaking, in CRSNs, each sensornode inherits important characteristics of WSNs, particularlythe limitation of hardware and energy resources. Hence, one ofthe main objectives of CRSNs is to reduce energy consumption.The main difference between CRSNs and the traditional CR net-works is that, data packets are generated by nodes that havedetected event(s) and are sent to a sink node; while in thetraditional CR networks, data packets may be generated by anynodes and are sent to a base station.

1.2. Clustering: an overview

Clustering, which is a topology management mechanism,organizes nodes into logical groups (or clusters) in order toprovide network-wide performance enhancement. Figure 2 showsan example of a cluster structure in which nodes in a CR networkare grouped to form clusters. Note that, the formed clusterstructure depends on the underlying network, such as the locationand channel availability (or white spaces) of the nodes. The SUsform three clusters (i.e. C1, C2, C3). Each cluster comprises threekinds of nodes, namely clusterhead, member node and gatewaynode. A clusterhead (i.e. CH1, CH2, CH3) serves as a local point ofprocess for various applications such as channel sensing, which isessential to CR operation, and routing. A member node associatesitself with a clusterhead. For instance, member nodes M1,1, M1,2,M1,3, M1,4, M1,5 are associated with clusterhead CH1. Clusterheadand member nodes communicate regularly among themselves,and these are called intra-cluster communications. The gatewaynodes, which are the member nodes located at the fringe of acluster, can hear from neighboring clusters, and so they provideinter-cluster communications. For instance, gateway node M1,2 isassociated with clusterhead CH1, and it provides inter-clustercommunications for clusterheads CH1 and CH2. Since the

Fig. 1. A SU exploits white spaces across various channels.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9580

Page 3: Clustering algorithms for Cognitive Radio networks: A survey

neighboring clusters may use distinctive channels, which iscommonplace in CR networks, a gateway node may need to switchits operating channel regularly. The clusterheads and gatewaynodes form a backbone to the base station. For instance, inFig. 2, nodes M1,5–CH1–M1,2–CH2–M3,3–CH3–SU BS form a back-bone. The number of hops between member nodes and cluster-head in a cluster (or cluster size) may be a single, two (Huang etal., 2011) or more.

The main difference among the clustering algorithms is thecluster formation procedure, which generally comprises cluster-head selection andmember node joining. A node initiates the clusterformation procedure whenever it fails to find any clusters nearby.Local information, such as a list of available channels, is exchangedamong neighboring nodes to form clusters in line with someglobal objectives, such as higher number of common channels in acluster and higher network-wide throughput (Badoi et al., 2011).During the clusterhead selection procedure, a clusterhead iselected from nodes in a cluster based on certain clustering metrics,such as channel availability. For instance, a clusterhead has thehighest number of single-hop neighbor nodes (or node degreelevel) in Jeng and Jan (2007) and Badoi et al. (2011), and the lowestnode ID in Ephremides et al. (1987). In Huang et al. (2011), aclusterhead has higher node degree level, as well as lower numberof hops and number of channel switches from member nodes to theclusterhead. During the member node joining procedure, a nodechooses a cluster to join based on some criteria, such as the numberof common channels between a node and a cluster (Badoi et al.,2011). Higher number of common channels in a cluster prevents re-clustering due to the lack of a common channel as a result of the re-appearance of PU activities. Note that, a node may reduce thenumber of common channels in a cluster upon joining the clustercausing instability, and so careful consideration must be made.

We present an example of cluster formation using Fig. 2.Denote a set of available channels, which are not occupied byPUs, by K. Suppose, the available channels at each node in clusterC1 is KCH1 ¼ f1;2; 3g; KM1;1 ¼ f1;2; 3g; KM1;2 ¼ f2;3;4g; KM1;3 ¼f2;3;4g; KM1;4 ¼ f2;3;5g and KM1;5 ¼ f2; 3;6g; while the node

degree level is DCH1 ¼ 5; DM1;1 ¼ 1; DM1;2 ¼ 2; DM1;3 ¼ 1; DM1;4 ¼ 1and DM1;5 ¼ 1. The local information (i.e. a set of available channelsand node degree level) is exchanged among neighboring nodes toform clusters in line with some global objectives such as improv-ing the network stability. To address the dynamicity of channelavailability and network topology, network stability can beachieved by increasing the number of common channels amongnodes in a cluster, and the node degree level of clusterhead andgateway nodes, respectively. Hence, clusterhead CH1 is selectedbecause it provides the highest number of common channels (i.e.channels 2 and 3), and it has the highest node degree levelDCH1 ¼ 5 among nodes in its neighborhood. Additionally, gatewaynode M1,2 is selected because it can hear from clusterheads CH1

and CH2.We present an example of the steps involved in the coordina-

tion mechanism for cluster formation in which single-hop clustersare constructed (Chen et al., 2007). Consider a new and indepen-dent node. Being unassociated with any clusters, the node aims tobecome a member node through associating itself with a cluster orbecome a clusterhead itself. It switches from channel to channel inorder to wait for and receive beacons from clusterheads or anymessages from its neighbor nodes. There are three cases in regardsto the reception of clusterhead's beacons and neighbor node'smessages. Firstly, the node receives both beacon and message, andso it forms a cluster itself and becomes a clusterhead. Secondly, thenode receives beacon only, and so it becomes a member node ofthe respective clusterhead. Thirdly, the node receives messageonly, and so it is located two hops away from the clusterhead;subsequently, it switches to another channel in search of a single-hop clusterhead, and if it still fails to join any clusters, it maybecome a clusterhead itself. A clusterhead discovers its neighborclusters, which may be a single, two or three hops away, andsubsequently selects gateway nodes to form inter-cluster commu-nications. Nodes in a cluster switch from channel to channel inorder to discover neighboring clusters, and exchange informationwith them. Generally speaking, the clusterhead assumes severalroles, including coordinating the member nodes and serving as

Fig. 2. A cluster structure in a CR network consists of PUs and SUs.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 81

Page 4: Clustering algorithms for Cognitive Radio networks: A survey

a local point of process for essential applications including channelsensing, channel access and routing (Talay and Altilar, 2013; Liminget al., 2012). Hence, the clusterhead incurs higher energy consump-tion compared to member nodes, and so the clusterhead role may berotated among nodes in a cluster. A considerable amount of literaturehas been published on coordination mechanisms for clusteringalgorithms in the context of mobile and static ad hoc (or mesh)networks. The rest of this article focuses on clustering algorithms,such as the use of clustering metrics in clusterhead selection andmember node joining, in the context of CR networks.

1.3. Clustering in cognitive radio networks: advantages

There are three main advantages associated with clustering inCR networks: scalability, stability and cooperative tasks support.

Firstly, clustering improves scalability through the reduction ofcommunication overhead and parallelism. Traditionally, in non-clustered networks, nodes exchange information with all othernodes in the networks. However, in clustered networks, clusterheadsand gateway nodes form a backbone to the base station andexchange information, such as routing messages, among themselves;and member nodes exchange information with their respectiveclusterheads only, and so the communication overhead can bereduced. Parallelism is important to CR networks. Firstly, in clusterednetworks, clusterheads serve as local points of process for coopera-tive tasks, such as channel sensing in which each member nodesenses for white spaces, and subsequently sends the sensing out-comes to their respective clusterheads for final decisions in thepresence of PU activities. Secondly, nodes in a cluster can select alocal common control channel, rather than a global common controlchannel, which may not exist, for intra-cluster communications.

Secondly, clustering improves stability through the reductionof global effects on network-wide performance as a result of anychanges to network dynamics, such as channel availability andnetwork topology. This means that, any changes on networkdynamics cause local updates among member nodes and theirrespective clusterheads only. This is because only the membernodes and their respective clusterheads are reconfigured inresponse to the changes. Since global update is not necessary,more white spaces are available to SUs for data transmission.

Thirdly, clustered networks support cooperative tasks, such aschannel switch, channel sensing, and routing, in order to improvenetwork performance. For instance, in the aforementioned exam-ple of channel sensing, clusterheads and member nodes cooperateto minimize miss detections and false alarms leading to lower SUs'interference levels to PUs. The clustered networks have also beenchosen as a framework to implement game theoretic-basedapproach, which is a popular technique to achieve the optimalnetwork performance in CR networks (Zou and Chigan, 2009).

In Zou and Chigan (2009), a geographical clustering approach isproposed to form single-hop clusters with minimal overlapping inwhich member nodes in a cluster play a single game, which isinitiated and terminated by the respective clusterheads. Nodesthat are physically close to each other join a cluster, and this helpsto reduce clustering overhead and delay associated with the gameprocess.

1.4. Clustering: challenges

The main challenge associated with clustering is the dynami-city of channel availability, which is an intrinsic characteristic ofCR networks. Traditionally, the cluster structure changes withnetwork topology; however, in CR networks, the channel avail-ability of each node changes with time and location. This hasbrought about new challenges to clustering in CR networks inwhich the lack of common channels among nodes in a cluster maycause loss of connectivity among clusterheads, as well as theconnectivity among clusterheads and their respective membernodes. As a consequence, cluster maintenance and re-clusteringmay need to be performed more often than that in the traditionalad hoc networks in order to optimize network performance ofclusters at most of the times. Most clustering algorithms in theliterature have been proposed to address this challenge. Forinstance, in Li and Gross (2011), the clustering algorithm increasesthe number of common channels among nodes in a cluster,particularly the gateway nodes, in order to prevent re-clusteringas a result of the lack of a common channel.

2. Taxonomy of clustering attributes in cognitive radionetworks

This section presents the taxonomy of attributes (see Fig. 3)relevant to clustering in CR networks. Generally speaking, cluster-ing algorithms have been designed for static and mobile networks.Both static and mobile networks must address the dynamicity ofchannel availability; while mobile networks must address topol-ogy changes caused by nodal mobility. The rest of this sectionpresents clustering objectives and clustering characteristics.

2.1. Clustering objectives

Nodes form clusters with the objective of network performanceenhancement. In general, there are five clustering objectives asfollows:

B.1 Establishment of Common Control Channel. The common con-trol channel is used by SUs to exchange essential control

Fig. 3. Taxonomy of clustering attributes in CR networks.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9582

Page 5: Clustering algorithms for Cognitive Radio networks: A survey

messages, such as channel sensing, channel access and routingmessages. The unlicensed channels, such as Industrial, Scien-tific and Medical (ISM), may be highly utilized and so anavailable licensed channel may be selected as a commoncontrol channel. A global common control channel may notexist due to the dynamicity of channel availability; however, agroup of geographically adjacent SUs tend to share a similarset of available channels. Hence, nodes form clusters, andselect a local common control channel, which is available to allnodes in a cluster (Chen et al., 2007; Li and Gross, 2011;Baddour et al., 2011; Gong et al., 2008; Bradonjic and Lazos,2012; Li et al., 2012; Liu et al., 2012). Before the establishmentof a common control channel in a cluster, nodes may hop intovarious channels in the hope of receiving information fromneighbor nodes and gathering information for clusteringpurpose (Li et al., 2012). It shall be noted that, this clusteringobjective focuses on clustering algorithms whose main objec-tive is to establish common control channels to form clusterednetworks, and the literature on the establishment of commoncontrol channels in non-clustered networks can be found in(Kim, 2009; Liu, 2010).

B.2 Enhancement on Cluster Stability. Cluster stability (or robust-ness) improves intra-cluster and inter-cluster connectivity inthe presence of the dynamicity of channel availability. Greatercluster stability minimizes the occurrence of re-clusteringbecause re-clustering may result in sub-optimal networkperformance due to the increment of clustering overhead.Increasing the number of common channels in a clusterincreases bandwidth availability for intra-cluster communica-tions (Huang et al., 2011); while increasing the number ofcommon channels with neighboring clusters increases band-width availability for inter-cluster communications and con-nectivity with neighboring clusters (Ozger and Akan, 2013).Higher bandwidth availability is important to applications thatrequire constant message exchange such as channel sensingand routing.

B.3 Enhancement on Energy Efficiency. Nodes form clusters withthe objective of reducing energy consumption in order toprolong network lifetime and improve intra-cluster and inter-cluster connectivity. This can be achieved by minimizingtransmission power, intra-cluster distance, and the Euclideandistance between member nodes and their respective cluster-heads (Zhang et al., 2011). Additionally, since a clusterheaddepletes energy faster compared to member nodes because ofits role as a local point of process for cooperative tasks, such aschannel sensing, channel access and routing, the role ofclusterhead is rotated among nodes in a cluster in order toachieve load balancing and a well-balanced energy consump-tion among nodes throughout the entire network Xu et al.(2010).

B.4 Enhancement on Cooperative Tasks. In channel sensing, eachmember node senses for white spaces, and subsequentlysends the sensing outcomes to its clusterhead, which makesthe final decisions on the presence of PU activities. Hence, acluster, which is comprised of a clusterhead and its membernodes, provides a suitable model for cooperative tasks.Generally speaking, clustering enhances channel sensing para-meters (e.g. bandwidth availability and probability of falsealarm), as well as the associated performance metrics (e.g.energy consumption). The application of clustering to enhancenetwork performances of channel sensing has been found inWei and Zhang (2010).

B.5 Minimizing Number of Clusters. With reduced number ofclusters (or increased number of member nodes in a cluster),clustering provides efficient coverage and minimizes over-lapping among clusters, while maintaining the original

network connectivity of non-clustered networks. Lower num-ber of clusters (or higher number of member nodes in acluster) reduces inter-cluster communication overheads, par-ticularly routing messages, leaving more white spaces for datatransmission (Chen et al., 2007; Baddour et al., 2011). Lowernumber of clusters has been achieved in most clusteringalgorithms (Huang et al., 2011; Chen et al., 2007; Li andGross, 2011; Baddour et al., 2011; Liu et al., 2012; Zhang etal., 2011, 2010; Ramli and Grace, 2010).

2.2. Clustering characteristics

There are two types of clustering characteristics, namelyclustering metrics and intra-cluster distance. Clustering metricshave been applied to perform cluster formation and clustermaintenance, such as clusterhead selection and member nodejoining. There are four kinds of clustering metrics, namely, channelavailability, geographical location, signal strength and channelquality, as well as node degree. Intra-cluster distance defines thenumber of hops between member nodes and their respectiveclusterheads. Nodes may form clusters with single or multiplehops, and both larger and smaller clusters have their pros and cons(see Section 5.6 for further descriptions).

C.1 Clustering metrics:C.1.1 Channel availability. Higher number of common channels

in a cluster increases cluster stability, and so it minimizesthe occurrence of re-clustering, which is caused by thelack of a common channel. Due to the dynamicity ofchannel availability, PU activities may reappear, and thiscauses the member nodes to switch the common controlchannel; however, if there is lack of a single channelcommonly available to all nodes in a cluster, migration ofclusterhead, or re-clustering may be triggered and thisincreases the clustering overhead. Hence, nodes formclusters and select clusterheads while ensuring highernumber of common channels among nodes in a cluster.This metric has been applied in Ozger and Akan (2013),Huang et al. (2011), Chen et al. (2007), Li and Gross(2011), Baddour et al. (2011), Bradonjic and Lazos (2012),Li et al. (2012), Liu et al. (2012), Zhang et al. (2011, 2010),and Asterjadhi et al. (2010).

C.1.2 Geographical location. SUs determine their respectivegeographical locations (e.g. through Global PositioningSystem (GPS)), and use this information in clusterformation and maintenance. Physically close nodes mayshare the same amount and characteristics of whitespaces; and so these nodes form a cluster to increasethe number of common channels among nodes in acluster. Additionally, physically close nodes may coop-eratively perform similar tasks, such as channel sensing,and so these nodes form a cluster to enhance networkperformance of the application. In (Ozger and Akan,2013), shorter Euclidean distance between a clusterheadand a sink node is preferred in CRSNs. This metric hasbeen applied in Ozger and Akan (2013), Zhang et al.(2010, 2011) and Wei and Zhang (2010).

C.1.3 Signal strength and channel quality. SUs select theirrespective clusterheads based on Received SignalStrength Indicator (RSSI) (Ramli and Grace, 2010;Alsarhan and Agarwal, 2009). In Ramli and Grace(2010), SUs form clusters with reduced average distancebetween member nodes and their respective cluster-heads, as well as minimum level of overlap amongclusters. This helps to reduce the inter-cluster and

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 83

Page 6: Clustering algorithms for Cognitive Radio networks: A survey

intra-cluster contentions, as well as the number ofclusters in the network. In Ramli and Grace (2010),nodes with higher RSSI indicate higher node degreelevels, and so they are selected as clusterheads; whilemember nodes connect to a single clusterhead with thestrongest RSSI. Higher signal strength may also indicatehigher channel quality (Li et al., 2012). This metric hasbeen applied in Li et al. (2012) and Ramli and Grace(2010).

C.1.4 Node degree. Higher level of node degree, which repre-sents the number of neighbor nodes, reduces intra-cluster distance; and subsequently reduces the amountof overhead associated with intra-cluster communica-tions. This metric has been applied in Ozger and Akan(2013), Huang et al. (2011), Li and Gross (2011), Baddouret al. (2011) and Asterjadhi et al. (2010).

C.2 Intra-cluster distance:C.2.1 Single hop. SUs form single-hop clusters in which each

member node communicates with its clusterhead in asingle hop (Ozger and Akan, 2013; Chen et al., 2007; Liand Gross, 2011; Baddour et al., 2011; Bradonjic andLazos, 2012; Li et al., 2012; Liu et al., 2012; Zhang et al.,2011; Wei and Zhang, 2010; Ramli and Grace, 2010).Generally speaking, single-hop clusters enhance net-work stability, parallelism and inter-cluster communica-tion delays.

C.2.2 Multiple hops. SUs form multiple-hop clusters in whicheach member node communicates with its clusterheadin multiple hops (e.g. two hops) (Huang et al., 2011;Zhang et al., 2010; Asterjadhi et al., 2010). Generallyspeaking, multiple-hop clusters reduce the number ofclusters in the network and hence, it provides lowerinter-cluster communication overhead, such as routingmessages.

3. Clustering algorithms in cognitive radio networks

This section presents existing work on clustering algorithms inCR networks, and they are presented in accordance to the cluster-ing objectives (see Section 2.1). Generally speaking, the clusteringalgorithms aim to achieve lower number of clusters (see Section2.1, B.5) (Huang et al., 2011; Chen et al., 2007; Li and Gross, 2011;Baddour et al., 2011; Li et al., 2012; Liu et al., 2012; Zhang et al.,2011, 2010; Ramli and Grace, 2010; Asterjadhi et al., 2010), inaddition to the other clustering objectives.

3.1. Establishment of common control channel

While there has been some literature on the establishmentof common control channel in non-clustered networks (Kim,2009; Liu, 2010); there is lack of clustering algorithm that helpsto achieve this goal in clustered networks, and this sectionpresents a clustering algorithm whose main objective is to estab-lish a common control channel (see Section 2.1, B.1) in a cluster.

3.1.1. Li's node ranking approachLi et al. (2012) proposed a cluster formation procedure for static

(or quasi-stationary) networks with the objective of establishingcommon control channels in single-hop clusters (see Section 2.2,C.2.1). The clustering metrics are channel availability (see Section2.2, C.1.1) as well as signal strength and channel quality (seeSection 2.2, C.1.3). The steps of the clustering algorithm for a node iare shown in Fig. 4. Generally speaking, each node i ranks itsneighbor nodes nAN based on the link characteristics, namely,number of common channels, as well as the capacity and quality of

the link. Lower rank value is favorable, and so node i associatesitself with the lowest ranked neighbor node. Note that, forsimplicity, the nodal representations (or the labels of the nodes),which may indicate the roles of the nodes in Fig. 5, remain thesame before and after the clustering procedure although the rolesof clusterhead and member node are assigned upon completion ofthe procedure.

Firstly, nodes exchange the local cluster formation information,such as a list of their respective available channels, with neighbornodes. Subsequently, each node computes a metric called reservedvalue for each link connecting to each of its neighbor nodes usingthe cluster formation information. The reserved value is computedusing the rank values; for instance, node i estimates the capacity oflink (i,n) based on Shannon's formula, and ranks the capacity ofeach link (i,nAN) connecting to each neighbor node. Hence, if alink (i,n) has the lowest rank of capacity or C(i,n)¼1, it representsthat link (i,n) is the most favorable one with the highest capacity.The reserved value of the link of node i connecting to one of itsneighbor node nAN is the weighted sum of the rank values asfollows:

Rði;nANÞ ¼w1Cði;nÞþw2Sði;nÞþw3Q ði;nÞ ð1Þ

where w1, w2 and w3 represent weights (e.g. w1¼w2¼w3¼1/3).Value S(i,n) represents the rank of the stability level. Lower valueof S(i,n) indicates higher number of common channels in link (i,n),and so it provides greater stability, thereby reducing the occur-rence of re-clustering. Value Q(i,n) represents the rank of thechannel quality, which is estimated based on the available dura-tion of the link for data transmission. Lower value of Q(i,n)indicates higher probability of two nodes i and n belonging tothe same cluster. For instance, in Fig. 5(a), the reserved values ofthe links of node M1,2 connecting to its neighbor nodes CH1 andCH1 are R(M1,2,CH1)¼1 and R(M1,2,CH2)¼3, respectively. Note that,each link has two ends, and so it has two reserved values: R(i,n)and R(n,i); for instance, R(M1,2,CH1)¼1 and R(CH1,M1,2)¼2.

Secondly, the reserved values are exchanged among neighbornodes, and each of them adds the reserved values of each link toprovide an aggregated reserved value as follows:

R0ði;nÞ ¼ R0ðn; iÞ ¼ Rði;nÞþRðn; iÞ ð2Þ

For instance, in Fig. 5(b), R0ðM1;2;CH1Þ ¼ R0ðCH1;M1;2Þ ¼ RðM1;2;

CH1Þ þRðCH1;M1;2Þ ¼ 1þ2¼ 3. A node becomes a clusterhead if ithas either the highest node degree level or the lowest node IDamong neighboring nodes or both.

Fig. 4. Steps of the Li's node ranking approach at node i.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9584

Page 7: Clustering algorithms for Cognitive Radio networks: A survey

Thirdly, nodes form clusters, and this is accomplished throughnodes associating themselves with clusterheads that provide linkswith the least aggregated reserved value R0(i,n) (or R0(n,i)). Forinstance, in Fig. 5(c), M1,2 is associated with CH1 becauseR0ðM1;2;CH1ÞoR0ðM1;2;CH2Þ. In other words, CH1 ¼ argminnANR

0

ðM1;2;nÞ. A common control channel is selected out of the availablecommon channels in a cluster.

Li's node ranking approach has been shown to provide lowernumber of clusters (see Section 4.1, P.1), higher number ofcommon channels in a cluster (see Section 4.1, P.2), lower cluster-ing overhead (see Section 4.1, P.3), and lower number of re-clustering (see Section 4.1, P.5).

3.2. Enhancement on cluster stability

This section presents clustering algorithms whose main objec-tive is to enhance cluster stability (see Section 2.1, B.2).

3.2.1. Huang's node importance degree approachHuang et al. (2011) proposed a cluster formation procedure

to form multiple-hop clusters (see Section 2.2, C.2.2) in mobilenetworks. The clustering metrics are channel availability (seeSection 2.2, C.1.1) and node degree level (see Section 2.2, C.1.4).There are two metrics pertinent to cluster stability at link andcluster levels.

Firstly, the probability of link availability, which is computedusing the nodal mobility characteristics (i.e. speed and direction),must be greater than a certain threshold within a long-enoughperiod of time T. Higher probability of link availability increasesthe link stability of each pair of nodes in a cluster.

Secondly, using channel m, each node i calculates a metriccalled node importance degree as follows:

Di;m ¼ ni;m=ð1þhi;mþsi;mÞ ð3Þwhere ni,m represents the number of one-hop and two-hopneighbor nodes on channel m; hi,m represents the average numberof hops (or intra-cluster distance), which affects the intra-clusterdelay, from node i to all one-hop and two-hop neighbor nodes;and si,m represents the average number of channel switches due tothe distinctive channels being selected by node i and its neighbornodes. Nodes with the highest node importance degree Di,m

among their respective two-hop neighbor nodes are selected as

clusterheads; while the rest of the nodes associate themselveswith clusterheads that provide the highest Di,m.

The clusterhead chooses a common control channel m; how-ever, this channel may not be available to some of its membernodes. Such cases may be reduced by minimizing the averagenumber of channel switches si,m; however, if this happens, anupstream member node must switch between channel m and acommon channel between the two nodes. For instance, in Fig. 6,clusterhead CH2 selects a common control channel m¼3; whilechannel availability at member node M2,4 is KM2;4 ¼ f3;4g and M2,5

is KM2;5 ¼ f4;5g. Since KM2;5 excludes the common control channelof the cluster, and the common channel between member nodesM2,4 and M2,5 is channel 4, member node M2,4 must constantlyswitch between channels 3 and 4.

Huang's node importance degree approach has been shownto provide lower number of clusters (see Section 4.1, P.1), lowerclustering overhead (see Section 4.1, P.3), and higher number ofcommon channels with neighboring clusters (see Section 4.1, P.9),which improves connectivity among clusters.

3.2.2. Baddour's affinity propagation message-passing approachBaddour et al. (2011) proposed a cluster formation procedure to

form single-hop clusters (see Section 2.2, C.2.1) in mobile net-works using the affinity propagation message-passing approach.The clustering metrics are channel availability (see Section 2.2,C.1.1) and node degree level (see Section 2.2, C.1.4). The steps of theclustering algorithm for a node i are shown in Fig. 7.

Firstly, nodes exchange the local cluster formation information,including a list of their respective available channels for eachneighbor node and node degree level, with neighbor nodes. Theaffinity propagation approach applies a metric that represents thesimilarity level of a node pair (i.e. nodes i and n), which indicatesthe number of common channels among node i and nAN.Specifically, the similarity between nodes i and n is sði;nÞ ¼sðn; iÞ ¼ jKi \ Knj, where Ki and Kn are the sets of available channelsat nodes i and n, respectively.

Secondly, nodes compute and further exchange the local clusterformation information including responsibilities and availabilities.Node i sends responsibility r(i,n) to node n, specifically

rði;nÞ’sði;nÞ�maxn0 anfaði;n0Þþsði;n0Þg ð4Þ

Fig. 5. The Li's node ranking approach. (a) Nodes calculate reserved values. (b) Nodes calculate aggregated reserved values. (c) Nodes form clusters.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 85

Page 8: Clustering algorithms for Cognitive Radio networks: A survey

and node n sends availability a(i,n) to node i, specifically

aði;nÞ’min 0; rðn;nÞþ ∑i0 =2fi;ng

max f0; rði0;nÞg( )

ð5Þ

Since s(i,n) represents the similarity level between nodes i and n,both responsibility r(i,n) and availability a(i,n) indicate how well-suited node n is to serve as clusterhead for node i from theperspective of nodes i and n, respectively.

Thirdly, nodes that fulfill r(i,i)þa(i,i)40 become potentialclusterheads. Among the potential clusterheads, nodes with thehighest node degree level among its single-hop neighborhoodbecome clusterheads.

Baddour's affinity propagation message-passing approach hasbeen shown to provide lower number of clusters (see Section 4.1,P.1).

3.2.3. Connectivity degree approachLi and Gross (2011) proposed a cluster formation procedure

to form single-hop clusters (see Section 2.2, C.2.1) in order toincrease the number of common channels among nodes in acluster in static (or quasi-stationary) networks. The objective isto prevent loss of connectivity among clusterheads, as well as theclusterheads and their respective member nodes. The clustering

metrics are channel availability (see Section 2.2, C.1.1) and nodedegree level (see Section 2.2, C.1.4). The steps of the clusteringalgorithm for a node i are shown in Fig. 8.

Firstly, nodes exchange the local cluster formation information,such as a list of their respective available channels for eachneighbor node, with neighbor nodes. Each node i keeps track ofconnectivity vector comprises two-tuple metrics ⟨Pi,Gi⟩ pertinentto cluster stability, where Pi represents spectrum connectivitydegree and Gi represents local connectivity degree. The spectrumconnectivity degree Pi is the total number of common channelsbetween node i and each of its neighbor nodes, and so it indicatesthe adhesiveness of node i to the network. For instance, node i hastwo neighbor nodes j and k, and the set of the available channels ateach node is Ki¼{1,2,3,4,5}, Kj¼{1,2,3} and Kk¼{2,4}; the spectrumconnectivity degree is computed as Pi¼Pi,jþPi,k¼3þ2¼5. Thelocal connectivity degree Gi is the number of common channelsamong node i and all of its neighbor nodes, and it indicates thesuitability of node i to form a robust cluster with its neighboringnodes. For instance, using the aforementioned example, Gi¼1because there is a single common channel among nodes i, j andk, specifically, channel 2. Figure 9(a) shows an original non-clustered network. The two-tuple metrics ⟨Pi,Gi⟩ are shown at eachnode, and the number of pairwise common channels for each nodepair Pi,nAN is shown at each link.

Secondly, nodes compute and further exchange the local clusterformation information, including spectrum connectivity degree Piand local connectivity degree Gi. Node i becomes a clusterhead ifnode I's spectrum connectivity degree is less than those of itsneighbor nodes, specifically, PioPnAN\CH, where N represents nodei's set of neighbor nodes and CH represents node i's set ofclusterheads among its neighbor nodes. The purpose of choosingnodes with lower spectrum connectivity degree as clusterheads isthat, it increases the number of common channels with neighbor-ing clusters, and so it increases connectivity among clusters.However, if node i's spectrum connectivity degree is similar toany of its neighboring nodes, specifically, Pi¼PnAN\CH, then thenode with higher local connectivity degree value, or Gi4GnAN\CH,becomes clusterhead. Non-clusterheads associate themselves withclusterheads and become member nodes. Note that, smallercluster size increases the number of common channels among

Fig. 6. The Huang's node importance degree approach. Node M2,4 must switchbetween channels 3 and 4 in order to communicate with CH2 and M2,5. The set ofavailable channels is shown at each node.

Fig. 7. Steps of the Baddour's affinity propagation message-passing approach atnode i.

Fig. 8. Steps of the Li and Gross's connectivity degree approach at node i.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9586

Page 9: Clustering algorithms for Cognitive Radio networks: A survey

nodes in a cluster, and so nodes may be eliminated from theirrespective clusters until there is at least a single common channelin the cluster. Figure 9(b) shows the updated values of two-tuplemetrics ⟨Pi,Gi⟩ and the newly formed clusters. For instance, sincePCH1 ¼ 2oPM1;1 ¼ 11, node CH1 becomes a clusterhead.

Thirdly, nodes that are physically located in more than a singlecluster, which are also called debatable nodes, choose to associatethemselves with a single cluster respectively. The debatable nodeschoose to associate themselves with clusterheads that can max-imize the number of common channels in a cluster in order toincrease cluster stability. For instance, in Fig. 9(c), node M1,1 is adebatable node, which may associate itself with either CH1, CH2,CH3 or CH5, and it chooses to associate itself with CH1.

Li and Gross's connectivity degree approach has been shown toprovide lower number of clusters (see Section 4.1, P.1), highernumber of common channels in a cluster (see Section 4.1, P.2), andhigher number of common channels with neighboring clusters(see Section 4.1, P.9).

A similar clustering algorithm that applies the local connectiv-ity degree parameter has been proposed in Asterjadhi et al. (2010)for cluster formation to form multiple-hop clusters (see Section2.2, C.2.2) in order to increase the number of common channelsamong nodes in a cluster in static networks. The clustering metricsare channel availability (see Section 2.2, C.1.1) and node degreelevel (see Section 2.2, C.1.4). Nodes exchange the local clusterformation information, including a list of the available channels foreach neighbor node nAN (see Sections 2.2, C.1.1) and node degreelevel (see Section 2.2, C.1.4) with k-hop neighbor nodes where N isa set of node i's neighbor nodes. Node i calculates the localconnectivity degree Gi, which is the minimum number of commonchannels among node i and all of its neighbor nodes, orGi ¼minnAN jKi \ Knj, where Ki is a set of available channels atnode i. Nodes with the largest f(Gi) value, which is computed usingGi, among their respective k-hop neighborhood become cluster-heads. Suppose, node i is a clusterhead. Neighbor node nAN thatreceives higher f(Gi), specifically f(Gn)o f(Gi), associates itself withcluster i and becomes its member node. Asterjadhi's connectivitydegree approach has been shown to provide lower number of

clusters (see Section 4.1, P.1), and higher number of commonchannels in a cluster (see Section 4.1, P.2).

3.2.4. Bipartite graphs approachBradonjic and Lazos (2012) proposed a cluster formation

procedure to form single-hop clusters (see Section 2.2, C.2.1) instatic networks using bipartite graphs. The clustering metric ischannel availability (see Section 2.2, C.1.1). The steps of theclustering algorithm for a node i are shown in Fig. 10.

Firstly, nodes exchange the local cluster formation information,such as a list of their respective available channels, with neighbornodes. Each node i represents its local network topology, whichconsists of neighboring nodes and the available channels for eachneighbor node, using a bipartite graph. The bipartite graph GðA [B; EÞ comprises two disjoint sets of vertices A [ B, and all edges inE connect vertices from A to B. For instance, Fig. 11(b) shows abipartite graph constructed by node CH1 in (a), where vertices A(or the upper row) represent node CH1 and its neighboring nodes,vertices B (or the lower row) represent the available channels at

Fig. 9. The Li and Gross's connectivity degree approach. (a) Nodes calculate two-tuple metrics ⟨Pi,Gi⟩. (b) Nodes form clusters. (c) Debatable nodes associate themselves withclusterheads.

Fig. 10. Steps of the bipartite graphs approach at node i.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 87

Page 10: Clustering algorithms for Cognitive Radio networks: A survey

node CH1, and each edge eAE indicates that node aAA has anavailable channel bAB. The objective of the approach is to achievea balanced tradeoff between cluster size and the number ofcommon channels among nodes in a cluster (see Section 5.3).Higher number of common channels among nodes in a clusterhelps to increase the amount of bandwidth availability for intra-cluster communications and reduce the occurrence of re-cluster-ing; however, the cluster size tends to be smaller and so the maindrawback is that there would be higher amount of bandwidthconsumption on inter-cluster communications. Three clusteringcriteria are proposed to achieve different levels of the aforemen-tioned tradeoff as follows:

� Maximum Node Biclique (MNB). MNB maximizes Qn

MNBðA;BÞ,which is the sum of the number of member nodes |A| and thenumber of common channels in the cluster |B|, as follows:

Qn

MNBðA;BÞ ¼ argmaxQ jAjþjBj ð6Þ

Higher Qn

MNBðA;BÞ indicates that member nodes share a highernumber of common channels in a cluster, or there is at least asingle common channel in a cluster with higher number ofmember nodes in a cluster. For instance, Fig. 11(c) shows thebipartite graph constructed by node CH1 using MNB in whichQn

MNBðA;BÞ ¼ 7þ1¼ 8.� Maximum Edge Biclique (MEB). MEB maximizes Qn

MEBðA;BÞ,which is the product of the number of member nodes |A| andthe number of common channels in the cluster |B|, as follows:

Qn

MEBðA;BÞ ¼ argmaxQ jAjU jBj ð7Þ

The nature of product in MEB exhibits a higher sensitivity tochanges in |A| and |B|; hence, MEB does not construct clusters withsmaller cluster size and lower number of common channels. Forinstance, Fig. 11(d) shows the bipartite graph constructed by nodeCH1 using MEB in which Qn

MEBðA;BÞ ¼ 4� 3¼ 12. Note that,bipartite graph in Fig. 11(d) provides higher Qn

MEBðA;BÞ ¼ 12 whenusing MEB, compared to the bipartite graph in Fig. 11(c) in whichQn

MEBðA;BÞ ¼ 7� 1¼ 7.� Maximum One-Sided Edge Cardinality Biclique (MECB). MECB

maximizes Qn

MECBðA;BÞ in order to form clusters with the high-est possible number of member nodes |A| under a constraint on

the number of common channels in a cluster |B| as follows:

Qn

MECBðA;BÞ ¼ argmaxQ jAj with jBj4BT ð8Þwhere BT is a threshold. This means that MECB maximizes thecluster size while providing a guarantee on the minimumamount of bandwidth availability in a cluster using thresholdBT. Hence, the bipartite graph is equivalent to Fig. 11(c) and(d) given BT¼1 and BT¼3, respectively. Note that, cluster sizebecomes smaller when the number of common channelsamong nodes in a cluster becomes higher.

Using one of the three proposed clustering algorithms (i.e.MNB, MEB and MECB), each node i segregates the bipartite graph(network) into biclique graphs (clusters). Each node i computeswi¼Q*(A,B).

Secondly, nodes exchange wi among themselves, and nodeswith the maximum wi among its neighbor nodes becomeclusterheads.

Thirdly, non-clusterhead nodes associate themselves with clus-terheads. Each non-clusterhead node selects clusterhead with themaximum wi among its single-hop neighbor clusterheads andbecomes its member node.

As part of the cluster maintenance, re-clustering may benecessary when the number of common channels in a clusterwith node i being clusterhead is less than a threshold, or |Bi|oBT.In this case, all nodes in the cluster switch into “undecided” state,and a new clusterhead j with |Bj|4BT is selected.

Bradonjić's bipartite graphs approach has been shown toprovide lower number of clusters (see Section 4.1, P.1), and highernumber of common channels in a cluster (see Sections 4.1, P.2).The investigation on achieving a balanced tradeoff between clustersize and the number of common channels among nodes in acluster is also presented. Similar approach has been applied in Liuet al. (2012) and Nafees et al. (2013).

3.3. Enhancement on energy efficiency

This section presents clustering algorithms whose main objec-tive is to enhance energy efficiency.

Fig. 11. Bipartite graphs approach. (a) Original network topology. (b) A bipartite graph constructed by node CH1. (c) A bipartite graph constructed by node CH1 using MNB.(d) A bipartite graph constructed by node CH1 using MEB.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9588

Page 11: Clustering algorithms for Cognitive Radio networks: A survey

3.3.1. Zhang's group-wise constrained approachZhang et al. (2011) proposed a cluster formation procedure to

form single-hop clusters (see Section 2.2, C.2.1) in static networksusing the group-wise constrained approach. The clustering metricsare channel availability (see Section 2.2, C.1.1) and geographicallocation (see Section 2.2, C.1.2). There are two factors pertinent tothe reduction of energy consumption.

Firstly, the transmission power can be reduced by minimizingthe distance between member nodes and their respective cluster-heads (or intra-cluster distance). Additionally, SUs' interference toPUs can be reduced with lower transmission power.

Secondly, the transmission power can be reduced by achievingthe optimal cluster size (or the number of clusters in the network),which affects intra-cluster and inter-cluster communications, andhence the transmission power. During intra-cluster communica-tions, member nodes exchange information among themselvesand their respective clusterheads; for instance, member nodesexchange the local cluster formation information (i.e. node loca-tion, the current cluster size and a list of common channels in thecluster) with neighbor nodes and clusterhead. During inter-clustercommunications, clusterheads send data to upstream clusterheadsusing the maximum transmission power in order to improvenetwork connectivity.

There are two main steps in the group-wise constrainedapproach for cluster formation. Firstly, all nodes form disjointclusters, and each cluster is comprised of a single node itself.Secondly, the clusters merge among themselves with adjacentclusters if they share at least a single common channel; and thisproceeds until the number of clusters reduces to the optimalnumber, which is dependent on the number of nodes in thenetwork, node density and the maximum transmission range.In other words, physically closest nodes with common channel(s) form clusters that grow in cluster size as more nodes join aparticular cluster one after another through cluster merging.

Since the clusterheads incur higher energy consumption com-pared to member nodes, the clusterhead role is rotated among allnodes with equal probability. Zhang's group-wise constrainedapproach has been shown to provide lower number of clusters(see Section 4.1, P.1), and lower energy consumption (see Section4.1, P.8).

3.3.2. Ozger's event-driven approachOzger and Akan (2013) proposed Event-driven Spectrum-

Aware Clustering (ESAC) protocol for cluster formation in staticCRSNs with the main objective of minimizing energy consumptionin single-hop clusters (see Section 2.2, C.2.1). The clusteringmetrics are channel availability (see Section 2.2, C.1.1) as well asgeographical location (see Section 2.2, C.1.2) and node degree (seeSection 2.2, C.1.4).

There are three strategies to minimize energy consumption.Firstly, ESAC is an event-driven approach that forms temporaryclusters that is initiated upon detection of an event, and it ismaintained until the termination of the event. This avoids energyconsumption required to form and maintain clusters at all times.Secondly, ESAC forms clusters in regions between event(s) andsink, rather than in the entire network. Thirdly, ESAC preventsre-clustering by maximizing the number of common channels in acluster, as well as with two-hop nodes, which are from neighbor-ing clusters. The rest of this section presents how the second andthird strategies are achieved. There are two main phases, namelyidentifying eligible nodes for clustering, and forming clusters.

In the first phase, a node that detects an event becomes aneligible node for clustering. The node chooses neighbor nodes thatare physically closer to the sink and farther from the eventto become eligible nodes as well. Next, the new eligible nodes

determine their next-hop eligible nodes, and this is performeduntil the sink node is found. The outcome is that, nodes in theregion between event(s) and sink node become eligible forclustering, and this achieves the aforementioned second strategy.

In the second phase, there are two main steps, namely cluster-head selection and cluster formation. The steps of the clusteringalgorithm for a node i are shown in Fig. 12. During clusterheadselection, an eligible node i with the maximum weightws;i ¼ jKij � Diþ10=di;sink among its single-hop neighborhood ischosen as clusterhead, where |Ki| is the number of availablechannels at node i, Di is the node degree level involving single-hop eligible nodes at node i, and di,sink is the Euclidean distancebetween node i and the sink node. Higher weight value indicatesthat a node has higher number of available channels, node degreelevel involving eligible nodes, and closer to the sink node amongits single-hop neighborhood nodes. During cluster formation, thepurpose is to discover a candidate (or potential) cluster with themaximum weight wf,i,j, and this requires |J| iterations. For eachiteration jA J¼{1,…, |K|} where Ki represents a set of availablechannels at clusterhead node i, clusterhead i creates a candidatecluster using a particular available channel kjAKi, and subse-quently, it calculates weight wf ;i;j ¼ jKjj � jni;1;jj � jni;2;jj for thecluster. Next, the sets Kj, ni,1,j and ni,2,j are updated. At eachiteration j, the respective channel kj is included in the set Kj,specifically jKjj [ kj. The ni,1,j represents the set of single-hopeligible nodes of node i using channel kj. The ni;2;j represents theset of two-hop eligible nodes in set ni,1,j using channels in set Kj. Inother words, by using the metric ni,2,j, a clusterhead considers thenumber of two-hop neighbors that are accessible by its one-hopneighbors using the cluster channels. This helps to increaseconnectivity among clusters. Higher weight wf,i,j values indicate acandidate cluster with higher number of common channels forintra-cluster and inter-cluster communications. Hence, thisachieves the aforementioned third strategy.

Fig. 12. Steps of the Ozger's ESAC protocol at node i.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 89

Page 12: Clustering algorithms for Cognitive Radio networks: A survey

Table 1Comparison of clustering algorithms in CR networks.

References Clustering objective Clustering attributes

Clustering metric Intra-cluster distance

B.1 Establishment of common controlchannel

B.2 Enhancement oncluster stability

B.3 Enhancement onenergy efficiency

B.4 Enhancement oncooperative tasks

B.5 Minimizingnumber of clusters

C.1.1Channelavailability

C.1.2Geographicallocation

C.1.3 Signal strengthand channel quality

C.1.4 Nodedegree

C.2.1Single hop

C.2.2Multiplehops

Li et al. (2012) � � � � �Huang et al. (2011) � � � � �Baddour et al. (2011) � � � � �Li and Gross (2011) � � � � �Ramli and Grace (2010) � � �Asterjadhi et al. (2010) � � � � �Bradonjic and Lazos (2012) � � �Liu et al. (2012) � � � �Zhang et al. (2011) � � � � �Ozger and Akan (2013) � � � � � �Wei and Zhang (2010) � � �Chen et al. (2007) � � �Zhang et al. (2010) � � � � �

References Performance enhancements compared to existing approaches Simulation tool

P.1 Lowernumber ofclusters

P.2 Highernumberof commonchannels in acluster

P.3 Lowerclusteringoverhead

P.4 Lowernumber ofunconnectednodes

P.5 Lowernumber ofre-clustering

P.6 Shorterintra-clusterdistance

P.7 Lowerlevel ofoverlapamongclusters

P.8 Lowerenergyconsump-tion

P.9 Higher number ofcommon channels withneighboring clusters

P.10 Lower errorprobability insensing outcomes

S.1 Self-developedtool

S.2NS-2 orNS-3

S.3OMNeTþþ

S.4Qualnet

Li et al. (2012) � � � � �Huang et al.(2011)

� � � �

Baddour et al.(2011)

� �

Li and Gross(2011)

� � � �

Ramli andGrace(2010)

� � � � �

Asterjadhiet al. (2010)

� � �

Bradonjic andLazos(2012)

� � �

Liu et al.(2012)

� � � � �

Zhang et al.(2011)

� � �

Ozger andAkan (2013)

� � �

Wei andZhang(2010)

� � �

Chen et al.(2007)

� �

Zhang et al.(2010)

� � �

K.-L.A

.Yauet

al./Journal

ofNetw

orkand

Computer

Applications

45(2014)

79–95

90

Page 13: Clustering algorithms for Cognitive Radio networks: A survey

Finally, clusterhead i selects the candidate cluster jA J with themaximum weight wf,i,j, which was created in one of the |J|iterations. Upon selecting a candidate cluster, the clusterhead isends request message to inform its single-hop neighborhood sothat nodes can join this cluster as member nodes.

Ozger's event-driven approach has been shown to providelower energy consumption (see Section 4.1, P.8), and highernumber of neighboring nodes with common channels (seeSection 4.1, P.9).

3.4. Enhancement on cooperative tasks

This section presents a clustering algorithm whose mainobjective is to enhance the performance of an application scheme,namely channel sensing. Channel sensing aims to detect whitespaces and the presence of PU activities. For instance, in Wei andZhang (2010), a clustering algorithm is proposed for channelsensing with the objective of reducing the error probability insensing outcomes.

3.4.1. Wei's contribution decision approachWei and Zhang (2010) proposed a cluster formation procedure

to form single-hop clusters (see Sections 2.2, C.2.1) in staticnetworks using the contribution decision approach. The clusteringmetric is geographical location (see Section 2.2, C.1.2), and it is alsobased on the accuracy of the sensing outcomes. A clusterednetwork is comprised of member nodes, clusterheads and a fusioncenter. Each cluster comprises member nodes and a clusterhead.Among the clusterheads, one is selected to serve as a fusion center.To obtain accurate decisions on sensing outcomes, clusterheadsmake local decisions based on sensing outcomes collected fromtheir respective member nodes; and subsequently, the fusioncenter makes global decisions based on local decisions collectedfrom clusterheads. There are two main steps in cluster formation.

Firstly, nodes form clusters based on their respective physicallocations, and so nodes that are physically close to each other areclustered.

Secondly, each SU keeps track of its own contribution value,which is calculated based on the consistency (or similarity) ofsensing outcomes compared to final decisions of the sensingoutcomes, which are made by the fusion center. The contributionvalue is positive if a member node's sensing outcome is consistentwith the final decision, and vice-versa. The contribution valueis further evaluated using a pre-defined threshold. Specifically, amember node sends its channel sensing outcomes to its cluster-head if its contribution value is greater than the pre-definedthreshold, otherwise it keeps silence due to its unreliability. Anode with the highest contribution value among its single-hopneighborhood is selected as the clusterhead. Since the cluster-heads incur higher energy consumption compared to membernodes, the clusterhead role is rotated among nodes, and this hasbeen shown to reduce energy consumption.

Wei's contribution decision approach has been shown toprovide lower energy consumption (see Section 4.1, P.8), andlower error probability in sensing outcomes (see Section 4.1,P.10). The algorithm achieves lower energy consumption becausemember nodes with contribution value lower than a pre-definedthreshold do not send their sensing outcomes to their respectiveclusterheads. Note that, achieving lower error probability insensing outcomes (see Section 4.1, P.10) requires higher numberof local sensing outcomes from more nodes, and so this incurshigher energy consumption. Hence, further research could bepursued to investigate approaches to achieve a balanced tradeoffbetween these two network performances.

4. Performance enhancements and complexity analysis ofclustering algorithms

This section presents performance enhancements and com-plexity analysis of various clustering algorithms.

4.1. Performance enhancements

Table 1 compares various aspects of clustering algorithms,including clustering objectives, clustering attributes and perfor-mance enhancements in comparison with conventional and tradi-tional approaches in CR networks. The comparison presents theadvantages brought about by various clustering algorithms; forinstance, single-hop clusters (see Section 2.2, C.2.1) enhancesstability, parallelism and inter-cluster communication delays whilemultiple-hop clusters (see Section 2.2, C.2.2) reduces the numberof clusters in the network. On the other hand, open issues (seeSection 5) presents the shortcomings of existing clustering algo-rithms and proposes mechanisms to address them. Additionally,Table 1 also presents simulation tools applied to investigate theclustering algorithms.

Simulation has been adopted to evaluate network performanceachieved by the clustering algorithms. There are various simula-tion tools including self-developed tool (S.1), and network simu-lators such as NS-2 or NS-3 (Network Simulator 3 (NS-3), 2013)(S.2), OMNeTþþ (OMNeTþþ , 2013) (S.3) and Qualnet (Qualnet,2013) (S.4). Generally speaking, the self-developed tool is devel-oped by the authors of the literature themselves using program-ming languages, such as C/Cþþ in Li and Gross (2011), and thetool may run on computer applications, such as MatLab in Ramliand Grace (2010). Most simulations in the literature use self-developed tool (S.1) (see Table 1).

The performance enhancements may be relevant to each other,and they are shown as follows:

P.1 Lower number of clusters provides efficient coverage byincreasing cluster size or the number of member nodes in acluster. Lower number of clusters reduces the high amount ofinter-cluster communications, particularly routing, and so itleaves more white spaces for control and data transmissions(Chen et al., 2007; Baddour et al., 2011). Additionally, itminimizes overlaps among clusters (see Section 4.1, P.7),and so it reduces channel contention among clusters.

P.2 Higher number of common channels in a cluster increases theavailability of at least a single common channel in a cluster,while leaving some common channels as backups. Hence,higher number of common channels in a cluster preventsre-clustering (see Section 4.1, P.5) as a result of there-appearance of PU activities, and subsequently improvescluster stability.

P.3 Lower clustering overhead reduces unnecessary exchanges ofclustering overhead. This may reduce energy consumption(see Section 4.1, P.8) and increase the amount of white spacesfor control and data transmissions. For instance, in Bradonjicand Lazos (2012), nodes broadcast clustering messages totheir respective single-hop neighborhood, rather thanmultiple-hop neighborhood in order to reduce clusteringoverhead.

P.4 Lower number of unconnected nodes reduces the number ofmember nodes that do not share any common channels in itscluster. This improves connectivity among member nodes in acluster, and subsequently reduces the number of channelswitches due to distinctive channels being selected by mem-ber nodes in a cluster. Note that, if there must be at least asingle common channel in a cluster, then either rotation ofclusterhead or re-clustering is necessary to achieve this.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 91

Page 14: Clustering algorithms for Cognitive Radio networks: A survey

P.5 Lower number of re-clustering reduces clustering overhead(see Section 4.1, P.3) and energy consumption (see Section 4.1,P.8) associated with re-clustering, and so it increases networklifetime. Re-clustering is necessary whenever a commonchannel cannot be found in a cluster, and it may be avoidedby the rotation of clusterhead.

P.6 Shorter intra-cluster distance reduces the number of hops orthe Euclidean distance (Ramli and Grace, 2010) betweenmember nodes and their respective clusterheads in a cluster.This helps to reduce the intra-cluster delay and energyconsumption due to lower transmission power.

P.7 Lower level of overlap among clusters reduces channel conten-tion among nodes in different clusters, and so it providesnetwork performance enhancement such as throughput andend-to-end delay.

P.8 Lower energy consumption increases network lifetime. In Weiand Zhang (2010), a threshold is chosen against the reliabilityof information in which member nodes do not send unreli-able sensing outcomes to their respective clusterheads inchannel sensing in order to reduce energy consumption.Higher threshold values reduce energy consumption.

P.9 Higher number of common channels with neighboring clustersindicates higher probability of connectivity among clusters.This prevents loss of connectivity with neighboring clusters,and reduces the number of channel switches due to distinc-tive channels being selected by a cluster and its neighboringclusters.

P.10 Lower error probability in sensing outcomes helps to improveaccuracy in the detection of PU activities and white spaces ina channel sensing scheme, which is one of the main functionsof CR networks. Note that, the enhancement of networkperformance in other kinds of applications is possible,although channel sensing has seemed to fit in well withclustered networks.

4.2. Complexity analysis

This section investigates the clustering algorithms with respect totime and message complexities associated with cluster formationstarting from initial networks comprised of non-clustered nodes. Thecomplexity analysis conducted in this section is inspired by similarinvestigation in Bettstetter and Konig (2013). This analysis is highlyrelevant to clustering, and so the exchanges of clustering information(e.g. a set of available channels) using Hello messages are not countedas these messages are incurred even in non-clustered networks. Thetime complexity is the number of time steps to form clusters in anetwork; and each time step involves a packet transmission from eachnode. The message complexity is the number of clustering messagesexchanged among non-clustered nodes to form clusters in a network.Note that, Wei and Zhang' (2010) contribution decision approach isnot analyzed as it is associated with a cooperative task inwhich nodeswith the highest channel sensing capability are selected to becomeclusterheads which can be rotated; while Zhang et al.'s (2010)approach is also not analyzed as it incorporates neighbor discoverymechanism in clusterhead selection. Denote the number of nodes in aCR networks by M. Table 2 compares the time and message complex-ities of the clustering algorithms.

The time and message complexities of several clustering algo-rithms are explained. In the Li et al.'s (2012) node ranking approach(see Section 3.1.1), each node broadcasts a reserved value, and itschoice of being a clusterhead or a member node, so time complex-ity is at most 2 time steps. The message complexity is at most 2Min which each node in a network, which comprises M nodes,exchanges at most 2 messages. In Huang et al.'s (2011) nodeimportance degree approach (see Section 3.2.1), 2-hop clusters

are formed. Each node calculates its own node importance degreeusing information which can be exchanged using Hello messages,and broadcasts its choice of being a clusterhead or a member nodewithin its two-hop neighborhood. Since a node may be two hopsaway from a clusterhead, it takes at most 2 time steps for aclusterhead to announce its role, as well as member nodes toannounce their respective choice of clusterheads, so time complex-ity is at most 2 time steps. The message complexity is atmost 2M messages. In Baddour et al.'s 2011 affinity propagationmessage-passing approach (see Section 3.2.2), each node broad-casts responsibilities, availabilities, and its choice of being aclusterhead or a member node, so time complexity is at most3 time steps. The message complexity is at most 3M. In Asterjadhiet al.'s (2010) approach (see Section 3.2.3), k-hop clusters areformed. Each node calculates its own local connectivity degreeand broadcasts this information, as well as its choice of being aclusterhead or a member node, within its k-hop neighborhood; anda non-clustered node takes at most k time steps to inform aclusterhead, which is k hops away to join the cluster. Hence, timecomplexity is at most 2k time steps, and message complexity is atmost kM messages.

In short, the time and message complexities increase with thenumber of hops between member nodes and their clusterhead in acluster (or cluster size), and the amount of clustering messagesassociated with a clustering algorithm.

5. Open issues

This section discusses open issues that can be pursued in thisresearch area in order to address the shortcomings of existingclustering algorithms.

5.1. Cluster maintenance: migration of clusterhead, cluster merging,cluster splitting, node joining and node leaving

The objective of cluster maintenance is to provide smoothprocedures of the migration of clusterhead in a cluster, clustermerging, cluster splitting, node joining and node leaving. Migra-tion of clusterhead is initiated to change the clusterhead node in acluster so that a more suitable member node is selected to betterserve as a clusterhead for the cluster. Cluster merging combinesadjacent clusters; while cluster splitting separates a single clusterinto more than ones. Both cluster merging and splitting require re-clustering, and the clusterhead should be maintained, re-electedor withdrawn to ensure that there is only a single clusterhead in acluster. Node joining and leaving affects the traffic amount. In CRnetworks, there has been limited literature on cluster maintenance,and the motivation of this investigation stems from the fact that thedynamicity of channel availability caused by PUs has introduced new

Table 2Comparison of time and message complexities among clustering algorithms.

References Time complexity Message complexity

Li et al. (2012) r2 r2MHuang et al. (2011) r2 r2MBaddour et al. (2011) r3 r3MLi and Gross (2011) r2 r2MRamli and Grace (2010) r2 r2MAsterjadhi et al. (2010) r2k rkMBradonjic and Lazos (2012) r2 r2MLiu et al. (2012) r2 r2MZhang et al. (2011) r2 r2MOzger and Akan (2013) r2 r2MChen et al. (2007) r2 r2MZhang et al. (2010) r2 r2M

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9592

Page 15: Clustering algorithms for Cognitive Radio networks: A survey

open issues, leading to the dynamic amount of available bandwidth atclusterheads and member nodes. Additionally, there are traditionalfactors that affect the amount of available bandwidth at clusterheads,such as the traffic amount and the effects of nodal mobility on theinterference to clusterheads. Due to the important role of clusterheadsand gateways as backbone nodes, channel availability at the cluster-heads should be well managed because insufficient amount ofbandwidth may cause packet loss and this affects Quality of Service(QoS) of the entire cluster. In the context of CR, cluster maintenanceaffects the number of common channels and the traffic amount in anexisting or a new cluster. Therefore, cluster maintenance must takeinto consideration the number of common channels in a cluster (andother performance metrics of interest) to ensure successful andsmooth cluster maintenance in order to minimize the occurrence ofre-clustering. In Sharma and Abrol (2013), an initial study on themigration of clusterhead in a cluster using thresholds based on missdetection, false alarm and the traditional signal-to-noise ratio ispresented for channel sensing. Further research could be pursued toinvestigate cluster maintenance while providing optimal networkperformance.

5.2. Common assumptions in cognitive radio networks

Future research could be pursued to relax the followingcommon assumptions applicable to the investigation of clusteringalgorithms in CR networks:

The links are considered bi-directional (or symmetric) (Zou andChigan, 2009; Baddour et al., 2011; Gong et al., 2008). However, inCR networks, links may be asymmetric because each node mayhave different sets of available channels. For instance, node i maycommunicate with node j using channel k; while node j can onlycommunicate with node i using channel m.

� Network dynamics, such as nodal mobility, channel availability andtraffic amount, change at a relatively slow rate or even staticthroughout the duration of cluster formation (Chen et al., 2007;Baddour et al., 2011). This means that most clustering algorithmsassume near-static scenarios so that clusters are formed usingstatic and the latest information. However, in real scenarios,clusters are formed in the presence of network dynamics.

� The common channels in a cluster are homogeneous in terms ofchannel quality, and this can be seen in the widely use of thenumber of common channels in a cluster (see Section 4.1, P.2) as aperformance metric. However, in real scenarios, clustering algo-rithms should consider the channel quality of the commonchannels.

� Performance metrics such as the number of clusters, thenumber of common channels in a cluster, and the number ofre-clustering (see Section 4.1), are commonplace to measurethe performance of clustering algorithms, rather than theoverall network-wide QoS performance, including throughput,end-to-end delay, jitter and packet loss rate. Hence, theinvestigation on achieving satisfactory QoS performance invarious clustering algorithms is necessary.

5.3. Tradeoff between various network performance metrics

Various network performance metrics (see Section 4.1) havebeen applied in the investigation of clustering algorithms in CRnetworks. Examples of the tradeoffs are as follows:

� Cluster size (see Section 4.1, P.1) and number of common channels in acluster (see Section 4.1, P.2). Larger cluster size increases the numberof member nodes in a cluster; and it has been shown to reduce theoverall network overhead (e.g. resource management and routing).

However, larger cluster size reduces the number of commonchannels in a cluster (Kim, 2009), and so it affects cluster stabilityand may increase the occurrence of re-clustering.

� Number of common channels in a cluster (see Section 4.1, P.2) andnumber of common channels with neighboring clusters (seeSection 4.1, P.9). While higher number of common channels ina cluster and with neighboring clusters are both favorable,achieving a balanced number of channels for both intra-clusterand inter-cluster communications are both important to ensureconnectivity throughout the entire networks. However, highernumber of common channels in a cluster may cause lowernumber of common channels with neighboring clusters.

Future research could be pursued to investigate approachesto achieve a balanced tradeoff among the network performancemetrics.

5.4. Gateway's schedule for intra-cluster and inter-clustercommunications

Gateway nodes, which are the member nodes located at thefringe of a cluster, can hear from neighboring clusters, and so theyprovide both intra-cluster and inter-cluster communications. Sinceadjacent clusters may use distinctive common channels, regularchannel switches may be necessary at gateway nodes. An optimalschedule for channel switches is necessary so that a gateway nodelistens to the right channel, and this ensures successful transmis-sions while minimizing packet collisions in both clusters.

5.5. Effects of clustering to cognitive radio schemes

Clustered networks support cooperative tasks, such as channelsensing, dynamic channel selection and routing; however, theeffects of clustering on the QoS of a wide range of applications areyet to be explored. Examples of the different applications are asfollows:

� In channel sensing, the SU member nodes sense for whitespaces and send the sensing outcomes to their respectiveclusterheads for final decisions on channel availability. Theaccuracy of the final decisions on sensing outcomes may beaffected by the number of member nodes in a cluster (or clustersize), with larger cluster size being favorable.

� In dynamic channel selection, nodes in a cluster choose acommon channel for control and data transmissions. Channelcontention may be affected by cluster size, with smaller clustersize being favorable. This means that smaller cluster sizereduces channel contention with neighboring clusters,although larger cluster size has been shown to reduce theoverall overhead (e.g. resource management and routing).

� In routing, larger cluster size reduces routing overheads, whichare broadcast among the clusterheads, although it may incurhigher amount of clustering and control overheads betweenclusterhead and member nodes. Cluster stability may bejeopardized due to the lack of a common channel becauselarger cluster size reduces the number of common channels ina cluster (Asterjadhi et al., 2010).

Future research could be pursued to investigate approaches toachieve a balanced tradeoff among the clustering metrics (e.g.number of clusters, and number of common channels in a cluster)and application metrics (e.g. accuracy of the sensing outcomes andchannel contention level). Further research could also be pursuedto investigate a clustering framework in which the effects ofclustering on network performance and QoS are investigated inregards to the operation of CR as a whole, rather than the effects of

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 93

Page 16: Clustering algorithms for Cognitive Radio networks: A survey

clustering on some particular objectives or applications (Guo et al.,2009).

5.6. Optimal cluster size

There are pros and cons with larger and smaller cluster sizes.Clustering algorithms may determine the optimal cluster size,although it may change with network scenarios. For instance,cluster size may reduce with higher nodal mobility (Huang et al.,2011).

There are three main advantages of smaller cluster size (orlarger number of clusters). Firstly, it reduces energy consumptionassociated with the exchanges of local clustering information, suchas the contribution values (Wei and Zhang, 2010) (see Section3.4.1). Secondly, it increases parallelism and so it reduces proces-sing delay. Thirdly, it increases cluster stability because it increasesthe number of common channels for intra-cluster communication,and reduces the occurrence of re-clustering.

The main advantage of larger cluster size is that, it reducesoverhead associated with inter-cluster communication, particu-larly resource management and routing overheads.

Future research could be pursued to achieve an optimal clustersize in order to achieve a balanced tradeoff.

5.7. Other open Issues

More in-depth research could be pursued in most clusteringalgorithms in order to further understand and enhance variousaspects of the new and existing algorithms. The following inves-tigations can be performed to enhance clustering algorithms:

� Generally speaking, Table 1 shows a wide range of potentialopen issues and enhancements that could be further investi-gated. For instance, there has been lack of research efforts inthe investigation of providing shorter intra-cluster distance(see Section 4.1, P.6), lower level of overlap among clusters (seeSection 4.1, P.7), and lower error probability in sensing out-comes (see Section 4.1, P.10).

� Cluster size tends to change with the number of commonchannels in a cluster, and so it varies with PU activities. Hence,a highly dynamic operating environment may affect clusterstability. This may cause high occurrence of re-clustering withconstant migration of clusterhead, cluster merging, clustersplitting, node joining and node leaving. Future research couldbe pursued to investigate clustering in highly dynamic operat-ing environment.

� All connected nodes must be clustered, or associated with atleast a single cluster, within a certain amount of time duringcluster formation. Further research could be pursued to ensurethat cluster formation in most clustering algorithms formsconnected and clustered networks if the equivalent flat net-works (or non-clustered networks) are connected.

� Mathematical models of the clustering algorithms for CR net-works can be developed so that the algorithms can be analyzedmathematically. Various analytical results, such as the mean,upper and lower bounds of network performances, achieved bythe algorithms can be derived. Additionally, the clusteringalgorithms can be analyzed with respect to various networkscenarios, such as single-hop or multiple-hop clusters.

6. Conclusions

In this article, we have presented a review on clusteringalgorithms to establish single-hop or multiple-hop clusters inCognitive Radio (CR) networks. There are five types of clustering

objectives, namely the establishment of common control channel,minimizing the number of clusters in the network, as well asenhancements on cluster stability, energy efficiency and coopera-tive tasks; and four types of clustering metrics, namely channelavailability (e.g. the number of common channels in a node pair),geographical location, signal strength and channel quality, as wellas node degree. The algorithms are characterized by clusteringobjectives and clustering metrics. Generally speaking, clusteringalgorithms aim to achieve lower number of clusters in a networkand higher number of common channels in each cluster. Certainly,there is plenty of future work in addressing the open issuesassociated with clustering algorithms in the conventional CRnetworks, and this article has laid a great foundation and sparkednew research interests in this area which has remained in theinfancy stage.

Acknowledgment

This work was supported by the Malaysian Ministry of Science,Technology and Innovation (MOSTI) under Science Fund 01-02-16-SF0027.

References

Akyildiz IF, Lee WY, Vuran MC, Mohanty S. Next generation/dynamic spectrumaccess/cognitive radio wireless networks: a survey. Comput Netw 2006;50(13):2127–59.

Alsarhan A, Agarwal A. Cluster-based spectrum management using cognitive radiosin wireless mesh networks. In: Proceedings of 18th international conference oncomputer communications and networks. (ICCCN), San Francisco, CA; 3–6August 2009. p. 1–6.

Asterjadhi A, Baldo N, Zorzi M. A cluster formation protocol for cognitive radio adhoc networks. In: Proceedings of the European wireless conference (EW’10),Lucca, Italy; 12–15 April 2010. p. 955–61.

Baddour KE, Ureten O, Willink TJ. A distributed message-passing approach forclustering in cognitive radio networks. Wirel Pers Commun 2011;57(1):119–33.

Badoi, C-I, Croitoru, V, Popescu, A. HC-IPSAG cognitive radio routing protocol:models and performance. In: Proceedings of the wireless and optical commu-nications networks (WOCN’11), Paris, France; 24–26 May 2011. p. 1–5.

Bettstetter C, Konig S. On the message and time complexity of a distributed mobility –

adaptive clustering algorithm in wireless ad hoc networks. Available online: ⟨http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.98.1082&rep=rep1&type=pdf⟩;2013 [accessed 08.08.13].

Bradonjic M, Lazos L. Graph-based criteria for spectrum-aware clustering incognitive radio networks. Ad Hoc Netw 2012;10(1):75–94.

Chen T, Zhang H, Maggio GM, Chlamtac I. Topology management in CogMesh: acluster-based cognitive radio mesh network. In: Proceedings of the IEEEinternational conference on communications (ICC’07), Glasgow, Scotland; 24–28 June 2007. p. 6516–21.

Ephremides A, Wieselthier JE, Baker DJ. A design concept for reliable mobile radionetworks with frequency hopping signaling. Proc IEEE 1987;75(1):56–73.

Fan Z, Rocky Z. Spectrum allocation and medium access in cognitive radio wirelessnetworks. In: Proceedings of the European wireless conference (EW’09),Aalborg, Denmark; 17–20 May 2009. p. 90–5.

Federal Communications Commission (FCC). First report and notice of rulemakingin the matter of unlicensed operation in TV broadcast bands. Resourcedocument. FCC. Available online: ⟨http://www.apwpt.org/downloads/fcc_2012-11906_exemptedt.pdf⟩; 2006 [accessed 16.08.13].

Gong L, Chen J, Tang W, Li S. A multi-channel access-based clustering protocol inhierarchical spectrum sharing network. In: Proceedings of 11th IEEE Singaporeinternational conference on communication systems (ICCS’08), Guangzhou,China; 19–21 November 2008. p. 1005–9.

Guo C, Peng T, Xu S, Wang H, Wang W. Cooperative spectrum sensing with cluster-based architecture in cognitive radio networks. In: Proceedings of the IEEE 69thvehicular technology conference (VTC Spring’09), Barcelona, Spain; 2009.p. 1–5.

Huang X-L, Wang G, Hu F, Kumar S. Stability–capacity–adaptive routing for high-mobility multihop cognitive radio networks. IEEE Trans Veh Technol 2011;60(6):2714–29.

Jeng AAK, Jan R-H. The r-neighborhood graph: an adjustable structure for topologycontrol in wireless ad hoc networks. IEEE Trans Parallel Distrib Syst 2007;18(4):536–49.

Khan AuR, Madani SA, Hayat K, Khan SU. Clustering-based power-controlledrouting for mobile wireless sensor networks. Int J Commun Syst 2011;25(4):529–42.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–9594

Page 17: Clustering algorithms for Cognitive Radio networks: A survey

Kim C-J, Kim S-W, Kim J, Pyo C. Dynamic spectrum access/cognitive radio activitiesin Korea. In: Proceedings of the IEEE symposium on new frontiers in dynamicspectrum access networks (DySPAN’10), Singapore; 6–9 April 2010. p. 1–5.

Kim M-R. Distributed coordination protocol for common control channel selectionin multichannel ad-hoc cognitive radio networks. In: Proceedings of the IEEEconference on wireless and mobile computing, networking and communica-tions (WIMOB’09), Marrakech, Morocco; 12–14 October 2009. p. 227–32.

Lee K, Lee H. Energy-efficient self-organized clustering with splitting and mergingfor wireless sensor networks. Int J Distrib Sens Netw 2013:1–11.

Li D, Gross J. Robust clustering of ad-hoc cognitive radio networks underopportunistic spectrum access. In: Proceedings of the IEEE internationalconference on communications (ICC’11), Kyoto, Japan; 5–9 June 2011. p. 1–6.

Li X, Hu F, Zhang H, Zhang X. A cluster-based MAC protocol for cognitive radio adhoc networks. Wirel Pers Commun 2012;69(2):937–55.

Liming X, Xiaohua J, Kunxiao Z. QoS multicast routing in cognitive radio ad hocnetworks. Int J Commun Syst 2012;25(1):30–46.

Liu C. Design on common control channel of MAC protocol of cognitive radionetworks. In: Proceedings of the international conference on electrical andcontrol engineering (ICECE’10), Wuhan, China; 25–27 June 2010. p. 3621–4.

Liu S, Lazos L, Krunz M. Cluster-based control channel allocation in opportunisticcognitive radio networks. IEEE Trans Mob Comput 2012;11(10):1436–49.

Nafees M, Islam AKMM, Zareei M, Baharun S, Komaki S. Spectrum aware cluster-based architecture for cognitive radio ad-hoc networks. In: Proceedings of 2ndinternational conference on advances in electrical engineering (ICAEE’13),Dhaka, Bangladesh; 2013. p. 181–5.

Nekovee M. Impact of cognitive radio on future management of spectrum. In:Proceedings of 3rd international conference on cognitive radio orientedwireless networks and communications (CROWNCOM’08), Singapore; 15–17May 2008. p. 1–6.

Network Simulator 3 (NS-3). Available online: ⟨http://www.nsnam.com⟩; 2013[accessed 16.08.13].

OMNeTþþ . Available online: ⟨http://www.omnetpp.org⟩; 2013 [accessed 16.08.13].Office of Communications (Ofcom). Digital dividend review, a statement on our

approach to awarding the digital dividend. Resource document. Ofcom. Avail-able online: ⟨http://stakeholders.ofcom.org.uk/binaries/consultations/ddr/statement/statement.pdf⟩; 200. [accessed 16.08.13].

Ozger M, Akan OB. Event-driven spectrum-aware clustering in cognitive radiosensor networks. In: Proceedings of the IEEE conference on computer andcommunications (INFOCOM’13), Turin, Italy; 14–19 April 2013. p. 1483�91.

Peiravi A, Mashhadi HR, Javadi SH. An optimal energy-efficient clustering methodin wireless sensor networks using multi-objective genetic algorithm. IntJ Commun Syst 2013;26(1):114–26.

Qualnet. Available online: ⟨http://web.scalable-networks.com/content/qualnet⟩;2013 [accessed 16.08.13].

Ramli A, Grace D. RF signal strength based clustering protocols for a self-organizingcognitive radio network. In: Proceedings of 7th international symposium onwireless communication systems (ISWCS’10), York, UK; 19–22 September 2010.p. 228–32.

Rayment SG, Eccelsine P, Chouniard G, Lubar D, Christensen M, Austin M,Sandmann M. Regulatory tutorial material. In: IEEE ECSG on whitespace sg-whitespace-09/0048r5. Available online: ⟨https://mentor.ieee.org/802-sg-whitespace/dcn/09/sg-whitespace-09-0048-05-0000-regulatory-tutorial-material.ppt⟩; 2009 [accessed 16.08.13].

Sharma P, Abrol V. Optimized cluster head selection & rotation for cooperativespectrum sensing in cognitive radio networks. In: Proceedings of 10th inter-national conference on wireless and optical communications networks(WOCN’13), Bhopal, India; 2013. p. 1–5.

Talar AC, Altilar DT. United nodes: cluster-based routing protocol for mobilecognitive radio networks. IET Commun 2011;5(15):2097–105. http://dx.doi.org/10.1002/dac.1280.

Talay AC, Altilar DT. Self adaptive routing for dynamic spectrum access in cognitiveradio networks. J Netw Comput Appl 2013;36(4):1140–51.

Wei J, Zhang X. Energy-efficient distributed spectrum sensing for wireless cognitiveradio networks. In: Proceedings of the IEEE conference on computer andcommunications (INFOCOM’10), San Diego, CA; 15–19 March 2010. p. 1–6.

Xu G, Tan X, Wei S, Guo S, Wang B. An energy-driven adaptive cluster-basedrotation algorithm for cognitive radio network. In: Proceedings of 1st interna-tional conference on pervasive computing signal processing and applications(PCSPA’10), Harbin, China; 17–19 September 2010. p. 138–41.

xG technology. ⟨http://www.xgtechnology.com⟩; 2013 [accessed 6.04.13].4.Zhang H, Zhang Z, Dai H, Yin R, Chen X. Distributed spectrum-aware clustering in

cognitive radio sensor networks. In: Proceedings of the IEEE global telecom-munications conference (GLOBECOM’11), Houston, Texas; 5–9 December 2011.p. 1–6.

Zhang J-Z, Fan W, Yao F-Q, Zhao H-S, Li Y-S. Cluster-based distributed topologymanagement in cognitive radio ad hoc networks. In: Proceedings of theinternational conference on computer applications and system modeling(ICCASM’10), Taiyuan, China; 22–24 October 2010. p. 544–8.

Zou C, Chigan C. On game theoretic DSA-driven MAC for cognitive radio networks.Comput Commun 2009;32(18):1944–54.

K.-L.A. Yau et al. / Journal of Network and Computer Applications 45 (2014) 79–95 95