Cloud Networking Formation in CogMesh Environment

1 Cloud Netw orking Forma tion in CogMesh En vironment T ao Chen, Member , IEEE, Honggan g Zhan g, Member , IEEE , Marcos D. Katz, Me mber , IEEE Abstract — As radio spectrum usage paradigm moving from the traditional command and control allocation sch eme to the open spectrum allocation scheme, wireless networks meet n ew opportunities and challenges. In thi s article we introduce the concept of cognitiv e wireless mesh (CogMesh) networks and address the unique problem in such a network. CogMesh is a self- orga nized distributed network ar chitecture combining co gnitive technologies with the mesh structure in order to pro vide a uniform service platform ov er a wid e range of networks. It is based on dyn amic sp ectrum access (DSA) and featured by self-organization, self-configuration and self-healing. The uniq ue problem in CogMesh is the common contro l channel problem, which is caused by the opportuni stic spectrum sharing nature of secondary users (S U) in the network. More precisely , si nce the channels of SUs are fluctuating according to the radio en vironment, it is difficult to find always a vailable global common control channels. Th is pu ts a significant challenge on th e network design. W e develop the control cloud b ased control ch annel selection and clu ster based n etwork formation techniques to tackle this problem. M oreo ver , we show in this article th at the swarm intelligence is a good cand idate t o deal with the control channel problem in CogMesh. Since the study of cognitive wireless networks (CW N) is still in it s early phase, th e ideas prov ided in this article act as a catalyst t o in spire new solutions in this field. I . I N T R O D U C T I O N Radio spectrum usage is undergoin g a paradigm shift from the tradition al co mmand an d contro l allocation to th e DSA [1]. Cognitive radio (CR) is a p romising approach to achieve open spectr um shar ing flexibly a nd efficiently [2] . The re- search on CR has alread y penetrate d in to different types of wireless networks, and covered almost e very aspect in wireless commun ications [ 3]. Follo wing the CR, the c oncept of CWN comes out with the emphasis on the network-wide cog nition and adap tation [4]-[5] . Although a CWN m ay not rely on CR technologies sole ly , it is a co mmon assum ption tha t CWNs are CR based to some extent and use DSA as the sp ectrum access scheme. W e follo w this assumption in the article. In a DSA based network the SUs of the spectrum opportu nistically access the spectrum based on the activities of the p rimary users (PU) as well as the r adio en vironm ent. The defin itions of PU and SU can be found in [3]. For the deep un derstandin g of DSA, please refer to the survey o f [6]. In this article, startin g fro m our p revious related work [7],[8 ], we first introduc e the c oncept of CogMesh , wh ich is defined as a self-organized distributed network arch itecture combinin g cogniti ve technologies with the mesh stru cture in T . Chen and M. Katz ar e with VT T , Finland. H. Zhang is wit h Zhejia ng Univ ersity , China. order to provide a u niform service platfo rm over a wide range of n etworks. A network based on this architec ture is featured by self-organizatio n, self-co nfiguratio n and self- healing. It shows similarities with wir eless ad ho c networks on the aspects of distributivity and self-organ ization. Ho wever , a CogMe sh network is more flexible on spectrum, energy a nd network resour ce usage, the refore being sup erior to wireless ad ho c networks on performan ce and resource efficiency . W e call th e environment where CogMesh networks are o perated the CogMesh environment. Considering that at present the study on CWNs is still o n its early phase, it is interesting to show rea ders the u nique problem s in the CogM esh en v ironmen t and potential solutions. W e iden tify the control chann el as one of the main challen ges in CogMesh. Indeed, in the CogMesh environment, because SUs oppo rtunistically share spectrum with PUs, the network cannot rely on a glo bal common control channel for coordi- nation. This is different from co n ventional wireless networks where the com mon control chann els are usually assumed. In this article, we will analy ze the commo n co ntrol chann el problem in CogMesh an d p ropose feasible solutio ns. For a distributed n etwork it is desirable that th e nod es sha re a common control channel in or der to p rovide reliable and efficient c ommun ications, and reduc e the control overhead . W e first pr opose a contro l cloud con cept fo r the contro l channel selection. A contr ol clo ud is a group of connected SUs that share a common control channel. Fu rthermo re, on e can find some similarities b etween CogMe sh networks and the collective behavior of social insects. From such an analogy , we introdu ce th e swarm intelligenc e mechanism in to Cog Mesh and use distributed algorith ms to for m the co ntrol clou ds as large as possible. A larger co ntrol clo ud means m ore nodes sharing the same control channel. Th en, based on the control clou ds, a cluster based network fo rmation scheme is employed to furth er con solidate th e spectru m m anagemen t. The main advantage of th e swarm intelligence based control channel selection and cluster b ased network formation is their adaptability to the radio and n etwork environment chang e, which is impo rtant for CWNs. In the re mainder of the article, we will first introdu ce the concept of Cog Mesh, and discu ss its differences with conv en- tional wireless networks. T hen we will descr ibe the contro l channel prob lem in CogMesh , and provide o ur solutions. A wireless MAC protoco l tailo red fo r CogMesh will be given to explain how th e network discovery and c luster f ormation are p erform ed. W e also gi ve the simu lation results illustrating the behavior of the proposed solutions. Finally , we draw the conclusion . 2 I I . C O N C E P T O F C O G M E S H CogMesh, as sho wn in Fig. 1, is a fle xible network archi- tecture exploiting a mesh topology to integrate he terogeno us wireless networks und er a un iform but loo sely organized control plane. It comb ines the advantages of CR systems and autonom ous networks in a seamless way with th e aim to provide a flexible n etwork platform ad aptive to a variety o f existing and emerging services. Primary User Seco nda ry U se r Lice ns ed Band I Lice ns ed Band II Unlice nsed B and Primary User Cog nitive wireles s netwo rk with Infras tructure CR Use r U s e r U s e r U s e r Us U s U s U s U s U s U s e r U s e r U s e r Spectrum b and Cogni tiv e Ad -Hoc network Dyna mic Spec trum Acc ess Fig. 1. A CogMesh net work. CogMesh is a wider concept than CR since the main concern of CR is the awareness, understan ding and adap tation o f radio resources, such as spectru m, tim e, space an d po wer . Sinc e, for instance, the p ower control is actually a network le vel prob lem [9], witho ut the necessary suppor t f rom the network level, the flexibility broug ht b y CR is limited. Moreover, th e new open spectrum access paradigm creates a wireless ecosystem in which sub-systems w ork in a highly coup led way . It requires new network d esign pr inciples to fu lly release the power of this new wireless ecosystem. As we kn ow , the wireless medium is unstru ctured in na- ture. The mesh structure is the natu ral way to match th is characteristic as it provides c hoices to inter connect wireless devices in e very possible way . It not only mean s the coverage extension o f wireless networks, but also acts as a metho d to efficiently utilize reso urces among networks. The me sh here means the network on deman d based on whatever and whenever services may r equire. Thus, the network topo logy control resu lts into a joint o ptimization p rocess as a func tion of service r equirem ents, cond itions of radio and network en vironm ents to ultimately fulfil the service go als and resource efficiency . In this pro cess the co gnition play s a fundamen tal role. CogMesh is a rather open architecture ha v ing v arious forms: it can be an integration of different centralized wireless net- works through a mesh stru cture; it can b e an ad ho c network built up on th e DSA parad igm; or it can b e a comb ination of aforemen tioned centralized and ad hoc networks. The common features of Cog Mesh networks are the self-organizatio n, self- optimization and self- healing capabilities. From the spectrum access perspectiv e, a CogMesh network is typ ically formed by PUs an d SUs o f the spectru m. A SU is allowed to access a spectrum b and only when causing tolerable interferen ce to the adjacent PUs on that frequ ency band. W e ca ll that freq uency ba nd a spectrum hole , wh ich is defined as a piece of spectrum no t occupied by any PU at a given time in a given geo graph ic area. Th e toler ant interferen ce of PUs can be well described by the concep t of in terfer ence temp eratur e [ 2], which is a metr ic u sed to quantify the interf erence in a radio en vironm ent. Althou gh having be en temporarily abando ned by the FCC spe ctrum policy task force, inter ference temp erature re mains pr oviding an accurate measure at a receiver for the acceptab le le vel of R F interferen ce in the spectrum band of interest, therefore playing an importan t r ole in the opportu nistic spe ctrum sharing. The interferen ce temp erature limit of a PU serves as a cap placed on p otential RF en ergy that could be introd uced by SUs o n a given band. Th e co ncepts of the spe ctrum hole and th e interferen ce temp erature are illustrated in Fig. 2. v v Ti me Power Spectrum Hol es Occupi ed b y PU f1 f2 f3 No ise flo or Power a t rec eiver Inte rf ere nce te mperatur e limit Licensed signal Opportunitie s of spectr um access for SU PU on frequency f1 PU on frequency f2 PU on frequency f3 SU Spectrum hole Interference temperature Fig. 2. S pectrum hol e and interfere nce te mperature. According to th e interference temp erature lim its o f PUs o n different frequen cy bands, which are pre-spe cified, spectrum holes are detected thr ough the spec trum sensin g pro cesses perfor med by SUs. Commun ications of SUs ar e con ducted throug h channels extracted from th ose sp ectrum holes. No te that the spectrum hole and cha nnel ar e different c oncepts: a spectrum ho le is a continuo us spectrum space of any size, wh ile a channel is a p re-specified spectru m agr eed by commun ication entities. According to the specifica tion on the size of channels, a spectrum hole may hold zero to multiple channels. For each SU, the ou tcome of the spectrum sensing is the a vailable channel set from multip le sp ectrum holes. From the ch annel set one chan nel is used as the con trol chan nel for the network coordin ation. Due to the comp lexity of the radio and network environment in CogMesh, the control chann el problem becomes a prom inent one. It is the focus of this article to analyze this problem and show p otential solution s. I I I . S T U D I E D S C E N A R I O O F C O G M E S H Under the self-o rganization frame work of CogMesh, there are two basic fo rms of ne tworking: centralized and a d hoc 3 networking. T o study the contro l chan nel pr oblem, we fo- cus our a ttention on the ad hoc networkin g of Co gMesh since it represen ts th e natur e of self- organization and self- reconfigu ration. By regarding the centra lized networking as a special form of clustering, we are able to study CogMesh as a distributed network in a wide sense, especially from the control viewpoint. Therefore in this ar ticle the stud ied scenario of the CogMesh is the follo w ing: the SUs o f CogMesh f orm an ad hoc network; the SUs coexist with PUs a nd oppo rtunistically share the spe ctrum; no g lobal commo n c hannel exists for th e control purpose of SUs; b y using local control ch annel, SUs forms a multi-h op ad hoc network. As we can see, the scenario becomes a m ulti-hop ad ho c n etwork u nder the DSA scenario . A. Multi-chan nel W ireless Systems A CogMesh network can be mo deled as a mu lti-channe l wireless system, in which the a vailable chann els of a node v ar y during the life time of the nod e. The network coordin ation over multi- channel systems h as been an interesting r esearch problem . Prop osals f or conventional mu lti-chann el wireless systems usually assume channels are av ailable all the time. T wo different assumptions are u sed when d esigning conven- tional multi-channel wireless system s: using sin gle transcei ver or usin g multip le transceiver . The former is the typical c on- figuration of cu rrent wireless devices. Only one tran sceiv er on th e d evice m eans it is no possible to transmit and recei ve simultaneou sly . T he latter assumes at least tw o transcei vers at one de v ice, ther efore bein g cap able of sending and receiving on different channels at the sam e time. For the sing le transceiver case, three basic methods are u sed to coo rdinate the multi-ch annel access: • Common control channel method , in which a commo n control chann el is u sed fo r the signaling o f all nodes. The typical case is m ulti-chan nel MAC (MMAC) [10 ], which defines a default con trol channel where all no des must periodically switch to an d sync hronize for a pr e- determined wind ow o f time. The channel fo r data tran s- mission is n egotiated in that pre-deter mined win dow . • Channel hopping method , in wh ich every nod e h ops its working chan nel according to certain pattern and has chance to m eet o ther nodes per iodically on different channels. Th e hop reservation multiple access ( HRMA) [11] is o ne instance o f this m ethod. In HRMA, all nodes ho p accord ing to a pr e-defined hopp ing pattern. Whenever a n ode ha s a data p acket to send , it exchan ges control messages with the intend receiver and both remain in the same hop p attern for th e entire data transmission. • Home cha nnel for receiver method , in wh ich a pre- defined home cha nnel is assigned to each nod e, and nodes a re switched to their home channels f or receiving immediately when they are idle. For instan ce, in [12 ], ev ery node is associated with a home channel based on node’ s MAC ad dress. After data transmission , a node immediately return to its home channel for in coming packets. For the multiple transceiv er case, the multi-channel access becomes s imple sin ce a dditional tran sceiv ers can be used for control purp ose. For instance, in [13], nodes are assumed to have a s many tran sceiv ers as the number of c hannels, bein g able to listen to all those chann els simultaneously . A nod e having data to send simply picks up a n idle chan nel for transmission. Obviously the cost for this approac h is extremely high. Other ap proach es use only two transceivers, on e f or the control and the o ther for the data tr ansmission [14]. The control transcei ver alw ays works on the default control channel to negotiate the data chan nels. As we can see, the multi-ch annel access solutio ns ar e all based on the assum ption that a ll channels ar e always a vailable. This is not th e case in Cog Mesh since th e availability o f a ch annel for a SU dep ends on th e r adio environment. I t means that we can no t use con ventional mu lti-chann el access solutions direc tly . Mo reover , we can not assume th at every node in Co gMesh has multiple tr ansceivers. Based on this, we need to d ev elop solution s fo r a gen eral case, i.e., the single transceiver working on the h alf duplex mode. B. Contr ol Channel Pr oblem in C ogMesh Networks It is well kn own th at the con trol pro blem is critical in distributed networks, due to the dynamics introduced by self- coordin ation activities. Un til n ow the major ity of pro posed spectrum con trol protoc ols d esigned for the DSA scenario assume the a vailability o f a common control channel [3]. For instance, Jing et al. [15 ] used common spectr um coo rdination channel (CSCC) etiqu ette protoco l f or coexistence of IEEE 802.1 1b and 802.1 6a network s; a cog nitiv e pilot chann el (CPC) conc ept was p ropo sed in [1 6] for exchan ging the spectrum inform ation among nodes. The use of the co mmon contr ol ch annel significantly r e- duces the comp lexity of the network coordination . However , the comm on control channels do no t alw a ys exist in CogMesh, since the SUs of Cog Mesh utilize spectr al h oles fo r com - munication s. Correspon dingly , the topolog y manageme nt of CogMesh is a ffected by two main factors: fir st, the ab sence of a comm on control channel in the network ; and second, the frequen t to pology changes accordin g to the presence of PUs and SUs. In the CogMesh environment, SUs use loc al con trol channels for the network coo rdination . Howe ver, u ntil now only fe w propo sals are mad e un der the non-co mmon control chan nel assumption . For instance, Zh ao et al. observed that though very limited numb er of globa l common c hannels exist in a network, local neigh bors share numero us chann els with others [1 7]. They prop osed a d is- tributed grouping scheme to solve the common control channel problem [17] ; Bian et al. [ 18] used th e con cept of the segmen t, which is a grou p of nodes who share comm on channels along a routing p ath, to organize contr ol ch annels. In [7], this pr oblem was tackled by a cluster-based appro ach, where the local u sers sharing common chann els form a dynamic one-hop cluster and the spectrum is m anaged by clu ster heads. In the following, we will use the contro l cloud conc ept and cluster based n etwork fo rmation to d eal with th e c ommon control channe l p roblem, and describe how the CogMesh network is co ordina ted for multi-hop commun ications based on the pro posed solutions. 4 I V . N E T W O R K C O O R D I N AT I O N I N C O G M E S H N E T W O R K S A. Contr ol Cloud Co ncept 2 1 2 1 2 3 1 {1,2} {1,2} {1,2,3} {1,2,3} {1,3} {1,3} {1,3} {1,3} {2,3} {2,3} {1,2,3} {1,2,3} {2,3} {2,3} {1,3} {1,3} {3} {3} Coverage of PU Cloud shares control channe l on channel 3 n - PU on channel n {1,2,3} - SU - Channel list of SU Fig. 3. Control cloud in CogMesh netw orks. A c ontr ol cloud is a collection of neig hbor SUs sharin g a co mmon contro l chan nel. W e call it clo ud becau se it m ay dynamica lly chang e its size accordin g t o the radio and network en vironm ent. If the whole SU network shares a c ommon control chann el, the who le network is u nder the c ontrol of a sing le co ntrol cloud. Otherwise, there ar e m ultiple co ntrol clouds separatin g the network. The reason to in troduce the c ontrol cloud in Cog Mesh is to provide a scalable control solution fo r CogMesh in the DSA scenario. The id ea is to make co ntrol clouds grow and cover as much SUs as possible while adapting to the radio en vironm ent. The e volution of co ntrol clo uds in the ne twork r esults from the self-organized activities of the network. Different control clouds are interconnected throug h gate way nodes between th e edges of the clouds. Control cloud s result fr om individual no des’ choices on their control channels. According to the channel quality , a node chooses a channel as its control channel, namely th e master channel , for signalin g. In case that two neigh bors choose different master channels, a proper li stening rule can be used to perfor m the neigh bor discovery o rderly o n other c hannels ac- cording to their chan nel qualities. On ce a n eighbo r is d etected, the proposed alg orithm is ru n to negotiate a co mmon master channel amo ng most of the neigh bors. According ly , channel clouds a re formed and evolved with a trend to form few and large clou ds as possible. Clearly , con trol messages run ning over few control channels redu ces the co ntrol overhead and delay . T o setup the master channel, we use the following neigh bor discovery pro cess. Suppo rted b y the layer two or three, a SU p eriodically broad casts HELLO messages over its master channel. The HE LLO m essage includes th e in formatio n of the node’ s master chan nel and all other available ch annels with the quantized quality values. The neighbor s of th e SU listen to their m aster chann els in most of time for HELLO me ssages, and shift the listening to other ch annels with pr obabilities propo rtional to their ch annel qua lities for a gi ven p eriod in a repeating man ner . Once the channel inform ation is exchanged among th e neighbors, a common master channel sha red b y th e neighbo rs will be negotiated by the propo sed algorithm. Therefo re the control chann el selection algor ithm deter- mines the for mation and evolution of control clouds. W e notice that th ere is a collective behavior in the control cloud formation , whe re each node makes its own decision. This effect is similar to the collective p henom enon widely seen in the biolo gic world. This mo tiv ates us to intro duce a swarm intelligence algor ithm into CogMesh targeting for con trol channel problems. The swarm intelligence is a well established science biolog ically inspired by the collectiv e b ehavior o f social insects, f or instance, ants o r bees so lving complex tasks like building nests or f oragin g [19]. It is based on the principle of th e division o f lab or wher e the higher efficiency is a chieved by specialized workers perform ing specialized tasks in par- allel. The advantages of swarm intelligenc e techn iques are scalability , f ault tolerance, p arallelism and autonomy . Swarm intelligence alg orithms have b een su ccessfully e mployed in telecommun ication n etworks for the perfor mance im prove- ment of routing protoc ols [ 20], [21]. Recently , its applications have been found on spectrum sen sing and reso urce allocation in CWNs [22 ]. In a typical swarm intelligence scenario , an agent deposits a small amo unt of pher omone on a trail and the trail with higher ph eromo ne level bec omes the choice of the working trail. T his d istributed optimiza tion appro ach relies on th e cooper ation of agents to ac hieve the co mmon optimizatio n goal with a collecti ve complexity o ut of individual simplicity . Considering each SU in a CogMesh n etwork as a simple a gent and its ch oice on the control channel as th e pher omone, the swarm intelligen ce matches well the dynamics in the CogMesh network. W e will describe the detail of th e swarm intelligence algorithm in Section V. B. Cluster Ba sed Networking In a ddition to th e con trol c loud concept, we use the c luster based network ing f or the network f ormation in CogMesh networks. The purpo se of using a cluster based approach is to make the spectru m access more man ageable in the DSA s ce- nario. Managing the spectrum as a whole in the cluster reduces the contr ol overhead when compare d with the way to do it in a fully ad hoc manner , especially when SUs are coexisting with PUs. Moreover , the cluster based app roach has advantages fo r routing in the multi-hop n etworking en viron ment. A large number of cluster formation alg orithms h av e been propo sed for a d hoc ne tworks so far [ 23],[2 4]. They are different on the criteria to select cluster h eads. Howe ver , there are some critical prob lems to utilize those ap proache s in CogMesh networks: 1) They a re usually designe d f or the single chan nel case while CogMesh is a multi-channe l case; 2) They are design ed fo r fixed network topology , a nd lack the cap ability to ad apt to dynam ic physical to polog y changes; 3) Most of them only guarantee the network c onnectivity , and theref ore the result may not be optimized; 4) In som e solution s the f ull k nowledge of the n etwork is required , whic h is no t realistic in CogMesh . As a conclusio n, a different ap proach is needed in CogMesh. In CogMesh, a cluster is a grou p of neighbor SUs controlled by a cluster head, wh ich is selected from that grou p of SUs. Normally the me mbers of th e c luster is one-hop away from 5 the clu ster head. Un der the co ntrol cloud concept, the SUs in a cluster ar e the membe rs of the same contro l clou d. Following the same r ule, we call the contro l ch annel of a cluster th e master chan nel of that cluster . The nod e forming the cluster becomes the cluster head, which is responsib le for intra-cluster channel access con trol and inter-cluster commun ications. By negotiating gateway node s b etween clu sters, clusters a re interconn ected to a large network , wher e multi-hop lin ks are used to deli ver data messages. A g atew ay node is a me mber of one cluster that can r each the memb er of anoth er cluster . The cluster in terconn ection is illustrated in Fig. 4, f rom which, we can see that clusters are interconne cted in tw o cases: two cluster heads are connected by one gate way node, or conn ected by two gate way node s when no node is one-hop neighbor of two cluster heads. There fore there are three types of mem bers in a cluster: the cluster head, ordinary node, and gatew ay node. Cluster A Gateway node Channel 1 Channel 2 Cluster B Cluster C Clusterhead Ordinary node Clusterhead Gateway node Ordinary node Fig. 4. Clusters intercon nected by gate way nodes. In this article, we will show a specific clu ster fo rmation algorithm design ed for CogMe sh, with the ab ility to adap t to the radio and netw ork environment. Before pro ceeding to the algorithm , we will introd uce the MA C f unctions th at assist the network formation in CogMesh . C. MAC Functions for Network F ormation Beacon Neighbor Discovery Period Data Period Private Random Access Public Random Access Superframe Spectrum Detection Period Member 1 Member 2 Member 3 Preamble Node ID Cluster Info Neigbhor ID Channel List Neigbhor ID Channel List ... Data frame Channel 1 Channel n ... Frame Map Fig. 5. S uperframe structure. The clu ster f ormation an d inter - cluster co nnection ar e p er- formed distributi vely based on the n eighbor inform ation of nodes. W e p rovide mechanisms in the MAC pr otocol to enable nodes exchange their one-hop and tw o-ho p neig hbors informa tion, which includes neigh bors’ s id entity and th eir channel list. In Cog Mesh, a node m ay only know partial of its neighbo rs at the in itial stage. The clusters are fo rmed based on the partial neighb or information. As nodes gradually collect more n eighbo r info rmation ba sed o n the ne ighbor discovery mechanism, clusters ar e reconstru cted and intercon nected to a more reliable network struc ture. For each cluster, ch annel access tim e is divided into a sequence of supe rframes. Each superfr ame is divided into se veral perio ds as shown in Fig. 5 . T he beac on period is issued by the cluster head. It contains the time synchronization , control and resource allocation informatio n o f the cluster . The following period is th e ne ighbor discovery period. It is divided into a number of fixed length min i-slots. Each mem ber of a cluster occup ies one m ini-slot and uses it to bro adcast the HELLO message, which inclu des its iden tity an d one-hop neighbo r list. An entry in the neigh bor list includes the identity of the neighbo r and its chann el list. The master channel of a neig hbor is ind icated in the cluster list, thro ugh which a node kn ows how to reach the neighbo r cluster . A prea mble is attached at the beginning of each mini slot for other nod es identifyin g the broadcasting m essage if they miss the beacon . Moreover , the time a nd dura tion o f rando m access periods in this superf rame is bro adcast in the Frame Map per iod of its mini-slot. A n eighbor of this no de, once receiving its HELLO message, has the chance to exchange its neighb or information with the node in the following ran dom acc ess period . The location of a membe r’ s mini-slot is announ ced by the cluster head in the bea con perio d. T he nu mber of m ini-slots in a superfram e is limited by a system parame ter in ord er to avoid too many nodes crowding in one cluster . The next per iod is the data pe riod. Parallel transmission s are permitted in this per iod if the transmission sessions use different channels. Following the data period , an intr a-cluster random access period is used for cluster members exchanging control messages. The superframe is ended by a public random access period. Its length is de termined by the cluster head a nd annou nced in the beaco n. This period h as multip le purposes. It uses f or a node jo ining the cluster, nodes exch anging neig hbor informa tion, or clusters exchangin g contro l information. Besides five main periods, there are one or several spectrum detection per iods schedu led in a supe rframe. Dur ing these periods, all membe rs of a cluster keep silence and detect spectrum holes. It is desirable to synch ronize the sp ectrum detection perio ds of adjacent clusters so as to re duce the false alarm of the PU detection. A false alarm occurs when a SU incorrectly reports the presence of PUs du e to the interference from other so urces. Since th e superframes o f different clusters are not requ ired to be synchroniz ed, the location o f the spec- trum d etection perio ds varies from cluster to cluster . Even in a cluster, th eir location v aries from super frame to super frame. 6 D. Neighb or Discovery and Cluster F ormation The neigh bor discovery and cluster for mation processes are introdu ced to gether since they are high ly co upled. For conv enience, we say the n eighbor clu ster of a nod e is the cluster that the node does not belo ng to , but has one-h op neighbo rs as its members, and the total neighb or clusters of all members of a cluster are called the cluster’ s neighbor c lusters. The neig hbor discovery is perfo rmed during clusters’ neig h- bor discovery periods. When a node wants to jo in the n etwork, it first detects the av ailab le channels. Then it scans on e of its channels for a giv en period o f time, waiting f or b eacons on that channel. Th e node starts the scannin g pro cess fr om the channel with the lo w est frequency , which is called the lowest channel. The scanning time o n a channel is ch osen so that it exceed s the period of the longest superfram e. W e call a scanning pe riod as a scanning interval, and the first scanning interval a new node starts as the first scann ing inter val. If there is a neighbo r cluster on the fr equency band a node listens on, it is ab le to captu re its beaco n during a scanning inter val. W e divide the first scanning interval into thre e cases: no message a rrives ; a b eacon ar rives ; or n eighbor messag es arrive but no beacon arri ves. In the first case, the node forms a cluster on the scann ing channel and becomes the cluster head. In th e second case, the no de requests to join the cluster th rough the public ran dom access p eriod o f the cluster . If the cluster head accepts the request, it assigns a min i-slot to th e requesting node. Star ting from next superfr ame, the new joining n ode broadc asts its HELLO messages in that mini- slot. Howe ver, if th ere is no empty m ini-slot in a cluster , the cluster h ead will reject the r equest. T he req uesting no de then chooses th e second lowest chann el to start a new scannin g process, or form its o wn cluster if finding the detected clusters are all full after iterating all ch annels. The third case m eans the no de has neighbo r clusters b ut it is two-hop away from cluster heads. The node then records th e n eighbo r inform ation, and tries to exchange n eighbor information with that neighbor th rough the public rando m access period of the corresp onding neigh bor cluster . Af ter that, it continues its scann ing pr ocess on the n ext av ailable channel. If the n ode can not find a channel satisfying the case one an d two after iterating all chann els, it starts its own cluster on a rand omly chosen channe l. After a no de joins a cluster , it periodically chooses from its channel list a non -master channel to scan so as to discover other n eighbo r no des. An algor ithm can b e developed to intelligently choose a non -master chan nel acco rding to the neighbo r inf ormation the nod e d etects. For instance, if it discovers new two-ho p n eighbo rs on a non -master ch annel, it listens on that channel first. Let us e xplain the neighbo r d iscovery and cluster formation by an example, as illustrated in Fig. 6. The number s in brackets close to each node r epresent a vailable ch annels o f that node. The smaller num ber represents the lower spectrum hole. W e assume the spectrum holes d o no t chang e during the cluster formation pro cedure. The edge between two nodes ind icates they ca n hear each o ther . Assume the node A is the first node formin g the cluster on the chann el 1. The cluster is labeled as th e cluster A. I ts one-hop neighbors B, C, D listen on their lowest f requen cy band , i. e., the channe l 1, d etect the beaco n issued by the cluster A. They join th e clu ster A thr ough a cor respondin g association process. Fro m the neighbo r discovery p rocess, the node B kno ws the node C is its o ne-hop neighb or, and the no de D is its two-h op neig hbor . Next, the nod e E, F , G form a cluster on the channel 2. Assume the node E for ms the cluster, lab eled as the cluster E. The node F , G join the cluster E right after . The node B listens on the non-m aster channel 2 . It discovers E, F as its one-hop neighbo rs, and G as its two-ho p neighb or . The cluster A and E therefor e ar e interconn ected by the node B. Then, assume the node I fo rms the cluster I on the ch annel 3. The node H r eceives B’ s HELLO message and detects B as its one- hop neighb or . Ho wever , H can not recei ve beacons from th e cluster A. It starts a new listen pro cess on the channel 3 a nd finally joins the cluster I. Th e node H inform s B that its new neighbo r list through th e public rand om a ccess period of the cluster A. The nod e B knows from H that there is a cluster on the channel 3. It knows the neighbor s H, I on the ch annel 3 throug h a scanning process on tha t chann el. Fu rthermo re, the node C will kn ow B has ne w neighbors H an d I fr om the neighbo r discov ery period of the clu ster A and finally know its n eighbo r I on th e ch annel 3. At this stage, three clu sters ar e formed , and the cluster heads has enoug h neighbo r information for inter-cluster connec tion. The clusters then negotiate with each o ther to form a network throu gh their public rand om access perio d. A B C D E F G H I {1,2,3} {1,3} {2,3} {1,2,3} {2,3} {1,3} {3} {1,2,3} A B C D E F G H I {1,2,3} {1,3} {2,3} {1,2,3} {2,3} {1,3} {3} {1,2,3} C(A) C(A) C(A) C(A) {2,3} {2,3} A B C D E F G H I {1,2,3} {1,3} {2,3} {1,2,3} {2,3} {3} {1,2,3} C(A) C(A) C(A) C(A) {2,3} C(E) C(E) C(E) A B C D E F G H I {1,2,3} {1,3} {2,3} {1,2,3} {2,3} {1,3} {3} {1,2,3} C(A) C(A) C(A) C(A) {2,3} C(E) C(E) C(E) C(I) C(I) 1. Connect Graph 2. Stage I, Cluster A formed 3. Stage II, Cluster E formed 4. Stage III, Cluster I formed {1,3} Fig. 6. E xample of a cluster formation proce ss. V . C O N T RO L C L O U D F O R M A T I O N B Y S W A R M I N T E L L IG E N C E A P P R O AC H The basic idea of the swarm intellige nce appr oach is to let a node select a channel with suf ficien t quality , me anwhile being preferr ed b y mo st of its neig hbors, as the master ch annel. Sufficient quality mean s th e quality of the chosen channel ranks a higher position a mong all av ailab le chann els. T he reason to cho ose a better chan nel as the contr ol chan nel is straightfor ward: transmission failures are reduc ed. The channel quality is measured in the spectrum sensing process and presented by a single v alue, Q , which is a non-n egati ve r eal 7 value in versely p ropor tional to the a ccumulated in terference imposed by the surrou nding in a g iv en time window . T o make it simple, we qua ntize the Q value into several stages. The Q value th at falls into one stage takes a fixed value which represents that stage. The preferenc e of the neig hbors on the master channel is reflected by the number of neig hbors who choo se the same master channel. W e use an approa ch that takes into account the freshness of the n eighbo rs’ choice. I n each n ode, we maintain a we ight, n amed W , f or each channe l, updating its on ce receiving a HELL O m essage. The channel with th e high est weight is selected as the master c hannel. I n a node, th e ov erall weights of all ch annel is equal to one. Perio dically , a no de selects its ma ster channel acco rding to its W list. The pro blem now becomes how to d etermine the W of each channel accord ing to the Q list of a node and its neig hbors. The W list, which is up dated freq uently according to the fluctuation of the channe l q uality and th e choice s of the neighbo rs on the m aster channe ls, becomes the key to reflect the radio en v ironmen t a nd d etermine the master chann el. W e apply the sw arm intelligence algorithm to update the weig hts. In o ur network, each node acts as an agent, u sing th e HELLO message a s the pheromone to influen ce its neig hbor nodes. A no de receiving a HELLO message updates its W list a s fo llows. T he chan nel equ al to th e m aster c hannel of the broadcasting neighbor recei ves a po siti ve reward with an amount propor tional to the difference of the master chann el quality between the neig hbor and the local no de. All other channels receive negative r ew a rds to m ake the sum of all weights r emain on e. This process ca n be mathematically presented as follows. The param eter W j , which is th e W value of the ch annel i on th e node A , is u pdated by: W i = W i + r (1 − W i ); (1) where r is a parame ter d etermined b y ∆ Q , which is the difference of the master channel qua lity between the neighb or sending the HEL LO message and the node A , i.e., r = f (∆ Q ); where r ∈ [0 , 1] (2) The r fu nction in (2) is a monotonic ally increased function . For all chan nels o ther than the ch annel i , their W values on the no de A are up dated by: W j = W j (1 − r ); for W j ∈ { W j | j = 1 , ..., N ; j 6 = i } (3 ) where N is the total number of chann els on the nod e A . This is a process in which a node persuades its n eighbo rs to mov e to its master chann el. The shift of a master channel happen s when sufficient ph eromon e is accumulate d on a non- master channe l. On the other hand, the channel q uality will be affected b y PUs and th us changed over time. A node updates its channel weight list per iodically a ccording to the refreshed channel quality list. It acts as the distur bed factor to push the master chann el b ack to the best quality channel. The amplification of disturbed factor makes the master channel ev olve with the radio environmen t. As we may notice, the channel weight is u pdated accordin g to the difference b etween chan nel qualities. By choosing different map functio ns between the chan nel quality and th e weight, we are able to contr ol the behavior of the swarm intelligence algorithm. For instance, if we reward the channel weight propor tional to the difference of chan nel qu alities, the control clo ud will keep stable und er the small variation of the radio environmen t; if we reward th e channel we ight in a n opposite w a y , the control clo ud will be m ore sensiti ve to the radio en vironm ent. The example of the r function can be: r = [a rctan( A ∗ ∆ Q ) + B )] / C ; (4) where A , B and C are the constants affecting the conv erging rate of the algorithm. (4) giv es a smaller ∆ Q mo re rew ard than a bigger ∆ Q . An onlin e learning strategy can be ap plied here to tune the map functio ns so that a node can reflect its desire on either the exploitation of the most common channels or the explo ration o f the best quality channels. The prop osed alg orithm has several advantages. First, it is independ ent of a spe cific physical an d MAC layer . As a result, it can be easily in tegrated into h eterogen eous wireless networks. Secon dly , the algorithm is flexible in th e sense that the parameter s o f the algorith m can be tu ned to suit different network scenar ios, fo r instance, adapting the HELLO m essage broadc asting ra te to the radio environment. Since the weigh t v alu e reflects the c hannel quality a nd will- ingness of the nodes to utilize the channels, it has added values for cluster ing, routing an d data tra nsmission. The weigh t value for the master chann el can ease the cluster manag ement in such a network, and then improve the spectrum efficiency . A routing protoco l in tegrating that weight value will be mo re intelligent to a dapt to the r adio en viro nment, the refore being more fle xible an d robust. In addition to using the weight value to choose the control channel, the neighbor SUs can u se it to select the transmission chann el as well, therefore increasing the spectrum efficiency . V I . M D S B A S E D C L U S T E R R E F O R M A T I O N A L G O R I T H M The cluster optimization problem can be considered as a dominatin g set (DS) problem in graph theory , whose objective is to find a subset o f no des called DS w ith the following proper ties: each nod e is either in the DS, or is adjacent to a node in the DS [ 25]. In our network, the DS is the collection of cluster heads. Th e cluster optimization problem is to find a minimal domin ating set (MDS) of the CogM esh network accordin g to its phy sical topolog y . A M DS is the minimal size DS among all possible DSs in the topolog y . T he MDS problem is proven to be a NP-h ard problem even when the complete network top ology is a vailable [ 24]. Ho wever , a sub-o ptimum DS can be ob tained throu gh a local m inimum election of the dominators b y a heuristic algo rithm. The algorithm is run periodically and distributiv ely on each n ode a nd only re lies o n the discovered neighb or inf ormation to determine th e locally optimized clu ster co nfiguratio n. As a result, the collection o f cluster heads is gradually co n verged to a sub-optimum DS. When the phy sical topolo gy changes du e to the ev ents such as new n odes joining the network , nodes leaving the network, or radio en vironm ent changing , the af fe cted n odes or clusters are reconfig ured to immediately ab sorb the changes. 8 0 30 60 90 120 0 10 20 30 PUs: 8 Channels: 5 PUs: 8 Channels: 5 PUs: 2 Channels: 5 W ithout swa rm intellige nce algorith m W ith swarm i ntelligenc e algorithm Standard deviation Number of SU s PUs: 2 Channels: 5 0 30 60 90 120 0 10 20 30 Standard deviation Number of SU s 0 30 60 90 120 0 20 40 60 80 Nodes on bigest cloud Number of SU s 0 30 60 90 120 0 20 40 60 80 C D A Nodes on bigest cloud Number of SU s B Fig. 7. P erformance of proposed al gorithms. The o ptimization a lgorithm is per formed ther eafter to opti- mize the ch anged ph ysical top ology . The basic rule for the reconfigu ration is wh en a n affected n ode curren tly belo ngs to no clu ster , it takes action to associate with o ne cluster o r start a ne w cluster . In other ca ses, the clu ster heads c oordin ate the changes. Note that after reco nfiguration , gateway no des of affected clu sters may ne ed reconfig uration. The algorithm w o rks as follo ws. From the ne ighbor list, a node, namely the working no de, obtain s a nod e set, which includes all member s o f its one-ho p neighb or clusters and its ho st cluster . It is the target no de set to b e optimized. A connectio n grap h is created based on the target n ode set. T he objective is to co nstruct clusters b ased on a MDS of the graph so th at the n umber o f c lusters in the target node set can be minimized. The MDS is obtained by a heuristic algorithm [26]. The algorithm takes the multiple channels of a node into account. First, a cluster is for med by taking the workin g node as th e cluster head a nd its contro l cha nnel as the master chan nel. The one -hop neighbo rs of the node are assigned to the cluster if th eir con trol chan nels are as same as the m aster ch annel of the w orkin g node. The members of the formed cluster ar e eliminated from the target node set. The remainin g no des are processed as following. As in a Max Degree algorithm [25], a node with max degree on its contro l channe l is chosen to form a cluster with cor respond ing neighbors in or der u ntil all n odes join the n etwork. Finally , the new c luster configu ration comes out with the c luster head s list, the master channels and the members of each cluster . If th e num ber of resultant clusters is smaller tha n the cur rent one, the work ing nod e starts a negotiation process to reco nfigure its surro unding clusters. T o start the negotiation process, th e n ode sends rear range- ment requests to the cluster heads it wants to reconfigur e, in- dicating the gain that can b e obtained fr om th e rearrangement, and the reconfigur ation in struction. T he gain is the total num- ber of clusters being red uced if the rearrangemen t is taken. A cluster head, o nce accepts the requ est, sends an acknowledge to the working n ode. The work ing node n egotiates with the target cluster head s to complete the remaining co nfiguratio n process only after receiving all acknowledges b ack. Otherwise, it cancels the process to avoid increasing the clu ster number by an in complete recon figuration . V I I . P E R F O R M A N C E A N A LY S I S In this section we simply show the perform ance of sw ar m intelligence algorithm by simulation. T he purpose is to gi ve readers a rough id ea on the beha viors of the alg orithm with out going deep ly in to the algo rithm. Simu lations are ru n under different network c ondition s. W e use the standard deviation of the SU numb er distributed on each ch annel to measure the trend of SUs to wards sharing the master chann els. A large standard deviation mean s the sizes of chann el clo uds are not evenly distributed, therefo re more nod es being aggr egated to fe w large clouds. Moreover we use th e size of the lar gest cloud, i.e. , th e n umber of SUs in the largest clou d, to illustra te th e impact of the swarm intelligence on the control cloud formatio n. 20 40 60 80 100 120 140 160 180 200 220 0 10 20 30 40 50 60 70 80 2 PUs, 2 channels 8 PUs, 2 channels 2 PUs, 4 channels 8 PUs, 4 channels 2 PUs, 8 channels 8 PUs, 8 channels Standard deviation Number of SUs Fig. 8. S tandard devia tion as a function of SU number . W e first compar e th e performa nce o f cluster formation with or without the swarm intelligence algorithm. The perfo rmance figures are shown in Fig. 7, in wh ich the sub-figur e A and B show the standard deviation, and the sub-figure C and D show the largest control clou d size. From those figu res, we conclud e that the swarm intelligen ce algorithm f orms larger control clouds, as e x pected. W e then show th e perform ance o f th e swarm intelligence algorithm un der different SU, PU and ch annel settings. Fig . 8 shows the be havior of the alg orithm c orrespon ding to different SU po pulation s. As seen from the figu re, the standard d evi- 9 0 5 10 15 20 25 30 35 40 45 50 0 2 4 6 8 10 12 14 16 30 SUs, 2 channels 60 SUs, 2 channels 30 SUs, 4 channels 60 SUs, 4 channels 30 SUs, 8 channels 60 SUs, 8 channels 8 PUs Standard deviation Time cycle Fig. 9. Dynamic behav ior of a swa rm in tellig ence algorithm in a Cog Mesh netw ork. ation is hig h in all cases, m eaning the algor ithm works as expected. The standard deviation in creases as th e SUs in crease. It implies mor e SUs are agg regated to few comm on master channels. The dynamic behaviors of the algorith m are shown in Fig. 9. The PUs in th is simulation setup change their o perating channels per iodically . As seen from this figure , in b oth cases after the channel fluctuatio n, th e standard deviation tu rns to a high s table v alu e shortly , m eaning the SUs a re aggregated to few co mmon master chan nels quickly . V I I I . C O N C L U S I O N The DSA paradigm br ings n ew o ppor tunities into the wire- less world. Howe ver , the ben efits of the ne w spectrum access paradigm do not come naturally . Th ere are plenty o f challenges need to be dealt with before releasing its f ull power . In this ar ticle, we intro duce a new con cept of the cognitiv e wireless n etwork, named CogMesh, to meet those challenges. W e id entify that CogMesh shares similarities with conven- tional wir eless ad hoc networks in terms of d istributi ve con trol and self-organizatio n, but at the same tim e there are significa nt difference betwee n them since in Cog Mesh SUs uses oppor- tunistic sp ectrum access. One of the prom inent challenge s in CogMesh is the contr ol c hannel prob lem. W e explain the problem in detail, and propo se corr espondin g so lutions. Inspired from the collective behaviors in the biolog y world, we introduce a swarm intelligence algorithm to form co ntrol clouds in CogMesh . On the top of the co ntrol cloud s, we propo se the cluster based netw ork formation solution to further consolidate the spectr um manag ement in the DSA scenario. Cluster based appro ach simplifies the r esource m anagem ent of multi-ho p networking in dynam ic radio en vir onments. The solutions pr ovided in this article aim to provid e some insights on the co ntrol pro blem in CWNs and show how a bio-inspir ed appro ach can be applied in CWNs. Actually the concept of CWNs open s up a n ew research area that covers almost every aspect of communication s. I nterdisciplina ry re- search is highly beneficial in this ne w area. It therefore calls for the con tributions fro m researcher s with di verse backg round . R E F E R E N C E S [1] Federal Communication s Co mmission, “Spectrum Polic y T aks Force, ” Rep. E T Doc ket No. 02-135, Nove mber , 2002. [2] S. Haykin, “Cog niti ve Radio : Brain-Empo wered Wir eless Communica - tions, ” Selec ted Ar eas i n Communications, IEE E J ournal on , vol . 23, no. 2, pp. 201–220 , 2005. [3] I. Akyi ldiz, W . Lee, M. V uran, and S. Mohanty , “Nex t Genera- tion/Dyn amic Spectrum Access/Co gniti ve Radio W ireless Netw orks: A Surve y , ” Comput er Network s , vol. 50, no. 13, pp. 2127–215 9, 2006. [4] R. Thomas, D. Friend, L. Dasilv a, and A. Mackenz ie, “Cognit iv e Networ ks: Adaptation and Learning to Achie ve End-to-End Performance Objecti ves, ” Communications Ma gazine , IEEE , vol. 44, no. 12, pp. 51– 57, 2006. [5] D. Bourse, K. El-Khazen, A. Dela utre, T . Wieb ke, M. Dillinger , J. Brak- ensiek, K. Moessne r , G. Vi vier , and N. Alonistioti , “European Resear ch in End-to-End Reconfigurabi lity , ” Pr oc. of IST Mobile & W irele ss Communicat ions Summit, L yon, F rance , pp. 27–30, June 2004. [6] Q. Zhao and B. Sadler , “A Survey of Dynamic Spectrum Access, ” Signal Pr ocessing Magazine , IEEE , v ol. 24, no. 3, pp. 79–89, 2007. [7] T . Chen, H. Zhang, G. Maggio, and I. Chlamtac, “T opology Mana gement in CogMesh: A Cluster-base d Cogniti ve Radio Mesh Network, ” in IEEE CogNet W orkshop in co njunctio n with IEE E ICC 2007 , Glasgo w , Scotlan d, June 2007. [8] T . Chen, H . Zhang , M. Katz, and Z. Zhou, “Swarm Inte lligen ce Based Dynamic Contro l Channel Assignment in CogMesh, ” in IEEE CoCoNet W orkshop in conjun ction with IEEE ICC 2008 , Be ijing, China , May 2008. [9] V . Ka wadia and P . Kumar , “Principle s and Protocols for Power C ontrol in W ireless Ad Hoc Netw orks, ” Selec ted Areas in Communication s, IEEE J ournal on , vol. 23, no. 1, pp. 76–88, 2005. [10] J. So and N. V aidya, “Multi- Channel MAC for Ad H oc Netwo rks: Handling Multi-Chann el Hidden T erminal s Using a Single Transcei ver, ” MobiHoc 2004, T ok yo, J apan , pp. 222–233, May 2004. [11] C. T ang and J. Garcia-Luna-Ace ves, “Hop-R eserv ation Multipl e Acce ss (HRMA) for Ad hoc Networks, ” IEEE INFOCOM 1999, New Y ork, USA , Marc h 1999. [12] S. Chaudhuri, R. K umar , and A. Saha, “A MA C protocol for Multi fre - quenc y P hysical l ayer, ” T echnical Report, January 2003, Ric e Univ ersity , Houston, TX, T ech. Re p. [13] A. Nasipuri , J. Zhuang, and S. Das, “A Multi -channel CSMA MA C Protocol for Multihop Wir eless Networks, ” IE EE WCNC 1999, New Orleans, USA , September 1999. [14] S. W u, C. Lin, Y . Tseng, and J. Sheu, “A Ne w Multi-c hannel MA C Protocol with On- demand Channel Assignment for Multi-hop Mo bile Ad Hoc Networks, ” I-SP AN 2000, Dallas/Richar dson, USA , December 2000. [15] X. Jing and D. Raych audhuri, “Spect rum Co-exi stence of IEEE 802 .11 b and 802.16 a Netwo rks Using the CSCC Etiquette Protocol, ” DySP AN 2005, Balti more , USA , Nov ember 2005. [16] P . Cordier , D. Bourse, K. Moessner et al. , “E2R Cogniti ve Pilot Channel concep t, ” P r oc. of IST Mobile & W irel ess Communications Summit 2006, Myconos, Gr eece , June 2006. [17] J. Zhao, H. Z heng, and G. Y ang, “Distrib uted Coordinati on In D ynamic Spectrum All ocation Netwo rks, ” IEEE DySP AN 2005, B altimor e, USA, Nov , 2005. [18] K. Bian and J. Park , “Se gment-Based Channe l As signment in Cogniti ve Radio Ad Hoc Netw orks, ” CR OWNC OM 2007, Orlando, USA, Au gust , 2007. [19] E. Bonabeau, M. Dorigo, G. T heraula z, and I. NetLibrary , Swarm Intell igenc e fr om Natural to Artificia l Systems . Oxford Univ ersity Press, 1999. [20] R. S choonde rwoerd, O. Hollan d, J. Brute n, and L. Rothkrantz , “Ant- Based Load Balancing in T elecommunica tions N etwo rks, ” Adaptive Behavior , vol. 5, no. 2, p. 169, 1997. [21] G. Di Caro and M. Dorigo, “AntNet : Distributed Stigmer getic Control for C ommunication s Networks, ” Jou rnal of Artificial Intell igen ce Re- sear ch , vol. 9, no. 2, pp. 317–3 65, 1998. [22] T . Renk, C. Kloeck, D. Burgkhardt , F . Jondral , D. Grandb laise, S. Gault, and J . Duna t, “Bio-Inspi red Algorithms for Dynamic Resource Alloc a- tion in Cogniti ve Wire less Networks, ” CRO WNCOM 2007, Orlando, USA, Aug ust , 200 7. [23] C. Lin and M. Gerla , “ Adapti ve Cl ustering For Mobile Wirel ess Net- works, ” Selec ted A rea s in Commun ications, IEE E J ournal on , vol. 15, no. 7, pp. 1265–12 75, 1997. 10 [24] A. Amis, R. Prakash, T . V uong, and D. Huynh, “Max-Min d-Cluster Formati on In W ireless Ad Hoc Networks, ” IEEE INFOCOM 2000, T el- A viv , Israel , March 2000. [25] L. Bao and J. Garcia-Luna-Ace ves, “T opology Managemen t in Ad Hoc Networ ks, ” MobiHoc 2003, Annapolis, USA , pp. 129–140, June 2003. [26] V . Chvata l, “A Greedy Heuristi c For T he Set-Cov ering Problem, ” Mathemat ics of Operation s Researc h , v ol. 4, no. 3, pp. 233–235, 1979.