Research Article Dynamical Spectrum Sharing and Medium Access Control for Heterogeneous Cognitive Radio Networks

Transcription

1 Hndaw Publshng Corporaton Internatonal Journal of Dstrbuted Sensor Networs Volume 16, Artcle ID 659, 15 pages Research Artcle Dynamcal Spectrum Sharng and Medum Access Control for Heterogeneous Cogntve Rado Networs Ahmed Mohamedou, 1 Aduwat Sal, 1 Borhanuddn Al, 1 and Mohamed Othman 1 Department of Computer and Communcaton Systems Engneerng, Unverst Putra Malaysa, 44 Serdang, Selangor, Malaysa Department of Communcaton Technology and Networs, Unverst Putra Malaysa, 44 Serdang, Selangor, Malaysa Correspondence should be addressed to Ahmed Mohamedou; ahmed.mohamedou@outloo.com Receved 18 August 15; Revsed 1 March 16; Accepted Aprl 16 Academc Edtor: Jaml Y. Khan Copyrght 16 Ahmed Mohamedou et al. Ths s an open access artcle dstrbuted under the Creatve Commons Attrbuton Lcense, whch permts unrestrcted use, dstrbuton, and reproducton n any medum, provded the orgnal wor s properly cted. Ths paper tacles the ssue of spectrum sharng and medum access control among heterogeneous secondary users. Two solutons are proposed n ths paper. The frst soluton can be used n centralzed fashon where a central entty exsts whch decdes transmsson power for all secondary users. Ths soluton tres to mnmze the tme requred by secondary users to clear ther queues. The second soluton assumes the autonomy of secondary users where the decson to update transmsson power s dstrbuted among users. Dynamcal system approach s used to model system behavor. The trajectory of nterference nose level suffered by secondary users s used to update transmsson power at the begnnng of each tme frame based on the proposed dynamc power assgnment rule. Ths rule couples the responses of all secondary users n a way whch smplfes future nterference nose forecastng. A forecastng engne based on deep neural networ s proposed. Ths engne gves secondary users the ablty to acqure useful nowledge from surroundng wreless envronment. As a result, better transmsson power allocaton s acheved. Evaluaton experments have confrmed that adoptng deep neural networ can mprove the performance by 46% on average. All of the proposed solutons have acheved an outstandng performance. 1. Introducton Cogntve rado s a major step n the evoluton of wreless communcatons [1]. It provdes wreless systems wth the ablty to adapt to the changes n the envronment and to ncrease ts utlzaton. By observng ther surroundngs, wreless systems should update ther communcaton parameters n a way that ncreases ther performance. The most mportant ssue n system utlzaton s spectrum usage whch canbeconsderedasthemanmotvatonbehndcogntve rado paradgm. The avalable wreless spectrum s a very scarce resource and hghly underutlzed []. Cogntve rado systems (secondary users) should tae the opportunty and they should try to use the underutlzed spectrum to mprove ther performance. Ths s easer sad than done. A serous problem wll arse f naïve polcy of unorganzed attempts by multple cogntve systems to utlze the same spectrum s adopted by these systems. Ths naïve way of utlzng Idle spectrum wll lead to chaos and hgh compettve stuatons among cogntve rado systems. Also, other resources wll be wasted such as power and computaton. Hence, secondary users should fgure out a way to share the underutlzed spectrum wthout addng more problems to each other. Spectrum sharng s a ey challenge n cogntve rado technologes. There are two types of spectrum sharng. The frst s prmary-secondary sharng where the prmary users allow secondary users to use ther spectrum under some condtons. Ths frst type s dvded nto two categores whch are prmary-secondary underlay sharng and prmary-secondary overlay sharng []. In underlay paradgm, the secondary user operates on the prmary user channel wth promse of not causng nterference hgher than some threshold. In overlay paradgm, the prmary user allows the secondary user to usetschannelnreturnforsomeservcesuchasrelayng. The second type of spectrum sharng s secondary-secondary sharng where secondary users utlze the CR spectrum based on some sharng polcy.

2 Internatonal Journal of Dstrbuted Sensor Networs 1.1. Related Wors. Numerous wors n lteratures tred to tacle the frst type (prmary-secondary) of spectrum sharng ssue [4 4]. For example, convex optmzaton was used by [4] to show the possblty of havng both of the prmary user and secondary user operatng n the same spectrum. Therapproachsbasedonusngmultpleantennasforthe secondary user transmsson so that spatal multplexng can be used to guarantee low nterference constrants at the prmary user. In addton, ther approach s able to operate on multple channels at the same tme whch means that ther approach s capable of havng hgh level of reconfgurablty n spatal, tme, and frequency domans. Another example of prmary-secondary sharng s provded by [5] that followed a game theory approach based on olgopoly maret strateges. In ther approach, prmary users are dynamcally evaluatng ther prces n whch they offer spectrum access to secondary users. The cost for prmary users s measured n terms of qualty of servce degradaton of prmary transmsson. Nyato and Hossan analyss of ther approach found that Nash equlbrum of ther game s not optmal from prmary users perspectve. Therefore, they forced prmary users to choose strateges whch acheve global optmalty and they provded a mechansm to punsh devated prmary users (.e., prmary users who do not follow globally optmal strateges) to ensure stablty. However, ther approach s not fully robust because t requres that all prmary users be aware of the punshment mechansm. A smlar maret-based approach s adopted by [8] where aucton theory s used. The proposed mechansm s based on secondary users submttng bds for auctoned spectrum by prmary users. Then, the prmary user wll choose the best bd whch acheves stable equlbrum. To generalze theproposedsolutontomultpleprmaryusers,several condtons have to be satsfed to acheve stablty whch s themanweanessofthsapproach.examplesofcondtons are the fact that prmary users always consder other wreless statons nterest; f they do not, there should be an entty that punshes uncooperatve prmary users. In addton, every prmary user s aware of all other prmary users. If t s not, the proposed soluton wll not be stable. The wor n [16] used smulated annealng to form ther soluton. It s based on prmary users allowng secondary users to operate n the spectrum whle obeyng nterference cap. In general ther approach had good performance. However, t requres a centralzed global optmzaton where smulated annealng s appled. Reference [15] followed a dfferent path to tacle spectrum sharng ssue. They assumed that secondary networs have relay nodes whch can be used to ease cogntve rado operatons. Each secondary user s supposed to select transmsson through a relay node whch guarantees nterference constrant of prmary users. Ther soluton showed very good performance. Nevertheless, t requres an exstng supportve relay nfrastructure to be effectve. Lterature lacs the needed research to tacle the second type (secondary-secondary) of spectrum sharng. Most of the references wored on the second-type assumed homogenety of secondary users and Idle prmary channels [5 ]. For nstance, [6] proposed a mechansm whch dstrbutes channels among secondary users n a way that mnmzes nterference among them. They followed game theoretc approach to develop ther soluton. To mprove the performance, Pgouvan taxaton was ncorporated nto the proposed mechansm. On the other hand, [] was concerned wth spectrum sharng and MAC n wde secondary networs where each secondary user can access a dfferent subset of avalable channels. They formulated such system as a mxednteger nonlnear problem whch s very hard to solve. As a result, lnearzaton and relaxaton technques were used to ease the problem. The proposed soluton s teratve n nature and t was able to acheve near-optmal performance accordng to ther problem formulaton. Some vew wors tred to handle heterogeneous cogntve rado settngs [4 41]. An example of such wors s proposed by [4]. Here, the authors suggested a two-phase approach. The frst phase dstrbutes channels among secondary users. Then, the second phase assgns transmsson power on these channels. Ther approach was able to perform well. However, t requres a centralzed scheduler whch ntroduces communcaton and cooperaton overhead. Smlarly, [5] tred to tacle heterogeneous spectrum sharng and MAC by proposng cross layer approach. It s based on classfyng channels probablstcally where Hungaran algorthm s used n the next step to schedule channels among secondary users. The proposed soluton was able to outperform greedy technques. 1.. Motvaton and Contrbutons. Most of the dscussed related wors assumed some sort of cooperaton among secondary users. These facts motvated the authors of ths paper to nvestgate the second type of spectrum sharng and medum access control n harsher wreless envronment. In ths envronment, secondary users are assumed to be totally heterogeneous wthout any cooperaton among them. In addton, prmary user channels are assumed to be heterogeneous as well, where they have dfferent bandwdth and channel characterstcs (.e., fadng and shadowng). Ths paper tres to contrbute to the exstng cogntve rado lterature by delverng three man contrbutons: (1) The concept of Spectrum-Tme Dualty s hghlghted whch states that t s benefcal for secondary users to reduce ther level of competton wth other secondary users by gvng spectrum n return for tme and vce versa. The dea depends on the fact that secondary users can be Idle f they transmtted all ther data. For a secondary user whch stll has data to send, ts jobs get easer after the others become Idle. Based on ths concept, centralzed transmsson power mechansm s ntroduced where the man objectvestoclearthesecondaryusersqueuesas soon as possble so that the competton for other secondary users s reduced. () Dstrbuted transmsson power rule s developed so that secondary users can assgn ther transmsson power accordng to closed dynamcal system evoluton. By usng ths rule, the trajectory evoluton of many system parameters can be predcted

3 Internatonal Journal of Dstrbuted Sensor Networs snce t follows strct dynamcs. Examples of system parameters are queue baclogs of secondary users, nterference nose caused by secondary users, and the actual transmsson power of all secondary users. ()Advancedmethodsproposedtoemploydeepneural networ (DNN) [4] whch s a very promsng machne learnng technque. Ths method utlzes trajectory evoluton of nterference nose levels to forecast future nose levels. As a result, performance of transmsson power rule s greatly enhanced. By employng DNN n cogntve rado technology, real cognton capabltes can be realzed. Secondary users wll be able to generate decsons wthout the nterventon of system desgner. To the best of our nowledge, ths s the frst attempt n lterature to employ DNN n cogntve rado networs. Note that the proposed solutons are desgned for cogntve rado technology snce t s the man technology of next generaton wreless communcatons. Nevertheless, these solutons can be easly mplemented n conventonal wreless technologes. Ths paper s organzed as follows: Secton defnes the adopted system model to develop the proposed solutons; Secton ntroduces the proposed centralzed spectrum sharng and access mechansm; Secton 4 develops dstrbuted transmsson power rule and nose forecastng engne as the second proposed soluton; ntensve evaluaton experments are dscussed n Secton 5 and the concluson s delvered as Secton 6.. System Model In ths secton, the general formulaton of the spectrum sharng problem n cogntve rado networs s ntroduced. Then, ths formulaton wll be modfed to ncorporate mportant aspectswhchcanbeusedtoeasetheproposedsoluton development..1. General Spectrum Sharng Problem. Consder a set of several secondary users pars (transmtter and recever) U= {u 1,u,...,u U } and a set of prmary user channels C = {c 1,c,...,c C }. Each prmary user channel has a state of beng ether Idle or Busy dependng on prmary user operatons. Every u perceves channels states accordng to ts sensng capablty where t depends on probabltes of detecton and false alarm. Let us tae O as the matrx whch represents how each user perceves channel occupancy states. Elements n O (o j denotes the jth channel state accordng to the th user) are ether f the channel state s Idle or 1 f t s Busy. For smplcty, o j =only f the actual channel state s Idle and the user thns channel state s Idle as well. Otherwse, o j =1 f the actual channel state s Busy or f the user thns that channel state s Busy even f t s not. Also, o j =1f the user thns that channel state s Idle whletheactualstatesbusy snce any transmsson attempt on the channel wll not be successful because of the hgh nterference from the prmary user. Let o denotetheactualstateofchannelc. The transmtter n every secondary user par u uses p as transmsson power on channel c,where p p max. Assume that prmary user transmsson power on channel c s much larger than the maxmum transmsson power of secondary users (p p max ). The channel gan from transmtter n secondary user par u j to the recever n secondary user par u on channel c s denoted by g j, whle the channel gan from prmary user transmtter to the recever n secondary user par u s denoted as g. Therefore, the Sgnal-to-Interference-Nose-Rato (SINR) experenced bytherecevernsecondaryuserparu on channel c s calculated as follows: γ (1 o = )g p U j= (1 o j)gj p j +(1 o ) g p +n, (1) where n represents the thermal nose at the recever. Every γ has to be larger than some threshold (γ γ thr )forths SINR to be useful. The threshold γ thr depends on the recever senstvty, codng, and error correcton mechansms adopted by u. Also, ths threshold can be consdered as a proxy for the mnmum qualty of servce. Most wors n lterature assume homogenety of channels bandwdth. However, future systems have to deal wth heterogeneous channels. Therefore, let us assume that every channel c has unque a bandwdth BW. The maxmum achevable data rate of user u on channel c canbemeasuredbyshannon formula r = BW log (1 + γ ). The general spectrum sharng problem s formulated as follows: max p U C r s.t.: γ γ thr p p max. Keep n mnd that ths formulaton assumes that every secondary user has somethng to transmt all the tme. In other words, secondary users do not have Idle perods. Such assumpton may be vald n controlled envronment of homogeneous secondary users (.e., wreless sensor networs). However, t s not vald for heterogeneous secondary users envronment (.e., mx of multple femtocells and WLANs). Therefore, we ntroduce another formulaton n the next subsecton... Spectrum-Tme Dualty. One of the most mportant observatons n heterogeneous secondary users envronment s that every user has ts own pacet sze (s ). As a result, dependng on the acheved r, there s a mnmum amount of tme that must be guaranteed for successful pacet transmsson on channel c (.e., t s /r, where pacet sze s n bts). Also, each u has several pacets watng for transmsson n ts queue (.e., Q pacets). The total tme needed to transmt all pacets n the queue can be approxmated by () Q T s C [BW log (1 + γ )]. ()

4 4 Internatonal Journal of Dstrbuted Sensor Networs Equaton () does not guarantee mnmum t.hence,t s only an approxmaton. For the tme beng, we wll use ths approxmaton to reformulate our problem. The man dea n the new formulaton s to mnmze the tme requred by u to clear ts queue. As a result, user u wll be Idle for some perod of tme. Consequently, the competton wll be reduced on the avalable channels for other secondary users. So far, we consdered γ as a fxed value. Conversely, γ s a functon of tme (.e., γ (t)) becausebothofp and g j change over tme. Whle p s a controllable parameter, g j s a random varable. Let us assume that channel gan s a statonary process, where E[g j ] represents ts fxed expectaton. Hence, the change n γ (t) s only due to change n p (.e., p (t)). Keep n mnd that γ (t) not only depends on the transmsson power of user u, but also depends on the transmsson power of other secondary users whch are operatng on the same channel. Accordngly, the maxmum amount of data that can be transferred over channel c n nterval [, τ] s r τ = BW log (1 + γ (t))dt. (4) If there s a central coordnator whch assgns the transmsson power to all users durng [, τ], the expected SINR for all users wll be fxed durng ths nterval γ (t) = γ snce only central coordnator can change the transmsson power for users. As a result r τ = BW log (1 + γ )dt. (5) Equaton (5) s for centralzed control, whle the prevous s for dstrbuted control. As a startng pont, we wll focus on the centralzed verson. Themaxmumnumberofpacetsthatcanbetransferred n nterval [, τ] s equal to [ [ C [ τ BW log (1 + γ )dt] ]. (6) s ] The next step s calculatng how user s queues wll evolve overtme.thearrvalrateofpacetsatuseru s denoted by λ pacets/second. The total number of pacets receved durng nterval [, τ] at user u s τ λ dt.letq () and Q (τ) denote the number of pacets watng n user u queue at the begnnngandattheendofnterval[, τ],respectvely.hence where τ ΔQ = Q (τ) = max [Q +ΔQ,], (7) λ dt [ [ C [ τ BW log (1 + γ )dt] ]. (8) s ] Assumng that the arrval rate of all users s a statonary process where E[λ ] represents the expectaton for user u, the stablty of all systems queues wll depend solely on ΔQ. If ΔQ >, u has more pacets n ts queue at the end of [, τ], whlefδq <,thequeuesclearngupanduser u has reduced the number of pacets n ts queue. Havng ΔQ =ndcates no change n the queue sze. Note that there s a relatonshp between the achevable data rate r = C r and ΔQ.Asr ncreases, ΔQ decreases and vce versa. Also, f ΔQ >Q,thenτ T >whchmeans that fracton of nterval [, τ] of user u wll be Idle. Ths fracton s equvalent to (τ T )/τ. Other secondary users wll beneft from the absence of Idle u due to reduced nterference whch was caused by u transmsson. However, the transmsson power assgnment of the remanng secondary users durng the absence of u s not optmal and can be mproved. Therefore, τ hastobechosentobeequaltothemnmumt among all users. Ths way, new power allocaton can tae place where the absence of u (wth the mnmum T ) s consdered. Thus, the new spectrum sharng problem s formulated based on clearng queues as soon as possble: mn p,τ U Q (τ) s.t.: γ γ thr p p max τ T. The am of the last constrant s to guarantee that no user s Idle durng the nterval [, τ] whch results n optmal power allocaton for all users durng the whole nterval.. Centralzed Medum Sharng and Access The man purpose of ths secton s to develop a general approach to allocate transmsson power n centralzed settng. However, such approach cannot be used n heterogeneous secondary users envronment. Therefore, the next secton wll develop a dstrbuted mechansm to acheve the same goal. The fndngs of ths secton can be easly mplemented n classcal wreless technologes where centralzed control entty s assumed. Frst, let us call the nterval [, τ] the decson tme frame whch means that durng ths tme frame the transmsson powers of all users are fxed. As a result, the tme s slotted nto frames where each frame has a perod length of τ n (.e., n represents the frame ndex accordng to the past and future tme frames). The perods lengths of tme frames are not assumedtobeequal.byfxngthetransmssonpowerdurng thetmeframe,δq at the end of tme frame n can be wrtten as follows: ΔQ = τ n [s s E [λ ] = τ n [s s E [λ ] C C [BW log (1 + γ )]] (9) r ]. (1) The value nsde the bracets can be controlled by only manpulatng the transmsson power levels. To hghlght

5 Internatonal Journal of Dstrbuted Sensor Networs 5 the relatonshp between ΔQ and any transmsson power manpulaton, γ as experenced durng the new tme frame canberewrttenas where γ +Δγ p =(1 o ) E [g ]p Δp =(1 o ) E [g ]Δp η = I +n I = ΔI = U j= U j= = p + Δp η + ΔI, (11) (1 o j ) E [g j ]p j (1 o j ) E [g j ]Δp j. (1) The update of transmsson power s represented by Δp whch s decded at the begnnng of the tme frame. Hence, the new transmsson power durng the new frame s equal to the transmsson power of the prevous frame plus Δp. The same goes for SINR level and nterference where ΔI represents the change of experenced nterference and Δγ represents the change of the experenced SINR level n comparson to the prevous tme frame. The next step s to calculate the change n data rate (Δr ). Basedontheprevoussecton,r = BW log (1 + γ );hence r +Δr = BW log (1 + γ +Δγ ). After several algebrac manpulatons, we fnd out that Δr = BW log (1 + Δγ ). (1) 1+γ Let us wrte both of γ and Δγ n terms of transmsson power and nterference: γ Δγ = p η (14) = p + Δp η + ΔI p η = η ( p + Δp ) p ( η + ΔI ). η ( η + ΔI ) (15) By substtutng both of (14) and (15) n (1) and performng numerous algebrac operatons, we have the followng equaton: Δr = BW log ( η [ p + Δp +1]). (16) p + η η + ΔI Equaton (16) gves us a drect ln between the change n data rate and the update of transmsson power. It can be shown that the relatonshp among transmsson power updates of all users s governed by (1 o ) E [g ]Δp [ (Δr /BW ) p U ] (1 o η j ) E [g j ]Δp j j= =( p + η )( (Δr /BW ) 1). (17) Note that (17) can be wrtten n matrx format for every th channel as C Δp = B,whereΔp s a vector of all users transmsson power updates, B s a vector where the th element s the left hand sde of (17), and C s the coeffcent matrx where ts elements are [C ] j = { (1 o ) E [g j ], =j { { (Δr /BW) γ (1 o j ) E [g j ], =j. (18) By usng the nvertble matrx theorem [4] whch states that matrx A s nvertble f and only f Ax = b has one soluton, snce we derved C from unque soluton (17), then C s nvertble. As a result, the requred transmsson power updates to acheve data rate change Δr basedonthecurrent transmsson power p arecalculatedasfollows: Δp = C 1 B. (19) Based on ths fndng, we reformulate the optmzaton problem n (9) by usng Δr wth τ n as the controllng varables nstead of p.notethatthechangenthelater varable (p ) can be expressed through the change n data rate..1. Reformulaton of Power Allocaton Problem. At the begnnng, we need to revst optmzaton problem (9) constrants. The frst constrant (γ γ thr ) ensures that data rate has the mnmum acceptable bt error rate. In the new tme frame, γ +Δγ γ thr p + Δp / η + ΔI γ thr.snce η + ΔI s always postve, we can rewrte the frst constrant for the th user on the th channel as Δp +γ thr ΔI +γ thr η p. Hence, the frst constrant for the th channel n matrx notaton s E Δp + D, where bt error rate coeffcent matrx (E ) has elements as follows: [E ] j = { (1 o ) E [g j ], =j { () { γ thr (1 o j ) E [g j ], =j. The vector D th element s γ thr η p.now,wecan state the frst constrant n terms of change of data rate nstead of transmsson power by substtutng (19) as follows: E C 1 B + D. (1) Smlarly, the second constrant of optmzaton problem ( p p max ) can be wrtten n terms of data rate change as follows: p + C 1 B p max 1. ()

6 6 Internatonal Journal of Dstrbuted Sensor Networs The thrd constrant needs to be changed to an equvalent constrant snce t depends only on τ n.notethatthemax operator n (8) mplctly states that Q +ΔQ.Bothof these nequaltes (Q +ΔQ and the thrd constrant τ n T ) are equvalent snce they guarantee that no channel s left underutlzed because some user occuped t and has no more data to transmt. Therefore, after substtuton of (1), we can reformulate the thrd constrant as follows: τ n s T (Δr 1) +τ n s T (r 1) Q τ n E [λ], () where Δr and r are matrces representng the requred change n data rate and the current data rate, respectvely. The th column n these matrces corresponds to the th channel and the th row corresponds to the th user. Expected pacets arrval rate of all users s donated by vector E[λ] and the th element n vector s represents the recprocal of the th user pacet sze (1/s ). The last step s to rewrte the global objectve functon. Note that each user would le to mnmze Q +τ n E[λ ] τ n /s C (r +Δr ) to zero whch means that the user has an empty queue and all of ts pacets were successfully transmtted. To combne the utlty functons of all users, the followng objectve functon s used: mn Δr,τ n U (w [Q +τ n E [λ ] τ C n s (r +Δr )]). (4) Tryng to fnd the optmal Δr and τ n of ths objectve functon subject to the mentoned constrants s trcy due to the nonconvexty of ths optmzaton problem. However, wecaneasetbyfxngτ n to a specfc value snce n realty τ n has to be a dscrete varable. It s well nown that any communcaton protocol has to transmt ts pacets as a whole at once. Therefore, t has to have a mnmum perod of tme to transmt a pacet. Ths mnmum perod depends on the used technology. For example, f the used technology s LTE-A, then τ n canbeonemllsecondwhchsthemnmum frame tme accordng to LTE-A standard [44]. Hereafter, our problemsredefnedasfollows: mn Δr U (w [Q +τ n E [λ ] τ C n s s.t.: E C 1 B + D p + C 1 B p max 1 (r +Δr )]) τ n s T (Δr 1) +τ n s T (r 1) Q τ n E [λ]. (5) Havng our problem n ths form, we can fnd the best possble Δr by usng Interor-Pont method [45]. Ths method s based on the dea of combnng the nequalty constrants wth the objectve functon so that Newton method can be used (please, refer to [45] for elaborated dscusson). Optmzng for Δr needs to be performed only once (at the frst tme frame), gven that the number of secondary users n subsequent tme frames s the same and ther pacet arrval rate process s statonary... Balancng Weght. In the problem of (5), due to the heterogenety of both prmary channels and secondary users, the obtaned soluton may lead to unfar power allocaton among secondary users. Therefore, w s used to guarantee a level of farness among them. Several formulas can be used to calculate w. Ths paper proposes three straghtforward approaches as follows: w = 1 (6) U w = Q U j Q j (7) E [λ w = ] U j E [λ ]. (8) Formula (6) descrbes the default approach where an equal treatment of all users s gven. Formula (7) gves users wth larger queues hgher prorty, whle formula (8) favors users wth faster pacets arrval rate. 4. Dstrbuted Medum Sharng and Access The prevous secton developed an approach to share and access medum among multple secondary users. However, such approach requres a centralzed coordnator to assgn transmsson power level to each secondary user. Ths approach can be mplemented only among cooperatve secondary users. For example, base statons n moble telecommuncaton system (.e., LTE or WMAX) whch belong to the same provder can adopt the centralzed approach snce they can coordnate among themselves easly. Conversely, an envronment wth noncooperatve secondary users wll requre a passve dstrbuted approach due to the lac of coordnaton among secondary users. Heterogeneous secondary users shall decde the approprate transmsson power based solely on the local acqured nformaton regardng the surroundng wreless envronment. Assumng that every user nows ts queue sze, expected pacets arrval rate, and Channel State Informaton (CSI), a relatonshp between the requred transmsson power and tme can be formulated based on the objectve of clearng the queue. In other words, each user would le to clear ts queue wthn a specfc maxmum tme frame (τ max ) to acheve the requred qualty of servce level. Therefore, thetransmssonpowershouldbeassgnedbasedonthe current queue sze, expected arrval rate, and nose so that the queuewllbeclearedbytheendofthetmeframe.noses calculatedbasedoncsi,whleτ max can be computed based on several factors such as delay senstvty of communcaton. Note that usng nose level as a parameter to decde the transmsson power wll lead to couplng all nterferng secondary users snce most of the nose s nterference nose generated by them. As a result, we wll have a dynamcal system whch evolves over tme. Ths dynamcal system s dscrete due to the fact that all dgtal communcaton technologes transmt data n pacets. The dscrete dynamcal system requres global tme frame to be defned whch can

7 Internatonal Journal of Dstrbuted Sensor Networs 7 be trcy n heterogeneous envronment. To smplfy the formulaton of transmsson power assgnment rule, we wll assume that such global tme frame s defned as Δt.Notethat τ s used to represent local tme frame for each secondary user, whle Δt represent the global tme frame of all secondary users. At the end of ths secton, a smple approach s proposed to defne the start and the end of ths global tme frame. Any transmsson power formula should try to honor four lmts whch are maxmum queue sze for the th secondary user (Q max ), the maxmum tme frame length for the th secondary user (τ max ), maxmum allowed transmsson power on the th channel (p max ), and maxmum nose for the th secondary user on the th channel (η max ). The last lmtscalculatedbasedonthefactthatsinrhastobe larger than some threshold (γ γ thr ). By usng p max as the transmsson power, η max = (E[g ]pmax )/γ thr can be consdered as the thrd lmt. Both of Q max and τ max depend on factors controlled by upper layers and applcaton requrements such as latency and delay. The thrd lmt p max s channel specfc whch s assgned by regulators or prmary users Tme Frame Length Assgnment. The frst step n desgnng the transmsson power allocaton rule s to ln the length of tme frame (τ )fortheth secondary user on the th channel wth the current state of the secondary user. Both of the current queue sze and current nose level can be used to represent current state. Ths wor proposes a tme frame length assgnment based on these parameters. By usng both η max and the nose level (η (t Δt)) experenced at the prevous tme frame, a relatonshp between nose and approprate tme frame length s formed as follows: τ (t) = η (t Δt) η max τ max. (9) Formula (9) does not honor maxmum tme frame lmt whch means that τ (t) canbelargerthanτ max.tohandle ths ssue, a modfed formula wll be used: τ (t) = η (t Δt)/η max η (t Δt)/η max +β Q (t) τmax, where β >. () The added parameter β controls the dffculty of assgnng τ max asthecurrenttmeframelength.ifβ s small, t wll be easy to acheve maxmum tme frame length and vce versa. Note that the current queue sze parameter s ntroduced n calculaton to add the effect of ths parameter. By usng the defnton of nose based on (1), the tme frame length can be assgned based on τ (t) = = [ U j η (t Δt) γ thr [ U j η (t Δt) γ thr τ max +β Q (t) E [g ]pmax = E [g j ]p j (t Δt) +n ]γ thr = E [g j ]p j (t Δt) +n ]γ thr τ max +β Q (t) E [g ]pmax. (1) The term (1 o j (t Δt)) s taen out because t wll be used n p j calculaton (Secton 4.). The calculated τ tres to reduce the requred tme to clear the queue when secondary user suffers from large queue or hgh level of nterference. By dong that, the secondary user gets more compettve whch results n hgher level of transmsson power. 4.. Transmsson Power Rules. As mentoned before, the prmary goal for any secondary user s to clear ts queue as soon as possble. Based on ths objectve, a general transmsson power assgnment rule on the th channel can be formed by algebracally manpulatng the followng equaton: BW (Q C +τ E [λ ]) τ BW l BW l s log (1 + γ ) =. () Equaton () s based on (8). The frst term dstrbutes the load (current and new arrvng pacets) among channels based on ther bandwdth. Other approaches to dstrbute theloadcanbeused.however,thsapproachsadopted for smplcty. After some algebrac manpulatons, general transmsson power s calculated as follows: p (t) where =[1 o (t)] U j= [E [g j ]p j (t Δt)]+n E [g ] δ (t), () δ (t) = [(s (Q (t)+e[λ ]))/τ C l BW l ] 1. (4) It s clear that p max lmt s not honored by ths transmsson power rule. A smlar approach used to modfy tme frame length assgnment τ canbeemployedhereaswell. Ths yelds p (t) = = [1 o (t)]η (t Δt) δ (t) p max η (t Δt) δ (t) +α [1 η (t Δt)/η max ] E [g ] ( U j [1 o (t)]( U j= [E [g j ]p j (t Δt)]+n )δ (t) p max = [E [g j ]p [1 ( j (t Δt)]+n )δ (t) +α U j = [E[g j ]p j(t Δt)]+n )/η max ] E [g ], where α >. (5)

8 8 Internatonal Journal of Dstrbuted Sensor Networs [1 η Term α (t Δt)/η max ] s added to ntroduce an exponental effect of current nose level. Whenever nose becomes very large (.e., system s n outage), the transmsson power reduces exponentally. Ths behavor favors secondary users whch are not n outage by reducng the nose caused by secondary users sufferng from outage due to hgh nose level. Also, whenever nose s very low, the transmsson power s prmarly computed based on the current queue sze through τ calculaton (assumng α s chosen properly). Note that when η (t Δt) = η max, transmsson power s computed purelybasedonqueuesze. 4.. Inference Engne. At ths pont, we constructed a model for heterogeneous cogntve rado networs where transmsson power of dfferent secondary users over dfferent channels forms a tghtly coupled dynamcal system. The next step s to develop an nference engne to forecast future nose level so that better values of transmsson power are assgned. Ths paper proposes the use of deep neural networ (DNN) [4] to construct such engne. The capablty of DNN to extract features from any seres of correlated nput data s more strengthened when nputs are comng from a tghtly coupled dynamcal system. Also, the nonlnearty of any dynamcal system can be approxmated by neural networ. These facts advocate for the use of DNN as base for the proposed engne Engne Input. The frst step to develop the proposed nose forecastng mechansm s to defne the approprate nput desgn. Any nput data should be based on local nformaton only snce secondary users are assumed to be heterogeneous wth no actve cooperaton. Inspred by Taen s theorem [46], the nput s defned as a vector of delayed nose values over tme (.e., nput = [η (t Δt), η (t Δt),...,η (t dδt)]).taenshowedthata local mage of dynamcal system evoluton manfold can be constructed based on delayed data from one dmenson of the manfold. The mappng between the orgnal manfold and the constructed local mage s one to one. By utlzng ths fndng n our case, every secondary user can construct an mage of the manfold descrbng total system nose level over tme.thsmanfoldmagecanbeusedasnowledgebaseto forecast future values of nose level. One queston stll needs an answer whch s how many of the past nose values should we use? In other words, what s the best value for d? BasedonTaen stheorem,thenumberofdelayedvalue samplesshouldbemorethantwcethespacedmensonalty whch contans the manfold. In our case, the dynamcal system forms U -dmenson space. Therefore, d= U +1 s the mnmum requred vector sze to guarantee manfold embeddng. Assumng that all nterferng secondary users arencloserange,eachoneofthemcanfgureout U by overhearng each other s dstnct transmsson patterns (.e., modulaton or codng settngs). Fnally, by feedng local mage of manfold values to the DNN, useful features of the manfold can be extracted. These features are used to forecast future nose level DNN Archtecture. Deep neural networ can be dvded nto two man categores: Deep Autoencoder (DAE) [47] and Deep Belef Networ (DBN) [48]. Both of these types use neurons where the actvaton functon s a sgmod functon. They dfferentate n how to nterpret the value of actvaton functon. DAE nterprets actvaton functon value as the neuron output, whle DBN uses ths value as probablty of selectng one of bnary values as the neuron output. Ths paper adopts the frst type as the man archtecture for nference engne due to ts determnstc behavor. Neurons n DNN are arranged n layers. Each layer represents an extracted feature of nput data. The frst layer s called nput layer, whle the last layer s the output layer. Any layer between the nput layer and the output layer s called a hdden layer. Classcal neural networs have one hdden layer because of extreme dffculty n tranng multple hdden layers [49] by usng the conventonal learnng technques. Ths was the case untl 6 when Hnton et al. [48] proposed a technque to tran deep neural networs where several hdden layers are staced over each other. Ths fndng revolutonzed the feld of neural networs and ts applcaton, especally after smlartes of ths tranng technque and how the mammalan bran operates were hghlghted. Stacng several hdden layers allows DNN to extract several features from nput data. To llustrate what we mean by features, let us use object recognton example snce t s easer to vsualze. Imagne the nput data s a stream of gray scale mages. After feedng these mages to DNN, the frst layer wll be able to dstngush dfferent types of edges, whle the second layer wll be able to dstngush dfferent combnatons of edges recognzed by the frst layer. By repeatng ths behavor, the thrd layer wll dfferentate between dfferent object subparts based on the second-layer output and the fourth layer wll use the thrd-layer output to recognze the object type. Ths herarchcal process can be repeated to extract hgher abstracton features dependng on the applcaton objectve. In our case, a smlar approach s used to extract several features of wreless envronment surroundng secondary users. The nose level can be forecasted locally by every secondary user based on these extracted features. In general, we do not really care what the exact nature of these features s as long as they enhance the nference engne forecastng capabltes. Smlarly, n prevous object recognton example, features such as lght ntensty may not have bg mpact on how well DNN can recognze objects. Nonetheless, these typesoffeaturesmayhelpemphaszemoremportantfeatures whch leads to mprovng DNN learnng capabltes. The adopted learnng technque has the ablty to extract any necessary features to acheve learnng objectves [47] Engne Integraton. At the begnnng, the nference engne wll not be able to accurately forecast future nose level because t requres tme to learn the surroundng wreless envronment dynamcs. Therefore, to calculate transmsson power usng (1) and (5), the forecasted nose level should be ntegrated n a way whch taes forecastng error nto consderaton. Let η be the forecasted nose level produced by nference engne for the th secondary user and let η

9 Internatonal Journal of Dstrbuted Sensor Networs 9 be the exponental movng average of the actual experenced nose whch s calculated as follows: η (t) = t Δt η (t jδt) j=1 ω j, (6) where ω can be any postve number larger than one and t represents the dependence of η on past values of η. Instead of usng η (t Δt) n (1) and (5) to calculate transmsson power, η... wll be used whch s a combnaton of the expected nose level n the current tme frame and the exponental movng average of nose level from prevous tme frames:... η (t) = (1 ε) η (t) +εη (t). (7) The weght term ε relates to the error n nose forecastng. Its value should be between zero and one where approachng zero means hgher confdence n nose forecastng accuracy. The adopted formula for ε s e[ η (t) η (t)] / η (t) 1 ε= η (t)] / η (t) +1. (8) e [ η (t) Update nose movng average Sense nose fgure Tran neural networ Inference engne Wreless envronment Combne nose fgures Predct nose fgure Fgure1:Inferenceengnentegraton. Sense spectrum channels Assgn transmsson power It s clear that as [ η (t) η (t)] (.e., forecastng error) decreases, dependence of η... on η ncreases. Another way to loo at ths approach s by magnng that we have two experts. One s totally past orented, whle the other s future orented. The forecastng of both of these experts s added n a way whch favors the expert wth better performance so far. To summarze, the ntegrated transmsson power allocaton rule for the th secondary user on the th channel s Wreless spectrum Avalable channels [1 o p (t) = (t)] η... (t) δ (t) p max... [1 η (t) δ (t) +α... (9) η (t)/η max ] E [g ]. Also, computng δ requres τ whch s redefned as follows: τ (t) = η (t) γ thr η (t) γ thr τ max +β Q (t) E [g ]pmax. (4) Notethatusngthsrulewllresultnatghtlycoupled dynamcal system as well. Consequently, nference engne performance s guaranteed to eep mprovng wth more experence. Both of Fgures 1 and show general llustraton of how the proposed solutons operate and ntegrate Global Tme Frame. So far, t s assumed that a global tme frame s well defned for all secondary users. However, such assumpton s not practcal n a heterogeneous envronment. Therefore, secondary users need to fnd out the global tme frame boundares based on local nformaton. The frst thng to note s that global tme frame boundares have to algn wth local tme frame boundares of all secondary users. Thus, from each secondary user perspectve, the end of local tmeframestheendoftheglobaltmeframeaswell.in addton, each secondary user wll update ts transmsson Queue sze SINR Arrval rate Queue sze SINR Arrval rate Power CMSA DMSA Transmt wth assgned transmsson power Fgure : Illustraton of the proposed system model. power by the end of ts local tme frame. Hence the nose level experenced by other secondary users wll change. Havng the last three facts n mnd, the global tme frame canbeapproxmatedlocallybysecondaryusers.thesmplest approach s to montor nose level varaton. Whenever a sudden change n nose level s notced, the tme of ths change wll be mared as a boundary of global tme frame. Usngthsapproachwllleadtorapdtransmssonpower

10 1 Internatonal Journal of Dstrbuted Sensor Networs updates for all secondary users. As a result, faster convergence of nference engne learnng can be acheved. 5. Evaluaton and Dscusson Several smulaton experments were conducted to evaluate the proposed solutons n ths paper. Parameters n these smulatons were chosen randomly. The reasonng behnd random choce s to test the evaluated solutons performance n the most general way. A unform dstrbuton s used to choose smulaton parameters values. Heterogeneous secondary users and heterogeneous wreless channels were assumed durng the smulaton experments. Secondary users have dfferent arrval rates and dfferent queue szes. Also, they have dfferent pacet szes and dfferent qualty of servce levels whch s represented by the mnmum allowed SINR. On the other hand, wreless channels have dfferent bandwdth and dfferent fadng behavors. The smulaton tme was set to 1 seconds. Tme frame value was chosen randomly between two extreme values. The mnmum value s one mllsecond and the maxmum value s.1 seconds. Each smulaton experment was repeated 1 tmes. The average of smulaton results was taen as the fnal result. Two parameters were chosen as the performance varables (x-axs). These are the number of secondary users and the number of wreless channels. On the other hand, threeperformancemetrcs (y-axs) ware selected to study the proposed solutons. These are queue sze, acheved data rate, and power consumpton. Secondary user pars (transmtter and recever) were dstrbuted randomly where the average dstance between any two pars s 5 m. The average dstance between the transmtter and recever n any secondary user par s m. Parameters such as mnmum SINR threshold, arrval rate, the maxmum queuesze,andchannelbandwdthwereassgnednaway that amplfes ther effect on the general performance. A random value for each one of these parameters was chosen between two extreme levels at the begnnng of each smulaton experment. Then, ths chosen value was multpled by some coeffcent whch depends on secondary user dentty. For example, the thrd secondary user may use 7 as ts coeffcent for maxmum queue sze whch means that the maxmum queue sze for the thrd secondary user s seven tmes larger than the chosen value for ths parameter at the begnnng of smulaton experment. These coeffcents were selected n an ordered fashon whch reflects secondary users and wreless channel ndexes. For nstance, the frst secondary user coeffcent s 1 and the second secondary user coeffcent s. By usng ths coeffcent approach to assgn parameter values, the effect of the parameter wll be apparent n the smulaton results. For example, coeffcents for the maxmum queue sze were set n ascendng order. Therefore, as the number of coexstng secondary users ncreases, bufferng capablty of overall system ncreases as well. Now, we can see f ntroducng more heterogeneous secondary users n terms of queue sze wll mprove the performance. Smlar to maxmum queue sze, mnmum SINR threshold and channel bandwdth were set n ascendng order, where pacets arrval rates were set Table 1: Smulaton parameters. Parameter name Value Tme slot.1.1 seconds Smulaton tme 1 seconds Queue sze 1 1 pacets SINR [, 1] db Pacets arrval rate [1, 1] pacets/s Carrer frequency GHz Channel bandwdth [1, ] Hz Pacet sze 64 bytes Transmsson power.1 watts Thermal nose densty 174 dbm/hz Fadng Raylegh Fadng standard devaton [5, 15] Path loss exponent [, 5] P d [.9,.99] n descendng order. The ntal maxmum queue sze s randomly chosen n nterval [1, 1] pacets. The ntal mnmum SINR threshold s randomly chosen n nterval [, 1] db. The ntal pacets arrval rate s randomly chosen n nterval [1, 1] pacets/s. The ntal channel bandwdthsrandomlychosennnterval[1, ] Hz. Pacet sze for all secondary users s 64 bytes. Ths parameter s dentcal for all secondary users to ncrease the nfluenceofarrvalrateandqueueszeonoverallsystem performance. Maxmum transmsson power for each wreless channel was set to.1 watts. Also, probablty detecton (P d ) for each secondary user, path loss exponent, and fadng standard devaton were chosen randomly n these ntervals [.9,.99], [, 5], and[5, 15], respectvely. Note that the fadng s modeled as Raylegh dstrbuted random varable. Idle and Busy ntervals for each wreless channel were dstrbuted exponentally where the mean (rate parameter) was chosen randomly n nterval [1, 1] mllseconds. The deep neural networ has three layers and movng average factor s.5. Alpha parameter n (9) s set to the movng average of SINR and beta parameter n (4) s set to.5. Note that the proposed soluton n Secton s denoted by CMSA n the fgures. Ths soluton has three versons dependng on weghtng approach defned by (6) (CMSA-E), (7) (CMSA- Q), and (8) (CMSA-L). For Secton 4 soluton, there are two versons. The default verson s based on (5) and (1). It s denoted by DMSA-D n the fgures, whle the second verson s based on deep neural networ ((9) and (4)) whch s denoted by DMSA-N n the fgures. Also, naïve transmsson power assgnment whch uses the maxmum allowed power s smulated as well. Tables 1 and provde summares of smulaton parameters and neural system parameters, respectvely. Smulaton experments were programmed usng MAT- LAB envronment. Several utltes were from LTE system level smulator [5]. Ths smulator has been developed by the Insttute of Telecommuncaton at Venna. It s freely avalable for noncommercal and academc usage. It s very flexble and t has the capablty of provdng more organzed and lesser complex smulaton methods.

11 Internatonal Journal of Dstrbuted Sensor Networs 11 6 Average queue sze (number of channels =1) 6 Average queue sze (number of SU =1) 5 5 Queue sze (pacets) 4 Queue sze (pacets) Number of secondary users Number of channels CMSA-E CMSA-Q CMSA-L (a) DMSA-D DMSA-N MAX-POWER CMSA-E CMSA-Q CMSA-L (b) DMSA-D DMSA-N MAX-POWER Fgure : Queue sze performance. (a) Average queue sze where the number of channels s 1 and the number of secondary users s ncreased. (b)averagequeueszewherethenumberofsecondaryuserss1andthenumberofchannelssncreased. Table:Neuralnetworparameters. Parameter name Value Number of layers Layer szes [, 15, 1] Actvaton functon Sgmod Learnng rate.1 Momentum.9 Regularzaton type L Regularzaton factor.1 Error functon Root mean squared 5.1. Queue Sze Evaluaton. The frst performance metrc to nvestgate s the average queue sze of all secondary users. Keep n mnd that secondary users have heterogeneous bufferng capabltes. Fgure (a) shows how the average queue sze for DMSA-D, DMSA-N, and MAX-POWER ncreases as the number of coexstng secondary users ncreases.thsbehavorsduetothefactthatntroducngmore secondary users ncreases the competton over the fxed avalable channels whch leads to lower data rate for all users. Furthermore, addng more users wth hgher maxmum queue sze as explaned n the prevous secton allows the average queue sze to ncrease. For MAX-POWER, the average queue sze ncreases almost lnearly, whle both of the DMSA versons have much better performance. It s clear that adoptng deep learnng approach has mproved the performance notceably. On the other hand, CMSA solutons have acheved the best performance. The average queue sze f these solutons were used s very close to zero especally when the number of avalable channels s larger than the number of secondary users as nferred from Fgure (b). In the latter (Fgure (b)), dfferent versons of CMSA have dfferent performance when thenumberofchannelssandthenumberofsecondary users s 1. Here, the Interor-Pont method was not able to fnd feasble transmsson power assgnment that clears all secondary user s queues. CMSA-Q had the best performance n ths harsh envronment snce ts optmzaton depends on thecurrentqueueszeofsecondaryusers(7).infgure(b), theaveragequeueszesdecreasngasthenumberofavalable channels ncreases. The reducton rate of MAX-POWER s very small compared to other solutons. Note that, n Fgures (a) and (b), DMSA-N performance approaches DMSA- D performance as overall system complexty ncreases. The acheved mprovement for DMSA-D through the use of DMSA-N shrns as the number of secondary users and channels ncreases. One way to eep DMSA-N mprovement from shrnng s by usng deeper archtecture. However, deeper archtecture requres more computaton power and tranng tme. In addton, DMSA-N s able to explot ncreased system complexty to fnd better transmsson power assgnment. However, as complexty passes some threshold, DMSA-N performance starts degradng smlar to other solutons. 5.. Data Rate Evaluaton. Data rate s measured for secondary users only f they are not n outage. Otherwse, t s assumed as zero. From wreless channel perspectve, the actual experenced data rate s used n the fgures. Farness evaluatons of the proposed solutons were conducted by usng Jan s ndex [51]. However, due to the lmted avalable space for ths paper, bref dscusson regardng farness evaluatons wll be provded wthout farness fgures. Average

12 1 Internatonal Journal of Dstrbuted Sensor Networs Average data rate (number of channels =1).5 Average data rate (number of SU =1) Data rate (Mbps) Data rate (Mbps) Number of secondary users Number of channels CMSA-E CMSA-Q CMSA-L DMSA-D DMSA-N MAX-POWER CMSA-E CMSA-Q CMSA-L DMSA-D DMSA-N MAX-POWER (a) (b) Fgure 4: Data rate performance. (a) Average data rate where the number of channels s 1 and the number of secondary users s ncreased. (b) Average data rate where the number of secondary users s 1 and the number of channels s ncreased. data rate acheved by CMSA-L s the hghest as seen n Fgures 4(a) and 4(b). However, the farness performance of CMSA- L s the lowest. It can be concluded that optmzng based on the arrval rate of secondary users acheves the hghest performance n terms of average data rate. CMSA-Q comes second after CMSA-L. It has better farness performance than CMSA-L due to ts dependency on the current queue sze of secondary users. Both of DMSA solutons acheve very hgh performance n terms of farness. DMSA-N has very close performance to CMSA-E n terms of average data rate when the system s smple. It s reasonable to hypothesze that ncreasng nference engne capablty by usng deeper archtecture wll reduce the gap between DMSA-N and CMSA-E performance n terms of average data rate. Performance of MAX-POWER s the worst. Secondary users were n outage durng most of Fgure 4(a) experments when MAX-POWER was used. 5.. Power Consumpton Evaluaton. The average power consumpton for MAX-POWER s fxed for all experments and t solely depends on the number of wreless channels. As ths number ncreases, the average power consumpton lnearly ncreases as well wth slope of one. Smlarly, the farness for MAX-POWER s fxed at the hghest possble value n all experments snce the same transmsson power s used by all users on all channels. For DMSA solutons, gradually ncreasng power consumpton s notced as well. However, therateofthspowerconsumptongrowthsshrnng as the number of secondary users and wreless channels ncreases. In Fgure 5(a), ths growth rate s decreasng due to the ncreased competton n the system as a result of the ncreased number of secondary users. Hgher competton forces secondary users to reduce ther transmsson power so that the nterference suffered by other secondary users s mtgated. The behavor observed n Fgure 5(a) s a testament for the proposed dynamcal system ablty to nfer the nterference and load states of the surroundng secondary users. Such nference ablty allows secondary users to be more reasonable when assgnng transmsson power. Performance n Fgure 5(b) states that as the number of wreless channels ncreases, the need for aggressve competton through hgher transmsson power decreases. Such behavor can be greatly mproved by usng deep learnng approach. Farness for DMSA soluton s very hgh compared to CMSA solutons. As the number of secondary users and wreless channels ncreases, DMSA farness decreases. Agan, such behavor s very desrable snce t affrms that DMSA solutons were able to ntellgently treat every secondary user dfferently based on ther dstnct stuatons. The average power consumpton of CMSA solutons s very tny compared to other solutons. To show the magntude and the general behavor of CMSA solutons n terms of average power consumpton, logarthmc scale was used as depcted n Fgures 5(c) and 5(d). The Interor-Pont method was able to fnd very tny transmsson power levels whch acheve hgh SINR values for secondary users. It seems that the thermal nosewasthereasonbehndeepngtransmssonpowerat these levels rather than reducng them more. Keep n mnd that SINR s a rato. It does not depend on the actual values of the numerator and denomnator as long as the rato between them s the same. Therefore, mathematcally speang, lower values of transmsson power can be found f thermal nose was not consdered whch s not realstc. Both of CMSA-Q and CMSA-L average transmsson power levels are decreasng as the number of secondary users

13 Internatonal Journal of Dstrbuted Sensor Networs 1 1 Average power consumpton (number of channels =1) 1 Average power consumpton (number of SU =1).9.9 Power consumpton (watt) Power consumpton (watt) Number of secondary users Number of channels CMSA-E CMSA-Q CMSA-L DMSA-D DMSA-N MAX-POWER CMSA-E CMSA-Q CMSA-L DMSA-D DMSA-N MAX-POWER (a) (b) Average power consumpton (log) (number of channels =1) Average power consumpton (log) (number of SU =1) 4 4 Power consumpton (log(watt)) Power consumpton (log(watt)) Number of secondary users Number of channels CMSA-E CMSA-Q CMSA-L CMSA-E CMSA-Q CMSA-L (c) (d) Fgure 5: Power consumpton performance. (a) Average power consumpton where the number of channels s 1 and the number of secondary users s ncreased. (b) Average power consumpton where the number of secondary users s 1 and the number of channels s ncreased. (c) Average logarthmc power consumpton where the number of channels s 1 and the number of secondary users s ncreased. (d) Average logarthmc power consumpton where the number of secondary users s 1 and the number of channels s ncreased. and wreless channels ncreases. CMSA-E has smlar behavor when the number of wreless channels s ncreasng. However, ncreased competton n Fgure 5(c) forces CMSA- E to ncrease transmsson power of secondary users. Both of CMSA-Q and CMSA-L are more senstve than CMSA-E to load states of secondary users. On the other hand, the equal treatment of CMSA-E for secondary users leads to ncreasng transmsson power when the number of secondary users s ncreased. Heterogenety of added wreless channels leads CMSA-E farness to decrease. As sad before, observng ths mpact of heterogenety confrms the dynamcty of the proposed solutons. 6. Concluson Ths paper proposed two solutons wth fve versons to allocate transmsson power for cogntve rado systems n heterogeneous envronment. Three of these versons are

14 14 Internatonal Journal of Dstrbuted Sensor Networs centralzed mechansms where the decson s generated by central entty whch has all the necessary nformaton. The remanng two versons are dstrbuted mechansms where each cogntve system observes ts surroundng wreless envronment and t uses only ts own sensory data to generate transmsson power decson. Desgns of all proposed solutons were nspred by Spectrum-Tme Dualty concept. Ths concept states that cogntve systems may ntentonally reduce ther spectrum utlzaton for other cogntve systems n return for longer usage tme after these cogntve systems clear ther queues. The optmzaton problem n centralzed solutons tres to clear queues of coexstng cogntve rado systems n a way that reduces competton n future tme frames. The dstrbuted solutons assgn transmsson power by tang nto consderaton the queue sze and the nterference level of all cogntve rado systems. Dynamcal system theory was used to desgn the dstrbuted solutons. Transmsson power allocaton rule was proposed. Ths rule guarantees the overall system evoluton n predcted fashon. In addton, a very powerful machne learnng technque was used. Ths technque s deep neural networ. Applyng ths technque on the proposed dynamcal system led to much better performance. To the best of our nowledge, ths paper s the frst attempt to utlze deep learnng n cogntve rado networs. Results suggest that usng deeper and more sophstcated neural networs wth local nformaton may produce comparable performance to the centralzed solutons where all global nformaton s avalable. One of the most mportant recommendatons of ths wor for future research s to utlze advanced machne learnng technques (.e., deep learnng) n cogntve rado networs. The ncreased complexty of these networs requres hgher level of sophstcaton and self-adaptablty n any proposed soluton whch can be provded to some extent by usng these technques. Competng Interests The authors declare that they have no competng nterests. Acnowledgments ThsresearchssupportedbyTWAS-COMSTECHresearch fund: Intellgent Spectrum Sensng and Sharng n Cogntve Rado Networs (Project Code: 1- RG/ITC/AS C). References [1] J. Mtola III and G. Q. Magure Jr., Cogntve rado: mang softwareradosmorepersonal, IEEE Personal Communcatons,vol.6,no.4,pp.1 18,1999. [] G. Staple and K. Werbach, The end of spectrum scarcty [spectrum allocaton and utlzaton], IEEE Spectrum,vol.41,no., pp.48 5,4. [] S. Srnvasa and S. Jafar, Cogntve rados for dynamc spectrum access the throughput potental of cogntve rado: a theoretcal perspectve, IEEE Communcatons Magazne, vol. 45,no.5,pp.7 79,7. [4] R. Zhang and Y.-C. Lang, Explotng mult-antennas for opportunstc spectrum sharng n cogntve rado networs, IEEE JournalonSelectedTopcsnSgnalProcessng,vol.,no.1,pp. 88 1, 8. [5] D. Nyato and E. Hossan, Compettve prcng for spectrum sharng n cogntve rado networs: dynamc game, neffcency of nash equlbrum, and colluson, IEEE Journal on Selected Areas n Communcatons,vol.6,no.1,pp.19,8. [6] D. Nyato and E. Hossan, Maret-equlbrum, compettve, and cooperatve prcng for spectrum sharng n cogntve rado networs: analyss and comparson, IEEE Transactons on Wreless Communcatons,vol.7,no.11,pp.47 48,8. [7] X.Kang,Y.-C.Lang,H.K.Garg,andL.Zhang, Sensng-based spectrum sharng n cogntve rado networs, IEEE Transactons on Vehcular Technology, vol.58,no.8,pp , 9. [8] X.Wang,Z.L,P.Xu,Y.Xu,X.Gao,andH.-H.Chen, Spectrum Sharng n cogntve rado networs: an aucton-based approach, IEEE Transactons on Systems, Man, and Cybernetcs, Part B: Cybernetcs,vol.4,no.,pp ,1. [9] Z. J and K. J. R. Lu, Cogntve rados for dynamc spectrum access dynamc spectrum sharng: a game theoretcal overvew, IEEE Communcatons Magazne,vol.45,no.5,pp.88 94, 7. [1] Y. Guo, G. Kang, N. Zhang, W. Zhou, and P. Zhang, Outage performance of relay-asssted cogntve-rado system under spectrum-sharng constrants, Electroncs Letters, vol. 46, no., pp , 1. [11] R. Zhang and Y.-C. Lang, Investgaton on multuser dversty n spectrum sharng based cogntve rado networs, IEEE Communcatons Letters,vol.14,no.,pp.1 15,1. [1] S. K. Jayaweera, G. Vazquez-Vlar, and C. Mosquera, Dynamc spectrum leasng: a new paradgm for spectrum sharng n cogntve rado networs, IEEE Transactons on Vehcular Technology,vol.59,no.5,pp.8 9,1. [1] D. Nyato and Z. Han, Dynamcs of multple-seller and multple-buyer spectrum tradng n cogntve rado networs: a game-theoretc modelng approach, IEEE Transactons on Moble Computng,vol.8,no.8,pp.19 1,9. [14] Y. J. Zhang and A. M.-C. So, Optmal spectrum sharng n MIMO cogntve rado networs va semdefnte programmng, IEEE Journal on Selected Areas n Communcatons, vol. 9, no., pp. 6 7, 11. [15] F.R.V.Gumaraes,D.B.daCosta,T.A.Tsftss,C.C.Cavalcante, and G. K. Karagannds, Multuser and multrelay cogntve rado networs under spectrum-sharng constrants, IEEE Transactons on Vehcular Technology,vol.6,no.1,pp.4 49, 14. [16] Q.Zhu,Y.Wu,D.Tsang,andH.Peng, Cooperatvespectrum sharng n cogntve rado networs wth proactve prmary system, n Proceedngs of the IEEE/CIC Internatonal Conference on Communcatons n Chna Worshops (CIC/ICCC 1), pp.8 87, X an, Chna, August 1. [17] F. Ye, Y. L, R. Yang, and Z. Sun, The user requrement based compettve prce model for spectrum sharng n cogntve rado networs, Internatonal Journal of Dstrbuted Sensor Networs, vol. 1, Artcle ID 74581, 8 pages, 1. [18] I. Flppn, M. Cesana, and I. Malanchn, Compettve spectrum sharng n cogntve rado networs: a queung theory based analyss, n Proceedngs of the 1st Internatonal Blac Sea Conference on Communcatons and Networng (BlacSeaCom 1), pp. 8 4, Batum, Georga, July 1.

15 Internatonal Journal of Dstrbuted Sensor Networs 15 [19] A. K. Farraj and E. M. Hammad, Performance of prmary users n spectrum sharng cogntve rado envronment, Wreless Personal Communcatons, vol. 68, no., pp , 1. [] D. Xu, Z. Feng, and P. Zhang, Effectve capacty of delay qualty-of-servce constraned spectrum sharng cogntve rado wth outdated channel feedbac, Scence Chna Informaton Scences,vol.56,no.6,1pages,1. [1] Y.-W. Chan, F.-T. Chen, R. Y. Chang, M.-K. Chang, and Y.- C. Chung, Spectrum sharng n mult-channel cooperatve cogntve rado networs: a coaltonal game approach, Wreless Networs,vol.19,no.7,pp ,1. [] L. Sbou, Z. Rez, and M.-S. Aloun, Capacty of spectrum sharng cogntve rado systems over naagam fadng channels at low SNR, n Proceedngs of the IEEE Internatonal Conference on Communcatons (ICC 1), pp , Budapest, Hungary, June 1. [] I. Chrstan, S. Moh, I. Chung, and J. Lee, Spectrum moblty n cogntve rado networs, IEEE Communcatons Magazne, vol.5,no.6,pp ,1. [4] C.-F. Shh, T. Y. Wu, and W. Lao, DH-MAC: a dynamc channel hoppng MAC protocol for cogntve rado networs, n Proceedngs of the IEEE Internatonal Conference on Communcatons (ICC 1), pp. 1 5, IEEE, Cape Town, South Afrca, May 1. [5] N. Ne and C. Comancu, Adaptve channel allocaton spectrum etquette for cogntve rado networs, Moble Networs and Applcatons,vol.11,no.6,pp ,6. [6] Y.WuandD.H.K.Tsang, Dstrbutedpowerallocatonalgorthm for spectrum sharng cogntve rado networs wth QoS guarantee, n Proceedngsofthe8thConferenceonComputer Communcatons (INFOCOM 9), pp , Ro de Janero, Brazl, Aprl 9. [7] A. Barş andö. B. Aan, BIOlogcally-nspred spectrum sharng n cogntve rado networs, n Proceedngs of the IEEE Wreless Communcatons and Networng Conference (WCNC 7), pp. 4 48, IEEE, Hong Kong, March 7. [8]J.Tang,S.Msra,andG.Xue, Jontspectrumallocatonand schedulng for far spectrum sharng n cogntve rado wreless networs, Computer Networs, vol. 5, no. 11, pp , 8. [9] F. Wang, M. Krunz, and S. Cu, Spectrum sharng n cogntve rado networs, n Proceedngs of the IEEE 7th Conference on Computer Communcatons (INFOCOM 8), Phoenx, Arz, USA, Aprl 8. [] L. Le and E. Hossan, Resource allocaton for spectrum underlay n cogntve rado networs, IEEE Transactons on Wreless Communcatons,vol.7,no.1,pp ,8. [1] H. Nan, T.-I. Hyon, and S.-J. Yoo, Dstrbuted coordnated spectrum sharng mac protocol for cogntve rado, n Proceedngs of the nd IEEE Internatonal Symposum on New Fronters n Dynamc Spectrum Access Networs (DySPAN 7), pp.4 49, IEEE, Dubln, the Republc of Ireland, Aprl 7. []F.Wang,M.Krunz,andS.Cu, Prce-basedspectrummanagement n cogntve rado networs, IEEE Journal on Selected Topcs n Sgnal Processng,vol.,no.1,pp.74 87,8. [] Y.T.Hou,Y.Sh,andH.D.Sheral, Spectrumsharngformulthop networng wth cogntve rados, IEEE Journal on Selected Areas n Communcatons,vol.6,no.1,pp ,8. [4] S. Wang, Z.-H. Zhou, M. Ge, and C. Wang, Resource allocaton for heterogeneous cogntve rado networs wth mperfect spectrum sensng, IEEE Journal on Selected Areas n Communcatons,vol.1,no.,pp ,1. [5] K. L. Haldar, C. Ghosh, and D. P. Agrawal, Dynamc spectrum access and networ selecton n heterogeneous cogntve wreless networs, Pervasve and Moble Computng, vol. 9, no. 4, pp , 1. [6] A. Vzzello, I. F. Ayldz, R. Agustí, L. Favall, and P. Savazz, Cogntve rado resource management explotng heterogeneous prmary users and a rado envronment map database, Wreless Networs,vol.19,no.6,pp.1 116,1. [7] M. Khaled and A. A. Abouzed, Aucton-based spectrum sharng n cogntve rado networs wth heterogeneous channels, n Proceedngs of the Informaton Theory and Applcatons Worshop (ITA 1), pp. 16, IEEE, San Dego, Calf, USA, February 1. [8] M. Neovee, Autonomous spectrum sharng n heterogeneous whte space networs, n Proceedngs of the IEEE Internatonal Conference on Communcatons Worshops (ICC 1),pp.16 16, Budapest, Hungary, June 1. [9] Z. Wang and Y. Zhang, Analyss of cogntve rado spectrum access wth heterogeneous traffc under buffer constrants, n Proceedngs of the rd Internatonal Conference on Consumer Electroncs, Communcatons and Networs (CECNet 1), pp , IEEE, Xannng, Chna, November 1. [4] J. M. Peha, Sharng spectrum through spectrum polcy reform and cogntve rado, Proceedngs of the IEEE, vol.97,no.4,pp , 9. [41] A. Durantn and M. Martno, The spectrum polcy reform pavng the way to cogntve rado enabled spectrum sharng, Telecommuncatons Polcy,vol.7,no.-,pp.87 95,1. [4] Y. Bengo, Learnng deep archtectures for AI, Foundatons and Trends n Machne Learnng,vol.,no.1,pp.1 17,9. [4] R. A. HornandC. R. Johnson, Matrx Analyss, Cambrdge Unversty Press, New Yor, NY, USA, nd edton, 1. [44] H. T. Cheng and W. Zhuang, Smple channel sensng order n cogntve rado networs, IEEEJournalonSelectedAreasn Communcatons, vol. 9, no. 4, pp , 11. [45] F. A. Potra and S. J. Wrght, Interor-pont methods, Journal of Computatonal and Appled Mathematcs,vol.14,no.1-,pp. 81,. [46] F. Taens, Detectng strange attractors n turbulence, n Dynamcal Systems and Turbulence, Warwc 198,D.Randand L.-S. Young, Eds., vol. 898 of Lecture Notes n Mathematcs,pp , Sprnger, Berln, Germany, [47] Y. Bengo, P. Lambln, D. Popovc et al., Greedy layer-wse tranng of deep networs, n Advances n Neural Informaton Processng Systems 19, pp , MIT Press, 7. [48] G. E. Hnton, S. Osndero, and Y.-W. Teh, A fast learnng algorthm for deep belef nets, Neural Computaton,vol.18,no. 7, pp , 6. [49] X. Glorot and Y. Bengo, Understandng the dffculty of tranng deep feedforward neural networs, n Proceedngs of the Internatonal Conference on Artfcal Intellgence and Statstcs (AISTATS 1), Socety for Artfcal Intellgence and Statstcs, Sardna, Italy, May 1. [5] C. Mehlführer, J. Colom Iuno, M. Šmo, S. Schwarz, M. Wrulch, and M. Rupp, The Venna LTE smulators enablng reproducblty n wreless communcatons research, EURA- SIP Journal on Advances n Sgnal Processng, vol.11,no.1, artcle 9, 11. [51] R. Jan, D.-M. Chu, and W. Hawe, A quanttatve measure of farness and dscrmnaton for resource allocaton n shared computer systems, Tech. Rep. DEC-TR-1, Dgtal Equpment Corporaton Research, 1984.

16 Internatonal Journal of Rotatng Machnery Engneerng Journal of Hndaw Publshng Corporaton The Scentfc World Journal Hndaw Publshng Corporaton Internatonal Journal of Dstrbuted Sensor Networs Journal of Sensors Hndaw Publshng Corporaton Hndaw Publshng Corporaton Hndaw Publshng Corporaton Journal of Control Scence and Engneerng Advances n Cvl Engneerng Hndaw Publshng Corporaton Hndaw Publshng Corporaton Submt your manuscrpts at Journal of Journal of Electrcal and Computer Engneerng Robotcs Hndaw Publshng Corporaton Hndaw Publshng Corporaton VLSI Desgn Advances n OptoElectroncs Internatonal Journal of Navgaton and Observaton Hndaw Publshng Corporaton Hndaw Publshng Corporaton Hndaw Publshng Corporaton Chemcal Engneerng Hndaw Publshng Corporaton Actve and Passve Electronc Components Antennas and Propagaton Hndaw Publshng Corporaton Aerospace Engneerng Hndaw Publshng Corporaton Hndaw Publshng Corporaton Internatonal Journal of Internatonal Journal of Internatonal Journal of Modellng & Smulaton n Engneerng Hndaw Publshng Corporaton Shoc and Vbraton Hndaw Publshng Corporaton Advances n Acoustcs and Vbraton Hndaw Publshng Corporaton