IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 1401 Qualiy of Experience Driven Muli-User Video Sreaming in Cellular Cogniive Radio Neworks Wih Single Channel Access Zhifeng He, Shiwen Mao, Senior Member, IEEE, and Sasry Kompella, Senior Member, IEEE Absrac We invesigae he problem of sreaming muli-user videos over he downlink of a cogniive radio nework (CRN, where each cogniive user (CU can access one channel a a ime. We firs consider he case where each CU can sense one channel a a ime slo a mos. To make he problem racable, we ackle he opimal specrum sensing and access problems separaely and develop maching-based opimal algorihms o he subproblems, which yield an overall subopimal soluion. We hen consider he case where each CU can sense muliple channels. We show ha under he assumpion ha all he specrum sensors work on he same operaing poin, a wo-sep approach can derive he opimal specrum sensing and access policies ha maximize he qualiy of experience (QoE of he sreaming videos. The superior performance of he proposed approaches is validaed wih simulaions and comparisons wih benchmark schemes, where a performance gain from 25% o 30% is demonsraed. Index Terms Qualiy of experience (QoE, cogniive radio nework (CRN, decomposiion, muli-user video sreaming, opimizaion. I. INTRODUCTION THE Cisco visual nework index repor predics a drasic increase in mobile daa, a dominan par of which is video relaed, in he near fuure [1]. Such dramaic increase in wireless video raffic, coupled wih he depleing specrum resource, poses grea challenges o oday s wireless neworks. I is of grea imporance o improve he wireless nework capaciy by promoing more efficien use of specrum. This goal can be accomplished by he cogniive radio (CR echnology, which is an evoluionary echnology for more efficien and flexible access o he radio specrum. In a cogniive radio nework (CRN, cogniive users (CUs search for unoccupied licensed specrum in he primary user (PU nework and hen opporunisically access deeced specrum holes in an unobrusive manner. CR has been recognized as an effecive Manuscrip received January 19, 2015; revised Ocober 13, 2015 and February 05, 2016; acceped May 03, 2016. Dae of publicaion May 05, 2016; dae of curren version June 15, 2016. This work was suppored in par by he U.S. Naional Science Foundaion under Gran CNS-0953513, and in par by he Wireless Engineering Research and Educaion Cener a Auburn Universiy. This work was presened in par a IEEE GLOBECOM 2014, Ausin, TX, USA, December 2014. The associae edior coordinaing he review of his manuscrip and approving i for publicaion was Prof. Maria Marini. Z. He and S. Mao are wih he Deparmen of Elecrical and Compuer Engineering, Auburn Universiy, Auburn, AL 36849-5201 USA (e-mail: zzh0008@igermail.auburn.edu; smao@ieee.org. S. Kompella is wih he Informaion Technology Division, Naval Research Laboraory, Washingon, DC 20375 USA (e-mail: sk@ieee.org. Color versions of one or more of he figures in his paper are available online a hp://ieeexplore.ieee.org. Digial Objec Idenifier 10.1109/TMM.2016.2564104 approach o suppor bandwidh-demanding mobile services such as wireless video sreaming [2], [3]. In he area of mulimedia communicaions, subjecive assessmen mehods have been sudied inensively [4], which is shown o reflec viewers percepual qualiy more accuraely han radiional objecive assessmen mehods. The Inernaional Telecommunicaion Union has proposed sandards on subjecive assessmen mehods for various applicaion scenarios [5]. For video ransmission, qualiy of experience (QoE is an effecive percepual qualiy assessmen approach for he percepual visual qualiy of video sequences. One of he mos widely used QoE meric is mean opinion score (MOS [6]. In he MOS model, he visual qualiy of a video sequence is no only dependen on he nework environmen such as packe loss rae, nework delay, bu also dependen on he conen ype. For example, under he same nework condiions, he visual qualiy of video conens wih fas moions (e.g., spors is generally worse han ha of video conens wih slow moions (e.g., news. Several QoE models have been presened in he lieraure (e.g., see [6] [9]. Since he ulimae goal of mos mulimedia communicaion services is o achieve high percepual qualiy for viewers, i is desirable o incorporae QoE models in such sysems. In his paper, we address he challenging problem of downlink muli-user video sreaming in cellular CRNs. We consider a CRN consising of one cogniive base saion (CBS and muliple CUs. Wihou loss of generaliy, we assume each CU can access one channel a a ime (i.e., wih a single anenna. The CUs cooperaively sense PU signals on licensed channels and he CBS infers he channel saes based on colleced CU sensing resuls wih an OR fusion rule. Once he idle channels are deeced, he CBS hen assigns hem o acive CUs for downlink muli-user video sreaming. We incorporae he video assessmen model proposed in [6], [11], aiming o maximize he CU QoE by opimal designs of specrum sensing and access policies. I is obviously a challenging problem o joinly design specrum sensing and access polices for QoE-aware muli-user video sreaming, due o he large number of design facors and he complex ineracions ha should be modeled in a cross-layer opimizaion framework. We firs consider he case where each CU can sense and access a mos one channel a a ime slo. To make he problem racable, we ake a divide-and-conquer approach o break i ino wo sub-problems: (i opimal assignmen sub-problem for specrum sensing (OAPSS: o discover a sufficien amoun of idle channels reliably and quickly o mee he bandwidh demand of he CUs; and (ii opimal assignmen sub-problem for video ransmission (OAPVT: o allocae available channels o CUs according o heir respecive QoE requiremens and nework saus. We 1520-9210 2016 IEEE. Personal use is permied, bu republicaion/redisribuion requires IEEE permission. See hp://www.ieee.org/publicaions sandards/publicaions/righs/index.hml for more informaion.
1402 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 propose a disribued greedy poly-maching algorihm (GPA ha can compue opimal soluion o he channel sensing sub-problem, and using he Hungarian Mehod o compue opimal soluion o he channel assignmen sub-problem. Furhermore, we examine he more general case where each CU can sense muliple channels (e.g., wih muliple specrum sensors bu can sill access only one channel a a ime slo. We formulae an inegraed problem ha maximizes he QoE of all he CUs by joinly opimizing specrum sensing and access policies. Under he assumpion ha all he specrum sensors work a he same operaing poin (i.e., wih he same probabiliy of deecion and he same probabiliy of false alarm, we show ha his challenging problem can be solved wih a wo-sep approach: firs, he specrum sensing scheduling problem is solved wih a greedy algorihm; Second, he channel allocaion problem, which is a Maximum Weigh Maching problem and can be solved opimally wih he Hungarian Mehod. We prove ha he wo-sep soluion algorihm is indeed opimal: decomposing he original problem ino wo sub-problems and solving hem sequenially do no sacrifice he opimaliy of he soluion. I is worh noing ha if we also assume idenical operaing poins for he specrum sensors, he single-channel sensing scenario is a special case of he muli-channel sensing scenario, o which he opimal soluion approach also applies. We validae he proposed schemes wih simulaions, and he simulaion resuls demonsrae heir superior performance in erms of he MOS ha CUs can achieve under various nework scenarios, when compared wih benchmark schemes. The remainder of his paper is organized as follows. The sysem model is presened in Secion II. The problem for he case of single channel sensing is formulaed and solved in Secion III. The problem for he case of muli-channel sensing is formulaed and solved in Secion IV. Simulaion resuls are discussed in Secion V. Secion VI reviews relaed work and Secion VII concludes he paper. II. SYSTEM MODEL We consider a primary nework operaing on N 1 orhogonal licensed channels. The primary nework is co-locaed wih a CR nework, which consiss of a CBS supporing M CUs. The CUs sense he PUs usage of he licensed channels and access he licensed channels in an opporunisic manner. As in prior work [15], [20], we assume he CUs, when hey are no receiving daa, measure he SNRs of PU ransmissions over all he licensed channels and repor he measured SNRs o he CBS hrough some feedback mechanism. Based on such feedback, he CBS hen assigns hose CUs wih good channel condiions o sense each licensed channel, in order o achieve a good sensing performance. We consider he downlink muli-user video sreaming scenario, where he CBS sreams a video o each acive CU using he license channels ha are deeced idle. We consider he mos general case where each CU is sreaming a differen video. We assume ime is divided ino a series of non-overlapping group of picures (GOP windows, each consising of T ime slos. Each ime slo can be furher divided ino four phases for specrum sensing and access for muli-user video sreaming, as shown in Algorihm 1. Noe ha a he very beginning of he firs GOP window, he SNR informaion used in Phase 1 may no be available ye. Algorihm 1 Specrum sensing and access for QoE-driven muli-user video sreaming 1 Phase 1: The CBS deermines for each CU which channel o sense based on SNR feedback, and broadcass he sensing schedule o he CUs 2 Phase 2: Each CU follows he sensing schedule o sense he channel o which i is assigned, and repors he sensing resul o he CBS 3 Phase 3: The CBS makes wo decisions: (i channel availabiliy a he curren ime slo, based on he sensing resuls and he fusion rule; and (ii channel assignmen o CUs for muli-user video ransmission a he curren ime slo, based on channel availabiliy, channel condiion, Conen Type (denoed as CT of each CU, and oher informaion. Then he CBS broadcass he channel access schedule o he CUs 4 Phase 4: The CBS uses he assigned channels o ransmi video daa; each CU follows he channel access schedule o receive video daa from is assigned channel. However, such informaion can be obained via esimaion or learning echniques, or by simply leing CUs probe he channels when hey are idle [16]. III. THE CASE OF SINGLE CHANNEL SENSING We firs consider he case ha each CU can only sense a single channel and access a single channel during a ime slo. We consider he case ha each specrum sensor has is own operaing poin, which may be differen from ha of oher specrum sensors. This urns ou o be an ineger programming (IP problem, which is NP-hard in general. We hen ake a divide-andconquer approach o break down he problem ino an OAPSS and an OAPVT. We develop effecive soluion algorihms for each sub-problem and prove heir opimaliy o each sub-problem. However, he overall soluion is near-opimal due o he divideand-conquer approach. A. Problem Formulaion 1 Opimal Assignmen Sub-Problem for Specrum Sensing: In a pracical wireless nework scenario, CUs are locaed a differen geographical posiions wih differen channel gains o primary ransmiers. Thus heir performance on deecing primary signals on a paricular licensed channel would be differen, e.g., a CU wih a beer channel gain o a primary ransmier may have a higher probabiliy of deecing he PU s single. By selecing a group of CUs which have beer channel gains o a PU, he probabiliy of deecing he PUs signal would be higher, and hen he probabiliy of inerfering PU ransmissions can be reduced [20], [21]. Usually cooperaive sensing is used o improve he deecion performance by fusing he sensing resuls from muliple CUs [22], and a cerain fusion rule is required o combine hese resuls. In his paper, he OR fusion rule is used a he CBS o deermine he presence or absence of PU signal on a paricular channel. Wih he OR rule, if any of he CUs repors he presence of a PU signal hen he CBS decides ha he channel is busy; oherwise, he CBS decides ha he channel is idle. We use an M N 1 marix X o denoe he sensing ask assignmen
HE e al.: QOE DRIVEN MULTI-USER VIDEO STREAMING IN CELLULAR CRNS WITH SINGLE CHANNEL ACCESS 1403 a ime slo, while he enry locaed a he ih row and jh column posiion is defined as { x 1, CU i senses channel j in ime slo (1 0, oherwise. A useful meric o evaluae he performance of deecing a PU signal is probabiliy of deecion, which is he probabiliy ha a CU successfully deecs he exisence of an exising PU signal. Le P denoe he probabiliy of deecion on channel d j by CU i a ime slo. For an energy deecor, we have [21] P d 1 2 erfc ( (λ σ 2 n ς K 1 2 ( 2ς +1 where λ is he hreshold of energy deecion on channel j by CU i a ime slo, σn 2 is he power of he i.i.d. addiive whie Gaussian noise a he CU, ς is he SNR of PU s signal on channel j a CU i, K is he number of samples on channel j by energy deecion. In (2, erfc(z 2 π z e u 2 du is he complemenary error funcion, and le erfc 1 ( denoe he inverse funcion of erfc (. According o he OR fusion rule, he probabiliy of deecion on channel j a ime slo is (2 M ( x Pd j 1 1 Pd. (3 In order o guaranee he proecion of he PUs, we se Pd P d by uning λ for all i, j. Thus he probabiliy of deecion of he aciviy of a PU will be greaer han P d if he channel is sensed by some CUs [according o (3]. In he case ha a channel is no sensed by any of he CUs, i will no be used for video sreaming. Under he assumpions ha he PU signal is complex valued phase-shif keying and he noise is circularly symmeric complex Gaussian, hen CU i s probabiliy of false alarm on channel j, denoed by P, can be expressed as [21] f P f ( (λ 1 2 erfc K σn 2 1 (4 2 ( 1 2 erfc 2ς +1erfc 1 ( 2 P K d + 2 ς. (5 The objecive of sensing ask assignmen is o maximize he probabiliy of deecing all he idle channels a ime slo, while mainaining fairness among he probabiliies of deecion of he N 1 licensed channels. I has been shown ha proporional fairness can be achieved by maximizing he sum of logarihmic funcions. 1 The opimal sensing ask assignmen problem is o 1 This is because of he concaviy of he logarihmic funcions. The marginal incremen of logarihmic funcions is decreasing. Therefore, when we maximize he sum of logarihmic funcions, i ends o allocae resources (CUs evenly o sense differen channels. maximize he following objecive funcion: N 1 log (1 P fj N 1 where ϕ. log alarm on channel j as N 1 M log ( 1 P f x ( log 1 P x f ( 1 P f and P f j N 1 ϕ x (6 is he probabiliy of false M ( x Pf j 1 1 Pf. (7 We assume ha each CU can sense one channel a each ime slo, and he number of CUs ha can be assigned o sense a channel i a each ime slo is unresriced. Therefore, he opimal sensing ask assignmen problem is formulaed as N 1 OAPSS:max ϕ x (8 s.. N 1 x 1, for all i (9 x {0, 1}, for all i, j. (10 2 Opimal Assignmen Problem for Video Transmission: For video qualiy assessmen, we adop he QoE model named mean score opinion (MOS ha was proposed in [11], where he MOS of CU i using channel j during ime slo, denoed by Ψ, can be expressed as ( Ψ α + CT iγ + (β + CT i δ ln SBR ( α + CT i γ + (β + CT i δ ln (Blog 2 1+SNR (11 where α 3.9860, β 0.0919, γ 5.8497, and δ 0.9844 are consans, CT i is he Conen Type of he video sequences required by CU i, B is he bandwidh of a channel in kbps, and SNR is he SNR of he video signal using channel j measured a CU i a ime slo [11]. The QoE model (11 is developed in [11]. This model has been used and shown feasible for video sreaming over CR neworks in prior works (e.g., [19]. As indicaed in [11] and [19], he SBR should be adjused according o he changing channel condiions, e.g., channel daa rae. Therefore, SBR is an adjusable parameer here in our problem. As in prior work [11], [19], [31], [32], we do no consider he packe level behavior. Insead, we use he effecive received daa rae (e.g., a he playou buffer as SBR (while packe level error conrol is implicily considered. We assume he adapive video sreaming scheme gracefully adaps he SBR o he available nework downlink bandwidh, as in [11], [31], [32]. Such graceful adapaion of he effecive video rae can be achieved wih he fine granulariy scalabiliy or scalable video coding [3]. We assume ha N 2 channels are deermined o be idle afer he sensing phase, where N 2 N 1. We consider a general case
1404 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 where no all he CUs have daa o receive a all imes. Insead, he probabiliy of a CU has daa o receive a each GOP window is 0 ξ 1. The number of CUs ha have daa o receive in a GOP window (called acive CUs is denoed as M 1, where M 1 M. AnM 1 N 2 marix Z is used o represen channel access assignmens on ime slo, while he enry locaed a he i-h row and jh column posiion is { z 1, assign channel j o CU i in ime slo (12 0, oherwise. We consider he case where each CU can use a mos one channel a each ime slo due o hardware consrains, and each channel can be used by a mos one CU a each ime slo. We aim o maximize he expeced average MOS of all he CUs during a GOP window by assigning he available channels max 1 T T M 1 E [ Ψ ] i 1 (13 where Ψ i is he MOS of CU i a ime slo. The above objecive funcion can be maximized if we maximize he expeced MOS incremen of he M 1 CUs during each ime slo [3], which can be wrien as M 1 E [ M ] 1 N 2 Ψ i E [ Ψ] z j 1 M 1 N 2 [P ( rj 0 s j 0 φ j 1 +P ( r j 1 s j 0 θ ]z (14 where s j 0indicaes he channel is sensed idle; P (r j 0 and P (rj 1 are he probabiliy of channel j o be idle or busy a ime slo, respecively; P (rj 0 s j 0 and P (rj 1 s j, j, j 0(denoed as P00 and P10, respecively are he condiional probabiliy for channel j o be idle or busy condiioned on he sensing resul, respecively. I follows ha P j, 00 P (r j 0 s j 0 (1 Pf j P (rj 0 (1 Pf j P (rj 0+(1 P d j P (rj 1 P j, 10 P (r j 1 s j 01 P (r j 0 s j 0. In (14, φ and θ are he effecive daa rae of he received video sequence a CU i using channel j which is indeed idle or busy a ime slo, respecively. Denoe μ and ν as he received SNR a CU i using channel j which is indeed idle or busy a ime slo, respecively. We hen have μ Γg i n 0 B ν Γg i n 0 B(1 + ς φ α + CT iγ +(β + CT i δln ( ( B j log 2 1+μ θ α + CT iγ +(β + CT i δln ( ( B j log 2 1+ν where Γ is he CBS ransmi power on channel j, for all j. Define ϖ as P j, 00 φ + P j, 10 θ. (15 ϖ The opimal channel access problem is formulaed as M 1 N 2 OAPVT: max ϖ z (16 s.. N 2 z 1, i {1,...,M 1 }. (17 M 1 z 1, j {1,...,N 2 } (18 z {0, 1}, for all i, j. (19 3 OAPVT Considering Fairness Among CUs: Now we consider achieving fairness among CUs for he channel allocaion problem. Considering he fac ha our objecive is o maximize he expeced average MOS of all he CUs during a GOP window by assigning he available channels, we propose o achieve a long erm fairness among CUs. In order o achieve long erm fairness among CUs, we propose o allocae channels more evenly o differen CUs in differen ime slos. For example, when here is only one available channel and wo CUs, A and B, if CU A is scheduled for video sreaming in he previous ime slo, hen a he curren ime slo, CU B should be scheduled for video sreaming. Generally speaking, a he curren ime slo, when he number of idle channels available for video sreaming, e.g., N 2, is less han he number of acive CUs, e.g., M 1,, he CUs ha have no been scheduled in previous ime slos will have a higher prioriy of being scheduled han he CUs ha have been scheduled in previous ime slos, while he objecive is sill o maximizing he MOS sum of all CUs. Therefore, we consider he following wo cases. a A he curren ime slo, N 2 M 1. In his case, all CUs can be scheduled for video sreaming, and he problem formulaion is he same as problem OAPVT. b N 2 <M 1. For ease of presenaion, denoe Θ as he se of acive CUs a he curren ime slo. Denoe Θ k, a subse of Θ, as he se of acive CUs whose oal imes of being scheduled so far is k; and denoe Θ k+1, also a subse of Θ, as he se of acive CUs whose oal imes of being scheduled so far is k +1, k 0, 1, 2, 3,...Le denoe cardinaliy of a se. Then we have Θ M 1, Θ k Θk+1 Θ, and Θ k + Θ k+1 M 1. i If N 2 Θ k M 1, hen he problem is o maximize he MOS sum for he CUs in Θ k, by choosing N 2 CUs from Θ k and allocae he N 2 available channels o he N 2 CUs. We hus have he following problem: P1 :max s.. N 2 ϖ z (20 i Θ k N 2 z 1, i Θ k. (21 z 1, j {1, 2,...,N 2 } (22 i Θ k z {0, 1}, for all i Θ k, 1 j N 2.
HE e al.: QOE DRIVEN MULTI-USER VIDEO STREAMING IN CELLULAR CRNS WITH SINGLE CHANNEL ACCESS 1405 ii If Θ k <N 2 <M 1, hen all he CUs in Θ k would be scheduled. In addiion, N 2 Θ k CUs would be chosen from se Θ k+1 o be scheduled as well. Then he N 2 channels is allocaed o he Θ k +(N 2 Θ k CUs o maximize he MOS sum. We have he following problem: P2:max i Θ s.. N 2 N 2 N 2 ϖ z (23 z 1,i Θ k (24 z 1, i Θ k+1 (25 z 1, j {1,...,N 2 } (26 i Θ z {0, 1}, for all i Θ, 1 j N 2. (27 Lemma 1: According o he definiions of Θ, Θ k, and Θ k+1, we have Θ k Θk+1 Θ, where k 0, 1, 2,..., a he beginning of every ime slo. The proof direcly follows he formulaions of problems P1 and P2, and he fac ha each CU can access a mos one channel a a ime. B. Soluion Algorihms and Analysis 1 Poly-Maching Based Soluion o OAPSS: We can see ha he OAPSS problem is formulaed as he well-known general assignmen problem (GAP, which is NP-hard in general. However, since here is no consrain on how many CUs can be assigned o a channel, he problem is acually a maximum weigh poly-maching problem on a biparie graph ha maches CUs o licensed channels wih edge weighs defined as ϕ. Furhermore, a channel can be mached o muliple CUs. I can be solved by he following greedy sraegy presened in Algorihm 2 [23]. Wih his algorihm, each CU selecs he channel wih he larges weigh, regardless wheher he seleced channel has been chosen by oher CUs or no [23]. This greedy sraegy has a ime complexiy of O(MN 1. In fac his is a disribued algorihm, since each CU can choose is bes channels o sense and here is no need o involve he CBS in his phase. Since he CUs can launch heir searching procedures in Line III-B1 in parallel, his disribued sraegy has a ime complexiy of O(N 1. In he following heorem, we also show ha he GPA is opimal. Theorem 1: The greedy poly-maching algorihm 2 achieves he opimal soluion o he OAPSS problem. Proof. Exchanging he summaion order, he objecive funcion of he OAPSS sub-problem (8 becomes M ( N 1 ϕ x, where N 1 ϕ x is he uiliy ha CU i can achieve under he wo consrains (9 and (10. Since each CU can have a mos one channel, he maximum uiliy CU i can achieve is max j {ϕ x }, which is accomplished in Line 4 of Algorihm 2. Since he opimal sraegies of he CUs do no conflic wih each oher and hus are independen o each oher, he maximum uiliy of he CUs are also independen of each oher. I follows ha max M ( N 1 ϕ x M ( { } maxj ϕ x, and Algorihm 2 is opimal. 2 Soluion o he Channel Accessing Problem: The hree problems, OAPVT, P1, and P2, are all IP problems, which are NP-hard in general. However, an ineresing characerisic of he hree problems is ha he coefficiens of he consrain marix in hese problems are all eiher 0 or 1, such ha he unimodulariy propery holds rue in hese problems [26], [27]. As a resul, he opimal and feasible soluions o hese problems can be obained by solving heir LP relaxaions. Thus hese problems can be solved wih he Simplex mehod [10], [28], which has a polynomial-ime average-case complexiy. IV. THE CASE OF MULTI-CHANNEL SENSING We nex consider he general case ha a CU can sense muliple channels bu can sill access one channel a a ime (e.g., each CU is equipped wih muliple specrum sensors bu wih only one ransceiver. To make he problem racable, we assume ha all he specrum sensors are uned o have he same probabiliy of deecion and he same probabiliy of false alarm. Under his assumpion, we presen a problem formulaion ha inegraes boh specrum sensing and access for QoE driven video sreaming. We hen develop a wo-sep algorihm wih proven opimaliy. Noe ha if his assumpion is made for he problem examined in Secion III, hen he single channel sensing problem becomes a special case of he muli-channel sensing problem, which can be solved wih he opimal soluion algorihms developed in his secion. For breviy, we omi he superscrip on all he relevan symbols in he res of his secion. A. Problem Formulaion We assume ha here are M CUs and N licensed channels. CU i can sense a mos C channels and access a mos one channel a a ime slo. Furhermore, each channel mus have Λ CUs o sense i o guaranee ha he cooperaive probabiliy of deecion on a channel saisfies P d 1 ( 1 P Λ.We d also have MC < NΛ, which means ha only pars of he N channels can be sensed a a ime slo. As discussed, in addiion o P d P d,wealsohavep f P f, for all i, j. From(2 and (4, his can be achieved by solving he sysem of hese wo equaions for he hreshold of energy deecion λ and he number of samples K for each specrum sensor i on channel
1406 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 j wih a differen SNR value ς.wehave ( erfc 1 (2 P f 2ς +1erfc 1 (2 P 2 d K 2 ς ( λ σn 2 ς erfc 1 (2 1+ P f erfc 1 (2 P f 2ς +1erfc 1 (2 P. d Le I ( i x Λ be an indicaor funcion defined as I ( i x Λ { 1, if i x Λ 0, oherwise. We hen have P (s j 0 { P (r j 0(1 P fj +P(r j 1(1 P dj } I ( i x Λ P (s j 11 P (s j 0. (28 Le S {sj,j 1, 2,...,N} represens he cooperaive sensing resuls on he N licensed channels. There are 2 N possible oucomes for S in oal. Le S h be he hh oucome, 0 h 2 N 1. Define Γ j (h o be he jh elemen in S h, j 1, 2,...,N. Assuming independen channel saes, he probabiliy of geing oucome S h as a sensing resul is P ( S S h N P (s j Γ j (h j 1 N [(1 Γ j (hp (s j 0+Γ j (hp (s j 1]. (29 For a sensing oucome S h,leφ h {j :Γ j (h 0, j 1, 2,...,N} be he se of channels ha are sensed idle. Le Y h [ y] h, 1 i M, j Φh, be he channel assignmen marix, where y h {0, 1} is he amoun of ime ha CBS ransmis o CU i on channel j in a ime slo, when he sensing oucome is S h. The channel assignmen sraegy can be expressed as Y [Y 0, Y 1, Y 2 N 1]. According o condiional expecaion, he expeced overall MOS can be derived as ( M 2 N 1 E Ψ i E(Ψ i S S h P ( S S h 2 N 1 h0 h0 E(Ψ i S S h P ( S S h. Wih he MOS model used in Secion III, we have [ E Ψ i ] S N (30 ( P j 00 φ + P j 10 θ y h (31 where P j 00 P (r j 0 s j 0 (1 P f P (r j 0 (1 P f P (r j 0+(1 P d P (r j 1, if i x Λ 0, oherwise (32 and { 1 P j P j 10 P (r 00 j 1 s j 0, if i x Λ 0, oherwise. (33 Define P j 00 φ + P j 10 θ, if i x Λand channel w j is sensed idle 0, oherwise. (34 Then he maser problem of maximizing he oal expeced QoE of all he video sessions, denoed as MP, can be formulaed as follows: 2 N 1 N MP : max : w y h P ( S S h (35 h0 s.. N y h 1, for all i, h (36 y h 1, for all j, h (37 x Λ, for all j (38 N x C, for all i (39 Equaion (34 x {0, 1}, for all i, j (40 y h {0, 1}, for all i, j, h. (41 I can be observed ha he formulaed problem MP is an ineger nonlinear programming problem, which is NP-hard in general, alhough a rigorous proof is no given in his paper. We nex show ha problem MP can be decomposed ino wo sub-problems and solved wih a wo-sep approach wihou sacrificing opimaliy. B. Soluion Algorihms Firs, we use Algorihm 3 o solve he specrum sensing subproblem, denoed as SP1, i.e., o deermine he sensing ask assignmen marix X. In Algorihm 3, we sor he N channels according o P (r j 0, j 1, 2,...,N, in descending order. We hen assign CUs o sense he sored channels sequenially
HE e al.: QOE DRIVEN MULTI-USER VIDEO STREAMING IN CELLULAR CRNS WITH SINGLE CHANNEL ACCESS 1407 s.. N y 1, i (43 y 1, j (44 y {0, 1}, for all i, j. (45 Clearly SP2 is also a maximum weigh maching problem and is he same as OAPVT. I can be solved wih opimal soluion using he Hungarian mehod. as follows. For he firs channel in he remaining channel lis, if here are no less han Λ CUs each of which can sill sense some exra channels, choose Λ CUs o sense he channel; oherwise, he channel is conservaively claimed o be busy in order o avoid poenial collision wih PUs using his channel. Iniially each CU can sense C channels, i.e., wih sensing capabiliy C. Each ime a CU is assigned o sense a channel, is sensing capabiliy will be reduced by 1. In Algorihm 3, line 1 sors he N channels, line 2 o line 4 iniialize he sensing capaciy of each CU, and line 5 o line 26 assign CUs o sense he N channels. Line 6 o line 17 is o choose λ CUs o sense a channel, and he sensing capaciy of a CU is reduced by 1 a each ime he CU is chosen o sense a channel. Line 18 o line 25 check if here remain a sufficien number of CUs o sense he nex channel. If rue, hen i will assign he CUs o sense he nex channel; oherwise, i will sop sensing he remaining channels. Line 27 o line 31 deermine he channels ha is no sensed by a sufficien number of CUs and herefore such channels are deermined o be busy. Afer obaining he sensing ask assignmen marix X from Algorihm (3, specrum sensing is conduced by CUs following he assignmens and sensing resuls are repored o he CBS. The CBS hen solves he following sub-problem, denoed as SP2, o obain he channel allocaion marix Y, which will be broadcas o he CUs for channel access: SP2 : max : N w y (42 C. Opimaliy Proof Alhough problem MP is solved wih he wo-sep approach in Secion IV-B, we show ha he soluion is acually opimal in Theorem 2. Theorem 2: Le [X, Y ] denoe he opimal soluion o problem MP, where X is he opimal specrum sensing sraegy and Y is he opimal channel allocaion sraegy. Then X can be obained by running Algorihm 3 and Y can be obained by solving problem SP2. Proof. Le j and j be he indexes of wo licensed channels such ha P (r j 0 P (r j 0. Define W [w 11,,w M 1,,w 1j,,w Mj,, w 1N,, w MN ] and W [w11,,wm 1,,w 1j,,w Mj,, w1n,,w MN ]. Also denoe X [x 11,,x M 1,, x 1j,,x Mj,, x 1N,, x MN ] and X [x 11,,x M 1,,x 1j,,x Mj,, x 1N,, x MN ] as he feasible sensing ask assignmen marices corresponding o W and W respecively. Le x 0in X, for all i, and x 0in X, for all i. Then we have w 0in W, for all i, and w 0in W, for all i.lex x, for all j j,j j, for all i. I follows ha w w (denoed as w, for all j j,j j, for all i.le i x i x Λ. Then according o (34, we have w w, for all i. We firs prove he following lemma, which will serve as a basis for he laer par of he proof for Theorem 2. Lemma 2: Denoe SP2 and SP2 as he channel allocaion problem corresponding o W and W as defined above, respecively. If here is a feasible soluion Y for SP2, hen here is always a feasible soluion, denoed as Y,forW, such ha W Y T W Y T, where ( T denoes he marix ranspose operaion. Proof. Le Y [y11,,ym 1,,y 1j,,y Mj,, y1n,, y MN ] be he feasible channel assignmen marix corresponding o W.Ley y, for all j j or j, y 0 or 1, for all i, y u i,u i 0or 1, for all i, y v îj î,v î 0 or 1, for an î I {1,...,M}, and y 0, for all i î, i I {1,...,M}. Then Y [y 11,,y M 1,,y 1j,,y Mj,, y 1N,, y MN ] wih y y, for all j j or j, for all i, y u i, for all i (recall ha w 0, for all i in W, y vî, and îj y 0, for all i î, i I {1,...,M}, will be a feasible
1408 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 soluion o SP2. This is because he consrains in SP2 are sill saisfied as follows. 1 The number of users on any channel j j or j in soluion Y is he same as ha in soluion Y, i.e., i y i y i y. 2 The number of users on channel j (or j in soluion Y is he same as ha on channel j (or j in soluion Y, i.e., i y i u i i y (or i y v î i y. Noe ha he consrain on he number of users on channel j is he same as ha on channel j. 3 The number of anennas ha CU î uses in soluion Y is he same as ha in soluion Y, i.e., j j,j (y + îj y + îj y îj j j,j (y îj + uî + vî j j,j (y + îj y + y îj îj. 4 The number of anennas ha CU i, for all i î, uses in soluion Y is he same as ha in soluion Y, i.e., j j,j (y + y + y j j,j (y îj + uî +0 j j,j (y + îj y + y. From he firs wo bulles above, i can be seen ha in SP2, he consrain for each channel j is saisfied. From he hird and fourh bulles above, we know ha he consrain for each CU is also saisfied. Therefore, we conclude ha Y is also a feasible soluion o problem SP2. I hen follows ha W Y T W Y T w y wy i j i j w y + w y + i j j,j i i ww + w y + i j j,j i i i w y i w y w y w y w y + w y w îj îj y + w y îj îj i î i î w 0+w y w îj îj 0+w y îj îj i î i î w y w y w îj îj îj îj îj vî w îj vî (w îj w îj vî 0 (46 where (46 is due o he fac ha w 0and w 0, for all i. Then he lemma immediaely follows. Denoe MP and MP as he original join-opimizaion problem wih {W, Y }, and {W, Y } as defined above, respecively, and Δ and Δ as he corresponding objecive funcion value of MP and MP, respecively. I follows Lemma (2 TABLE I SIMULATION PARAMETERS Parameers Value Parameers Value M 30 μ 21 db o 11 db N 1 30 ν 80 db o 60 db K 10 4 ς 30 db o 10 db ( } f s 10 6 Hz max j {Pr H 0 j 0.9 T 10 Pd 0.95 B 10 6 Hz Pf 0.1 C 3 Λ 4 ha Δ Δ 2 N 1 h0 P ( S ( S h W Y T W Y T 0. The above inequaliy indicaes ha when we have limied specrum sensing capabiliy and canno guaranee a saisfacory probabiliy of deecion o all he channels, in order o maximize he expeced uiliy we can obain from he possible sensing resuls and he corresponding opimal ransmission sraegy, we should assign he highes prioriy o he channel ha has he highes probabiliy of being idle, and allocae CUs ha sill have sensing capabiliy o sense his channel. I would be subopimal if we allocae CUs wih exra sensing capabiliy (if hey do exis o sense oher channel(s ha has(have a lower probabiliy of being idle. This is exacly he same sraegy used in Algorihm 3, i.e., assigning CUs o sense he channels in a decreasing order of heir probabiliies of being idle. This concludes he proof of he heorem. V. SIMULATION STUDY A. Parameers Configuraion The performance of he proposed algorihms is validaed wih MATLAB simulaions. We consider a scenario in which he PUs and CUs are randomly disribued around a CBS wihin he service radius of he CBS. Each CU requess a video sequence of a cerain conen ype, while differen CUs may reques differn videos. The reques is sen o he CBS and he CBS decides he channel allocaion based on he objecive of maximizing he MOS sum of all CUs. Table I presens he values of simulaion parameers used in he simulaions, where f s is he sampling frequency a he CUs wih energy deecion. I is verified ha he range of MOS is wihin 1 o 5 wih he value of he parameers provided as in [11]. As in [11], H.264 video is used for he subjecive es of building he QoE model, where he frame srucure is IPPP for all he sequences. The GOP size (i.e., he oal number of frames in a group is 4 and he inerval beween P-frames is 1. B. Benchmark Schemes We firs examine he performance of he proposed algorihms for he single-channel sensing case, which solves he OAPSS and OAPVT problems separaely. We erm his scheme proposed scheme 1 (PS1 in his secion. We compare PS1 wih hree benchmark schemes presened in [30] (ermed Benchmark 1, [19] (ermed Benchmark 2, and [33] (ermed Benchmark 3, respecively.
HE e al.: QOE DRIVEN MULTI-USER VIDEO STREAMING IN CELLULAR CRNS WITH SINGLE CHANNEL ACCESS 1409 Fig. 1. MOS and daa rae relaionship for hree reference video sequences. Fig. 2. Performance of channel sensing versus he minimum channel idle probabiliy (95% confidence inervals ploed as error bars. In paricular, in [30], he auhors assume ha he QoE model is known bu he parameers are unknown. The algorihm esimaes he QoE model parameers by observing he realized qualiy sum of all user. I hen dynamically changes he channel allocaion based on he esimaed QoE parameers, in order o maximize he qualiy sum of all users. Since our QoE model is adoped from [11], where packe error rae (PER is no considered, while model in [30] considers PER, we ake he approach in [19] o se he PER for Benchmark 1 model for a fair comparison. In [19], he enire group of CUs are caegorized ino hree classes of differen prioriy. The prioriy of a CU is deermined by he video sequence ha i acquires from he CBS. The CUs acquiring Suzie have he highes prioriy of accessing a channel, he CUs acquiring Carphone have he second highes prioriy, and he CUs acquiring Fooball have he lowes prioriy, where Suzie, Carphone, and Fooball are hree video sequences of differen conen ypes used in our simulaions. However, his scheme does no consider he variabiliy of channel gains among he CUs. We also compare he proposed algorihm wih he algorihm proposed in [33]. which consiss of a se of accepabiliy-based QoE models, denoed as A-QoE, based on he resuls of comprehensive user sudies on subjecive qualiy accepance assessmens. The models are able o predic users accepabiliy and pleasanness in various mobile video usage scenarios. In our proposed and he benchmark schemes, only he channel assignmen algorihm is differen. The number of CUs engaged for each simulaion is he same for all he schemes. Our proposed approach and he benchmark approach share he same channel sensing resul and he same video raffic reques for each CU. And hen channel allocaion is conduced for he wo differen approaches. C. Simulaion Resuls and Discussions As a basis for our simulaions and discussions, Fig. 1 plos he relaionship beween MOS and daa rae according o (11 for hree widely used es video sequences wih differen conen ypes, including Suzie, Carphone, and Fooball. The parameers are obained from [11]. The resuls are as expeced since generally for he same daa rae, he MOS of a slow moion video sequence is higher han ha of a high moion video sequence. We use hese video sequences in he simulaions presened in he res of his secion. The effeciveness of he sensing algorihm componen of PS1 is presened in Fig. { 2. ( We increase he minimum channel idle probabiliy min j P r j 0 } from 0.1 o 0.47 and plo he real channel saes and he sensed channel saes. As a benchmark, we also presen he simulaion resuls wih he random sensing scheme used in [25] and [29]. Wih random sensing, each CU randomly and independenly selecs one of he N 1 channels o sense wih equal probabiliy. As uilizaion of he channels decreases, he number of idle channels increases. The proposed sensing algorihm can discover more idle channels for CUs o use. Moreover, he number of channels ha miss deecion is less han 0.5 on average, which is less han N 1 (1 P 1.5. d j Recall ha N 1 is he oal number of channels and P is he d j probabiliy of deecion. So N 1 (1 P is he expeced d j number of channels ha miss deecion. The sensing algorihm offers an accepable level of proecion o he PUs, and is effecive in deecing idle channels for he CUs. We nex compare he expeced MOS of all he CUs a each ime slo (denoed as Ψ during an enire GOP window. In our simulaions, each CU requess a video sequence of a cerain conen ype (differen CUs may reques videos of differen conen ype, and he reques is sen o he CBS. The CBS decides he channel allocaion based on he objecive of maximizing he MOS sum of all CUs. In Fig. 3, we plo he achieved MOS sum of all he CUs { achieved ( by PS1 and he Benchmark schemes. We se min j P r j 0 } 0.5and raffic load ξ 1in his simulaion. Fig. 3 shows ha he proposed QoE-aware scheme achieves a consisenly higher sum han all he hree benchmark schemes during he enire GOP window. The main reason is ha Benchmark schemes 1 and 3 only consider channel gain diversiy among he CUs while allocaing channels, and Benchmark scheme 2 assigns channels o he CUs based on heir respecive prioriies only and channel gain diversiy is no considered among he CUs, which may resul in a subopimal sraegy o he objecive of maximizing he MOS sum of all CUs. Fig. 4 demonsraes how he CU video qualiy is affeced by he raffic load of he CUs (i.e., ξ. The average MOS per CU
1410 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 Fig. 3. Average MOS per CU over ime during 10 GOP windows. Fig. 5. Average MOS sum of he CUs over an enire{ GOP( window, } avg Ψ, versus he minimum channel idle probabiliy, min i,j Pr H 0j,andhe { } minimum SNR of CUs, min i,j μ. from (15 ha ϖ P (r j 0 s j 0 (φ θ + θ (1 Pf j P (rj 0 (φ θ (1 Pf j P (rj 0+(1 P d j P (rj 1+ θ. Fig. 4. Average MOS per CU over ime during 10 GOP windows for differen CU raffic loads ξ (wih 95% confidence inervals. during 10 GOP windows achieved by PS1 and he benchmark schemes are ploed, where 95% confidence inervals are ploed as error bars. As he CU raffic load is increased, more CUs need channels for video ransmission. We can see ha while he number of he really idle channels is greaer han he number of acive CUs, he average MOS per CU of all schemes increases wih ξ, and he performance gap beween our proposed scheme and he benchmark schemes grows larger. While he number of really idle channels is no greaer han he number of acive CUs, he average MOS sum of boh schemes remain he same, since no more channel resource is available o saisfy he need of he exra CUs. In Fig. 5, we examine he impac of PU channel uilizaion and he SNR a he CUs on CU video qualiy. In he hree-dimensional plos, he x-axis is he minimum channel idle probabiliy, i.e., min j { P ( r j 0 }, and he y-axis is he minimum SNR of CUs, i.e., min i,j { μ }. I can be observed from he figure ha as channel uilizaion is decreased, a channel has a higher probabiliy of being idle and here will be more channels available for CUs in he ransmission phase. Thus he average MOS sum of he CUs is improved. Furhermore, i follows Since φ and θ are he MOS gain when channel j is idle and busy a ime slo, respecively, we have (φ θ > 0. Therefore, w is an increasing funcion of P ( rj 0 and he overall MOS sum is improved wih P ( rj 0. On he oher hand, an increased minimum SNR a he CUs leads o a higher daa rae (i.e., a higher SBR in (11, and resuls in a higher MOS value for he CUs according o he MOS model given in (11. We also find PS1 ouperforms { } Benchmark { ( scheme 2 for he enire range of min i,j μ and minj P r j 0 } in erms of he average MOS sum over an GOP window in his simulaion. We nex show how he raffic load affecs he performance of PS1 in Fig. 6. In paricular, we simulaed wo raffic loads, i.e., ξ 0.5 and ξ 0.9, and plo he disribuion of he CU MOS values. The enire MOS range (1 o 5 is evenly divided ino 4 ranges wih uni spans, and he number of CU MOS falling ino each range is ploed in he sacked manner. We find ha when he raffic load is ligh, mos of he acive CUs ge he opporuniy o receive video daa, hus yielding a comparaively higher MOS value in his case. When he raffic load is heavy, he amoun of idle channels becomes lower han he amoun of acive CUs, and hus some CUs are no scheduled for video sreaming. The proposed scheme ouperforms he benchmark schemes in boh cases. In he following we examine he performance of he proposed wo-sep approach for he muli-channel sensing phase, which is ermed PS2 in he simulaions. In Figs. 7 and 8, we plo he number of idle channels deeced and he achieved MOS values. We use a modified version of he GPA ha solves he OAPSS problem in Secion III-A.1 as a benchmark scheme where he channel allocaion sraegy is he same as ha of our proposed wo-sep approach, since we only wan o show he performance of he proposed wo-sep approach. We change he algorihm by
HE e al.: QOE DRIVEN MULTI-USER VIDEO STREAMING IN CELLULAR CRNS WITH SINGLE CHANNEL ACCESS 1411 Fig. 6. Disribuion of he CU MOS values ha fall wihin a cerain region under wo cases: raffic load ξ 0.5, and raffic load ξ 0.9. Fig. 8. MOS performance comparison beween PS2 and GPA. Fig. 7. Sensing performance comparison beween PS2 and GPA. leing he righ hand side of consrain (9 be C, which is he specrum sensing capaciy of each CU, as N 1 x C, for all i, and add a consrain M x Λ. This way, since all he P s are idenical, for all i, j,, all he ϕ s are also idenical, for all i, j,, and he algorihm will choose MC/Λ channels o sense. In Fig. 7 we can see ha he number of idle channels deeced by PS2 is considerably greaer han ha by GPA. This is because in PS2, channels wih a greaer probabiliy of being idle will have a higher probabiliy of being sensed, while in GPA, all channels, regardless of he probabiliy of being idle, have he same probabiliy of being sensed. I is common sense ha if a channel has a high probabiliy of being idle, hen i will have a high probabiliy of being found idle by specrum sensing. Noe ha in he muli-channel sensing case, a mos a number of MC/Λ channels will be sensed a a ime slo. In Fig. 8, we compare he MOS performance of PS2 and he channel assignmen algorihm as in SP2 combined wih GPA. Since he proposed scheme ends o find more idle channels han GPA does, more channels will be used for video sreaming, leading o beer QoE performance. This resul also validaes Fig. 9. Comparison of fairness performance beween channel allocaion sraegy considering fairness among CUs and channel allocaion sraegy wihou considering fairness among CUs, for differen min j P (r j 0. he fac ha he proposed wo-sep approach which reas he specrum-sensing-and-accessing-join-opimizaion problem as an inac problem and solves for he joinly opimized sensing and accessing sraegy, will achieve he opimaliy, he approach of decoupling he join-opimizaion problem ino wo subproblems, as Secion III does, will los opimaliy in some exen. Finally, we examine in Fig. 9 he fairness performance of he channel allocaion sraegy considering fairness among CUs, i.e., P1 and P2 (see (20 and (23, respecively. We adop Jain s fairness index as in [38], [39]: f(e 1,e 2,...,e M ( M e i 2 /(M M e2 i, where e i is he average MOS of CU i during a period of ime (10 GOP windows in our simulaion, i 1, 2,...,M. The fairness index ranges from 0 (wors o 1 (bes. The benchmark scheme is he channel allocaion sraegy wihou considering fairness among CUs, i.e., OAPVT. I can be seen ha P1 and P2 can achieve an improve of over 0.1 on fairness performance over OAPVT in his simulaion. VI. RELATED WORK This paper is relaed o he prior work on qualiy of service (QoS and QoE provisioning and video sreaming over CRNs. We briefly review he relaed work in he following.
1412 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 18, NO. 7, JULY 2016 CR research has been largely focused on he aspecs of specrum sensing and dynamic specrum access [12]. In [13], he auhors sudy he sensing-hroughpu radeoff problem ha opimizes he specrum sensing ime so ha he CU s hroughpu can be maximized wih resriced inerference o he PUs. Unlike [13], he proocol proposed in [14] also considers he problem of which channel o sense, in addiion o sensing parameers and access sraegy opimizaion. Moreover, i is shown ha he design of sensing sraegy is independen o sensing parameers design and he access sraegy, as specified in a principle of separaion [14]. These works focus on he opimizaion of sensing parameers only, and here is no collaboraion beween CUs. Considering he fac ha differen CUs may have differen specrum sensing performance, he algorihm proposed in [15] forms groups of CUs for cooperaive sensing, aiming o find he bes grouping scheme o discover mos idle channels. Moreover, he problem of sensing parameer opimizaion in addiion o opimal sensor selecion is invesigaed in [16], wih he objecive o achieve a rade-off beween deecion performance and sensing overhead. The problem of video sreaming over CRNs has been sudied in a few prior works. The ransmission of mulimedia over CRN is firs proposed by Miola in [17]. In [3] he qualiy opimizaion problem is formulaed as a mixed ineger nonlinear programming problem and solved wih effecive algorihms. Auhors of [18] develop an aucion game model o deliver conen-aware mulimedia. The auhors in [19] consider he scenario where mulimedia ransmission is scheduled in CRN and a QoE Driven channel allocaion scheme is proposed o opimize he mulimedia ransmission of prioriy-based CUs, where he MOS model proposed in [11] is used. Specifically, each CU has differen QoE requiremens and hus has differen prioriy in uilizing he idle channels of he PU sysem. Upon he re-appearance of an acive PU on he idle licensed channel, each CU uilizing he idle licensed channels will evacuae from he curren channel i is using o avoid conflic wih he acive PU. The auhors of [34] propose a a learning-based QoE-driven specrum handoff scheme for mulimedia ransmissions over CR neworks. Reinforcemen learning is incorporaed in he specrum handoff scheme o maximize he QoE of video ransmissions in he long erm. The proposed learning scheme is asympoically opimal, model-free, and can adapively perform specrum handoff for he changing channel condiions and raffic load. To exend he video sreaming ime for he CUs, he auhors of [35] propose a flexible sensing scheme o reduce he need for unnecessary channel sensing. Furhermore, he nework absracion layer unis in he SVC video are assigned uiliies which accuraely reflec heir conribuions o he video qualiy, and differen layers are sreamed over differen channels based on heir conribuions o maximizing he oal uiliy of he received video. In order o comprehensively evaluae he uiliy of he CUs in video sreaming, he auhors of [36] propose o no only consider he video qualiy of each CU, bu also consider he number of saisfied CUs. A hree-dimensional scalable qualiy of he H.264/SVC video ransmission problem is formulaed and solved wih an subopimal soluion. In [37], he auhors consider he case ha he fuure Inerne nework may become highly heerogeneous, and herefore an efficien cogniive nework managemen is proposed for he opimizaion of nework operaions like managemen of resources, mobiliy or QoS in order o ensure smooh nework operaion and high user saisfacion. VII. CONCLUSION In his paper, we invesigaed he problem of QoE-aware video sreaming over CRNs where each CU can access one channel a a ime. For he case where each CU can sense and access a mos one channel a a ime, we formulaed an IP problem on specrum sensing and solved i wih a opimal Greedy Poly-maching Algorihm. We hen formulaed a channel assignmen problem and solved i wih he Hungarian Mehod ha is also opimal wih respec o QoE of he muli-user videos. For he case where each CU can sense muliple channels bu access only one channel, we presened a more general, inegraed formulaion. Based on an assumpion on he specrum sensor configuraion, we developed a wo-sep approach o solve he inegraed problem and proved is opimaliy. The proposed schemes were shown o ouperform several alernaive schemes in he simulaion sudy. REFERENCES [1] Cisco Sys., Inc., San Jose, CA, USA, Visual Neworking Index (VNI, Feb. 2014. [Online]. Available: hp://www.cisco.com/ [2] Z. He, S. Mao, and S. Kompella, QoE driven video sreaming in cogniive radio neworks: Case of single channel access, in Proc. IEEE Global Telecommun. Conf., Dec. 2014, pp. 1388 1393. [3] D. Hu and S. Mao, Sreaming scalable videos over muli-hop cogniive radio neworks, IEEE Trans. Wireless Commun., vol. 11, no. 9, pp. 3501 3511, Nov. 2011. [4] K. Yamagishi and T. Hayashi, Opinion model using psychological facors for ineracive mulimodal services, IEICE Trans. Commun.,vol.E89-B, no. 2, pp. 281 288, Feb. 2006. [5] J. You e al., Percepual-based qualiy assessmen for audio-visual services: A survey, Signal Process. Image Commun., vol. 25, no. 7, pp. 482 501, Aug. 2010. [6] A. Khan, L. Sun, and E. Ifeachor, Conen clusering based video qualiy predicion model for MPEG4 video sreaming over wireless neworks, in Proc. IEEE In. Conf. Commun., Jun. 2009, pp. 1 5. [7] K. Yamagishi, T. Tominaga, T. Hayashi, and A. Takashi, Objecive qualiy evaluaion model for videophone services, NTT Tech. Rev., vol. 5, no. 6, pp. 1 5, Jun. 2007. [8] A. Eden, No-reference image qualiy analysis for compressed video sequences, IEEE Trans. Broadcas., vol. 54,no.3,pp.691 697,Sep. 2008. [9] Q. Huynh-Thu and M. Ghanbari, Temporal aspec of perceived qualiy in mobile video broadcasing, IEEE Trans. Broadcas., vol. 54, no. 3, pp. 641 651, Sep. 2008. [10] M. S. Bazaraa, J. J. Jarvis, and H. D. Sherali, Linear Programming and Nework Flows, 4h ed., New York, NY, USA: Wiley, 2009. [11] A. Khan, L. Sun, and E. Ifeachor, QoE predicion model and is applicaion in video qualiy adapaion over UMTS neworks, IEEE Trans. Mulimedia, vol. 14, no. 2, pp. 431 442, Apr. 2012. [12] Y. Zhao, S. Mao, J. Neel, and J. H. Reed, Performance evaluaion of cogniive radios: Merics, uiliy funcions, and mehodologies, Proc. IEEE, vol. 97, no. 4, pp. 642 659, Apr. 2009. [13] Y.C. Liang, Y. Zheng, E. Peh, and A. T. Hoang, Sensing-hroughpu radeoff for cogniive radio neworks, IEEE Trans. Wireless Commun., vol. 7, no. 4, pp. 1326 1337, Apr. 2008. [14] Y. Chen, Q. Zhao, and A. Swami, Join design and separaion principle for opporunisic specrum access in he presence of sensing errors, IEEE Trans. Inf. Theory, vol. 54, no. 5, pp. 2053 2071, May 2008. [15] W. Wang, B. Kasiri, J. Cai, and A.S. Alfa, Channel assignmen of cooperaive specrum sensing in muli-channel cogniive radio neworks, in Proc. IEEE In. Commun. Conf., Jun. 2011, pp. 1 5. [16] A. W. Min and K. G. Shin, Join opimal sensor selecion and scheduling in dynamic specrum access neworks, IEEE Trans. Mobile Compu., vol. 12, no. 8, pp. 1532 1545, Aug. 2013. [17] J. Miola III, Cogniive radio for flexible mobile mulimedia communicaions, Mobile New. Appl. J., vol. 6, no. 5, pp. 435 441, Sep. 2001.
HE e al.: QOE DRIVEN MULTI-USER VIDEO STREAMING IN CELLULAR CRNS WITH SINGLE CHANNEL ACCESS 1413 [18] Y. Chen, Y. Wu, B. Wang, and K. J. R. Liu, Specrum aucion games for mulimedia sreaming over cogniive radio neworks, IEEE Trans. Commun., vol. 58, no. 8, pp. 2381 2390, Aug. 2010. [19] T. Jiang, H. Wang, and A. Vasilakos, QoE-driven channel allocaion schemes for mulimedia ransmission of prioriy-based secondary users over cogniive radio neworks, IEEE J. Sel. Areas Commun., vol. 30, no. 7, pp. 1215 1224, Aug. 2012. [20] W. Wang, B. Kasiri, J. Cai, and A. S. Alfa, Disribued cooperaive mulichannel specrum sensing based on dynamic coaliional game. in Proc. IEEE Global Telecommun. Conf., Dec. 2010, pp. 1 5. [21] B. Wang, K. J. Ray Liu, and T. C. Clancy, Evoluionary cooperaive specrum sensing game: How o collaborae? IEEE Trans. Commun., vol. 58, no. 3, pp. 890 900, Mar. 2010. [22] S. Zhang, T. Wu, and R. K. N. Lau, A low-overhead energy deecion based cooperaive sensing proocol for cogniive radio sysems, IEEE Trans. Wireless Commun., vol. 11, no. 8, pp. 5575 5581, Nov. 2009. [23] K. Kar, L. Xiang, and S. Sarkar, Throughpu-opimal scheduling in mulichannel access poin neworks under infrequen channel measuremens, IEEE Trans. Wireless Commun., vol. 7, no. 7, pp. 2619 2629, Jul. 2008. [24] H. W. Kuhn, The Hungarian mehod for he assignmen problem, Naval Res. Logisics Quarerly, vol. 2, nos. 1/2, pp. 83 97, Mar. 1955. [25] H. Su and X. Zhang, Cross-layer based opporunisic MAC proocols for QoS provisionings over cogniive radio wireless neworks, IEEE J. Sel. Areas Commun., vol. 26, no. 1, pp. 118 129, Jan. 2008. [26] R. Garfinkel and G. L. Nemhauser, Ineger Programming. New York, NY, USA: Wiley, 1972. [27] Y. Xu and S. Mao, User associaion in massive MIMO HeNes, IEEE Sys. J., o be published, doi: 10.1109/JSYST.2015.2475702. [28] I. Griva, S. G. Nash, and A. Sofer, Linear and Nonlinear Opimizaion, 2nd ed., Philadelphia, PA, USA: SIAM, 2009. [29] D. Hu and S. Mao, A sensing error aware MAC proocol for cogniive radio neworks, EAI Endorsed Trans. Mobile Commun. Appl., vol. 12, no. 7/9, pp. e1 e12, Aug. 2012. [30] L. Zhou e al., Resource allocaion wih incomplee informaion for QoEdriven mulimedia communicaions, IEEE Trans. Wireless Commun., vol. 12, no. 8, pp. 3733 3745, Aug. 2013. [31] A. Khan, I. H. Mkwawa, L. Sun, and E. Ifeachor, QoE-driven sender birae adapaion scheme for video applicaions over IP mulimedia subsysem, in Proc. IEEE In. Conf. Commun., Jun. 2011, p pp. 1 6. [32] A. Khan, L. Sun, E. Jammeh, and E. Ifeachor, Conen classificaionbased and QoE-driven video send birae adapaion scheme, in Proc. 5h In. ICST Mobile Mulimedia Commun. Conf., Sep. 2009, pp. 16 19. [33] W. Song and D. Tjondronegoro, Accepabiliy-based QoE models for mobile video, IEEE Trans. Mulimedia, vol. 16, no. 3, pp. 738 750, Mar. 2014. [34] Y. Wu, F. Hu, S. Kumar, and Y. Zhu, A learning-based QoE-driven specrum handoff scheme for mulimedia ransmissions over cogniive radio neworks, IEEE J. Sel. Areas Commun., vol. 32, no. 11, pp. 2134 2148, Nov. 2014. [35] R. Yao e al., Uiliy-based H.264/SVC video sreaming over mulichannel cogniive radio neworks, IEEE Trans. Mulimedia, vol. 17, no. 3, pp. 434 449, Mar. 2015. [36] H. Saki, M. G. Marini, and M. Shikh-Bahaei, Muli-user scalable video ransmission over cogniive radio neworks, in Proc. IEEE In. Conf. Commun., Jun. 2015, pp. 7564 7569. [37] T. Ojanpera e al., Cogniive nework managemen framework and approach for video sreaming opimizaion in heerogeneous neworks, Wireless Pers. Commun., vol. 84, no. 3, pp. 1739 1769, Mar. 2015. [38] I. Son, S. Mao, M. X. Gong, and Y. Li, On frame-based scheduling for direcional mmwave WPANs, in Proc. IEEE INFOCOM,Mar.2012,pp. 2149 2157. [39] I.-K. Son e al., Frame-based medium access conrol for 5G wireless neworks, Mobile New. Appl., vol. 20, no. 6, pp. 763 772, Dec. 2015. Zhifeng He received he B.S. degree in elecronics informaion science and echnology from he Shandong Universiy of Technology, Zibo, China, in 2009, he M.S. degree in microelecronics and solid-sae elecronics from he Being Universiy of Poss and Telecommunicaions, Being, China, in 2012, and is currenly working oward he Ph.D. degree in elecrical and compuer engineering a Auburn Universiy, Auburn, AL, USA. His curren research ineress include cogniive radio, mm-wave communicaions and neworking, mulimedia communicaions, and opimizaion. Mr. He was he recipien of he IEEE GLOBECOM 2015 Bes Paper Award. Shiwen Mao (S 99 M 03 SM 09 received he Ph.D. degree in elecrical and compuer engineering from Polyechnic Universiy, Brooklyn, NY, USA. He is currenly he Samuel Ginn Disinguished Professor wih he Deparmen of Elecrical and Compuer Engineering, Auburn Universiy, Auburn, AL, USA. His research ineress include wireless neworks and mulimedia communicaions. Prof. Mao is a Disinguished Lecurer of he IEEE Vehicular Technology Sociey in he Class of 2014. He is on he Ediorial Board of he IEEE TRANSAC- TIONS ON MULTIMEDIA, he IEEE INTERNET OF THINGS JOURNAL, he IEEE COMMUNICATIONS SURVEYS AND TUTORIALS,heIEEE MuliMedia Magazine, ec. He is he Chair of he IEEE ComSoc Mulimedia Communicaions Technical Commiee. He was he recipien of he 2015 IEEE ComSoC TC-CSR Disinguished Service Award, he 2013 IEEE ComSoc MMTC Ousanding Leadership Award, he NSF CAREER Award in 2010, he IEEE GLOBECOM 2015 Bes Paper Award, he IEEE WCNC 2015 Bes Paper Award, he IEEE ICC 2013 Bes Paper Award, and he 2004 IEEE Communicaions Sociey Leonard G. Abraham Prize in he field of communicaions sysems. Sasry Kompella (S 04 M 06 SM 12 received he B.E. degree in elecronics and communicaion engineering from Andhra Universiy, Vishakhapanam, India, in 1996, he M.S. degree in elecrical engineering from Texas Tech Universiy, Lubbock, TX, USA, in 1998, and he Ph.D. degree in elecrical and compuer engineering from he Virginia Polyechnic Insiue and Sae Universiy, Blacksburg, VA, USA, in 2006. He is currenly he Head of he Wireless Nework Research Secion under he Informaion Technology Division, U.S. Naval Research Laboraory, Washingon, DC, USA. His research ineress include wireless neworks, from mobile ad hoc o underwaer acousic neworks, wih specific focus oward he developmen of cogniive and cooperaive nework opimizaion echniques for efficien wireless link scheduling.