Throughput Optimization in High Speed Downlink Packet Access (HSDPA)

Transcription

1 Throughput Optimizatio i High Speed Dowlik Packet Access (HSDPA) Tao Cui Departmet of Electrical Egieerig Califoria Istitute of Techology Pasadea, CA 925, USA taocui@caltech.edu Feg Lu, Ail Goteti, V. Sethurama, S. P. Rao ad P. Subrahmaya Qualcomm Ic. 365 Kifer Road Sata Clara, CA 9505, USA feglu, agoteti, vigeshs, sprao, psubrahm}@qualcomm.com Abstract I this paper, we ivestigate throughput optimizatio i High Speed Dowlik Packet Access (HSDPA). Specifically, we propose offlie ad olie algorithms for adjustig the Chael Quality Idicator (CQI) used by the etwork to schedule data trasmissio. I the offlie algorithm, a give target BLER is achieved by adjustig CQI based o ACK/NAK history. By sweepig through differet target BLERs, we ca fid the throughput optimal BLER offlie. This algorithm could be used ot oly to optimize throughput but also to eable fair resource allocatio amog mobile users i HSDPA. I the olie algorithm, the CQI offset is adapted usig a estimated short term throughput gradiet without specifyig a target BLER. A adaptive stepsize mechaism is proposed to track temporal variatio of the eviromet. We ivestigate covergece behavior of both algorithms. Simulatio results show that the proposed offlie algorithm ca achieve the give target BLER with good accuracy. Both algorithms yield up to 30% HSDPA throughput improvemet over that with 0% target BLER. I. INTRODUCTION The success of 3rd geeratio wireless cellular etworks is maily based o efficiet provisioig of the expected wide variety of services requirig differet Quality of Service with respect to data rate, delay ad error rate. I order to improve support for high data rate packet switched services, 3GPP has developed a evolutio of UMTS based o WCDMA kow as High Speed Dowlik Packet Access (HSDPA) which was icluded i the Release 5 specificatios ]. HSDPA targets icreased capacity, reduced roud trip delay, ad higher peak dowlik (DL) data rates. Evolutios of HSDPA featurig data rates up to 84 Mbps are uder developmet. I HSDPA, the user equipmet (UE) (also kow as mobile statio) moitors the quality of the dowlik wireless chael ad periodically reports this iformatio to the base statio (referred to here as NodeB) o the uplik. This feedback, called Chael Quality Idicator (CQI), is a idicatio of the highest data rate that the UE ca reliably receive i the existig coditios o the dowlik wireless chael. The frequecy of reportig CQI is cofigured by the etwork, ad is typically set to oce every few millisecods. Usig the chael quality reports, the NodeB accordigly schedules data o the High Speed Physical Dowlik Shared Chael (HS-PDSCH). The NodeB s selectio of the trasport block The views ad coclusios cotaied i this paper are those of the authors ad should ot be iterpreted as represetig the official policies, either expressed or implied, of Qualcomm Ic. size (umber of iformatio bits per packet), umber of chaelizatio codes, modulatio ad resource allocatio choices such as HS-PDSCH trasmit power allocatio are guided by the NodeB s iterpretatio of the reported CQI. CQI reports are iteded to accurately reflect the HS- PDSCH performace that the UE ca support i the existig wireless chael coditios. It is recommeded i 2] that, i static chael coditios, the UE report CQI such that it achieves a block error rate (BLER) close to 0% whe scheduled data correspodig to the media reported CQI. I practice, the accuracy of CQI reports i reflectig HS-PDSCH performace is iflueced by the wireless chael coditios such as the speed of the mobile user ad the dispersive ature of the chael. Achievig a certai target BLER at a give scheduled data rate requires differet average HS-DSCH SNR uder differet chael coditios. Also, the NodeB ofte uses differet trasport block sizes, umber of codes ad modulatio, collectively referred to as the trasport format resource combiatio (TFRC), to achieve similar data rates. The exact choice of TFRC that the NodeB uses affects the required HS-PDSCH SNR to achieve a certai target BLER. These variabilities may cause the actual BLER to deviate from the 0% target. Moreover, the 0% target BLER may ot yield maximum throughput uder all coditios of the wireless chael. The cell throughput optimizatio i HSDPA ca be cosidered a two part problem: oe is code ad power allocatio across users, ad the other is maximizig the lik throughput for each user for a give resource allocatio. I this paper, we focus o the lik throughput optimizatio ad cosider throughput optimizatio through simple adjustmets to the reported CQI. We propose offlie ad olie algorithms for adjustig the CQI. I the offlie algorithm, we first propose a adaptive algorithm to achieve a give target BLER usig the stochastic gradiet descet method, which adjusts the CQI offset adaptively based o the short term BLER obtaied from the ACK/NACK history. By searchig through differet target BLERs, we ca fid the throughput optimal BLER offlie. The proposed algorithm ca be implemeted at the UE as well as at the Node B. Whe applied at the Node B, i additio to achievig the target BLER, it ca also save trasmit power. This algorithm could be used ot oly to refie CQI-BLER aligmet but also to eable fair resource allocatio amog mobile users i HSDPA. Stadard stochastic approximatio /09/$ IEEE

2 (SA) algorithms typically require a decreasig stepsize 3]. We show the covergece of the offlie algorithm with a costat stepsize. I the olie algorithm, we use a variatio of the Kiefer- Wolfowitz algorithm 4] i SA, which does ot eed to specify a target BLER. The CQI offset is adapted gradually usig a estimated short term throughput gradiet. Ulike 4], the stepsize i the proposed algorithm does ot decrease to zero. I additio, a adaptive stepsize mechaism is proposed to track temporal variatio of the eviromet. With a costat stepsize, we show that the proposed olie algorithm coverges to a small eighborhood of the local optimal solutio. Our simulatio results show that the proposed offlie algorithm ca achieve the give target BLER with good accuracy. Both throughput optimizatio algorithms are show to improve the throughput by up to 30% i simulatio. The throughput optimal BLER is calculated for popular chael path profiles. I geeral, the throughput optimal BLER is ot always 0% ad depeds o the chael path profile. For AWGN chaels, it is about 0%, as is implied i 5]. Cosiderig that the UE implemetatio i the simulatio closely mirrors commercially shippig devices ad already icludes several receiver optimizatios, the additioal gai obtaied through the algorithm is idicative of potetial HSDPA throughput ehacemet realizable i practice. II. LINK THROUGHPUT EXPRESSION IN HSDPA Throughput i HSDPA is the rate of trasfer of iformatio bits over the wireless chael i uits of bits per subframe. For a give resource allocatio across users, the etwork may schedule differet TFRCs to a user depedig o the perceived chael coditios of the user. We model the schedular usig a 30 etry table of CQI vs TFRC such as 2, Table 7D]. For a give SNR g i db, let p cqi (g) deote the probability that a ew trasmissio with TFRC correspodig to CQI idex cqi is scheduled, b cqi (g) be the first BLER, the BLER correspodig to the first trasmissio, whe the TBS correspodig to cqi is scheduled ad TBS cqi deote the TBS correspodig to cqi. We assume that the secod BLER (retrasmissio BLER) is zero. I the followig, BLER deotes first BLER. Assumig that N subframes are to be received by UE, the average umber of ew trasmissios with TBS cqi is p cqi (g)n. The average umber of subframes required to sed p cqi (g)n subframes is p cqi (g)(+b cqi (g))n, assumig the secod BLER is zero. Therefore, the average throughput at SNR g is 30 cqi= pcqi(g)tbscqi T (g)= 30 cqi= pcqi(g)(+bcqi(g)). () We model the estimated CQI as a radom variable with mea m(g)=g+κ, where κ is a fixed offset, ad probability desity fuctio (pdf) f(cqi,m(g)). I practice, we fid that the Gaussia distributio is a reasoable approximatio of the real CQI distributio, i.e., (cqi m(g)) 2 f(cqi,m(g))= e σ 2 (g). (2) 2πσ2 (g) Give g, we ca compute p cqi (g) as + f(x,m(g))dx, if cqi=30, cqi 0.5 cqi+0.5 p cqi(g)= f(x,m(g))dx, if cqi=, (3) cqi+0.5 f(x,m(g))dx, otherwise. cqi 0.5 Whe we apply a CQI offset Δ o κ or shift the distributio by Δ, (3) is modified to p cqi(δ)= cqi+0.5 cqi 0.5 f(x Δ,m(g))dx= cqi+0.5 Δ cqi 0.5 Δ f(x,m(g))dx, (4) for cqi =,30 ad T (g) i () will also be a fuctio of Δ, i.e., T (Δ). I the followig, for clarity i writig expressios such as T (Δ), we omit the depedece o g. I fadig chaels, the average throughput T (Δ) is a average of () over all fadig states ad it depeds o the average SNR. Use of CQI adjustmet mechaisms i commercial etworks has bee reported i the literature 5]. Applicatio of a CQI offset Δ is oe such mechaism. The objective of this paper is to maximize the average throughput over Δ, i.e., maxt (Δ). (5) Δ Note that whe b cqi (g) is kow for example by experimet ad curve fittig. We ca solve (5) umerically. III. OFFLINE THROUGHOUT OPTIMIZATION A. Algorithm The algorithm compares the curret short term BLER with the target BLER ad updates the CQI offset accordig to Δ += Δ α(ˆb(δ ) b )] U, (6) where Δ is the CQI offset at the -th iteratio, ˆb (Δ ) is the short term BLER at the -th iteratio, ad ] U deotes the mappig oto set U to limit the rage of CQI offset. For example, we could choose U = 2,2] to limit the CQI offset betwee 2 ad 2. Though the mappig ] U is ot ecessary to achieve covergece, we use it to avoid udesired large CQI offset. The short term average BLER ˆb(Δ) is estimated usig the ACK/NACK history withi a slidig widow of size w. Let t be the startig time of the -th widow for CQI offset update with t =, ad X(t) ad Y (t) deote the status of the first trasmissio ACK ad NACK i subframe t, respectively, where X(t)= if a first trasmissio ACK is received ad X(t)=0 otherwise, ad Y (t)= if a first trasmissio NACK is received ad Y (t)=0 otherwise. After choosig t, t + is chose such that the umber of first time trasmissios betwee t + ad t + is w, i.e., t + t=t + (X(t)+Y (t))=w. The short term first time BLER is the estimated by t+ t=t ˆb(Δ)= Y (t) +. (7) w As the status of ACK/NACK is available at both Node B ad UE, the CQI update algorithm ca be implemeted at both sides. At the UE, sice uquatized (or raw) CQI is available, we could add Δ to the raw CQI, CQI raw, directly, i.e., the reported CQI is the quatizatio of Δ +CQI raw. However, i high geometry, CQI raw may be greater tha 30. Because

3 the maximum CQI is 30, UE should report 30 eve though the curret chael coditio could support a data rate (TBS) higher tha that correspodig to CQI 30. I this case, we caot achieve the target first BLER o matter what CQI offset is applied because b cqi (g)<b, cqi,...,30} ad b(δ) i (8) is less tha b for all Δ. This problem ca be resolved by usig the Node B algorithm. At Node B, oly reported quatized CQI from UE is available. After obtaiig CQI offset Δ, it performs quatizatio Q(Δ +CQI reported ) ad trasmits TBS correspodig to the CQI after quatizatio, where CQI reported is the reported CQI from UE ad Q( ) is the CQI quatizatio fuctio which maps its iput to a iteger betwee ad 30. To get a fie cotrol of the BLER ad throughput ad achieve the target BLER eve at high SNR, we ca also chage the power of the HS-PDSCH chael by usig the residual CQI, i.e., the HS-PDSCH chael power is reduced by (Δ+CQI reported ) Q(Δ+CQI reported ). By applyig this method at Node B, whe Δ+CQI reported >30, the HS-PDSCH power is reduced such that there exists a cqi such that b cqi (g)>b. Thus, the target BLER ca still be achieved. This algorithm also leads to power savig at Node B. From the system poit of view, this meas that Node B could support more users ad the overall system s performace is improved. The proposed algorithm i this subsectio ca be used to eable fair resource allocatio amog UEs by assigig a fixed target BLER for each UE. I this case, eve though a UE may report a higher CQI, the proposed algorithm will add a egative CQI offset to the reported CQI such that the scheduled TBS achieves the target BLER. Thus, the Node B actually moitors whether the reported CQI from each UE matches the curret chael coditio the UE suffers. B. Mootoicity of Average First BLER To derive the algorithm (6), we first show that the first BLER is a mootoic fuctio i the CQI offset Δ. The average first BLER ca be writte as 30 b(δ)= p cqi(δ)b cqi(g). (8) cqi= Takig the derivative of b(δ) with respect to Δ ad takig ito accout the expressio of p cqi (Δ) i (3), we obtai 30 db(δ) = f(cqi 0.5 Δ,m(g))(b cqi(g) b cqi (g)). (9) cqi=2 Due to CQI table desig, a higher cqi correspods to a higher TBS, which cosists of higher modulatio ad codig rate. Typically, b cqi (g)>b cqi (g), db(δ) >0 ad b(δ) is a strictly icreasig fuctio i Δ. Thus, i this case, give throughput optimal CQI offset Δ, there exists a uique BLER b such that b(δ )=b. This result also holds i fadig chaels after averagig over all fadig states. As b(δ) is mootoe, there is a oe-to-oe correspodece betwee BLER ad CQI offset. Therefore, fidig the throughput optimal BLER is equivalet to fidig the optimal CQI offset. I the followig, we develop adaptive algorithms to achieve a give target first BLER. The optimal throughput ca the be foud by searchig all possible target BLERs. C. Algorithm Derivatio I this subsectio, we derive the adaptive algorithm i (6) that adjusts the CQI offset Δ to achieve the target BLER b by miimizig the mea-squared error (MSE) betwee b(δ) ad b, i.e., f(δ)= b(δ) b 2. We eed to solve mi Δ b(δ) b 2. (0) To solve (0), we use gradiet descet method via df (Δ) Δ +=Δ α =Δ 2 α db(δ) (b(δ) b ), () where α>0 is a stepsize. As b cqi (g) ad σcqi 2 are geerally ukow, it is hard to compute db(δ) especially i fadig chaels where the BLER should be averaged ( over ) all the fadig states. We replace db(δ) with sig db(δ) =, where sig( ) deotes the sig of its iput. We thus modify () to ( ) db(δ) Δ +=Δ α(b(δ) b )sig =Δ α(b(δ) b ), (2) where α>0 is a stepsize. We fid that this modificatio works well i practice eve though the covergece rate may be slower as compared to (). Exact implemetatio of (2) requires the kowledge of b(δ), which is hard to obtai olie. We propose replacig the log term BLER b(δ) i (2) with the short term BLER ˆb (Δ ), which gives (6). D. Covergece Aalysis I this subsectio, we aalyze the algorithm (6) without cosiderig the mappig ] U. The aalysis ca be exteded to the case with ] U as i 3]. We oly give the results without proof due to the space limitatio. Let M max deote the maximum value of db(δ) over all possible Δ. We ca show that if 0<α< 2 M max, there exists a costat 0<η< such that EΔ +} Δ <η EΔ } Δ. (3) Therefore, EΔ } coverges to Δ expoetially ad a larger stepsize α idicates a faster covergece rate whe 0<α< M max. Secod order aalysis shows that E(Δ Δ ) 2 } also coverges, ad its limitig poit satisfies lim E Δ + Δ 2} α < (2+αδ)4wδ. (4) where δ>0 depeds o db(δ). The previous two results cocer the covergece of Δ, which is the mai focus of SA 3]. I our problem, we do ot really care about whether Δ coverges. We are oly iterested i whether the time average of ˆb (Δ ) coverges. By usig the martigale iequality 6], we get lim ˆbi(Δ i)=b with probability. (5) i= IV. ONLINE THROUGHOUT OPTIMIZATION I this sectio, we develop a olie throughput optimizatio algorithm, which does ot eed to specify a target BLER. The olie algorithm works for ay CQI table ad ay SNR.

4 A. Algorithm To maximize the average throughput T (Δ) i (5), we cosider usig gradiet descet method to fid the maximum throughput ad its correspodig CQI offset. We assume that g is fixed util the algorithm coverges. Let Δ be the CQI dt (Δ) offset at the -th iteratio ad be the derivative of T (Δ) with respect to Δ. By usig gradiet descet method, the CQI offset update is give by Δ +=Δ +α dt (Δ ), (6) where α >0 is a stepsize. Usig (6) directly requires the kowledge of both the exact derivative ad the average throughput fuctio. We propose to replacig (6) with Δ ˆT +=Δ +α (Δ +ɛ) ˆT (Δ ), (7) ɛ where ˆT (Δ ) is a estimate of T (Δ ). As i Sectio III- C, we use a widow based algorithm to estimate the average throughput. Let TBS(t) be the TBS trasmitted i subframe t ad R(t) ad F (t) deote the status of the ACK/NACK, where R(t)= if a ACK is received ad R(t)=0 otherwise (icludig NTX subframes), ad F (t)= if a ACK or NACK is received ad F (t)=0 otherwise. Suppose that the CQI offset Δ is updated every w subframes ad the average throughput T (Δ ) is computed usig the past w received subframes. At the begiig of the -th iteratio, we compute T (Δ ) via w t=( )w+ ˆT (Δ R(t)TBS(t) )= w t=( )w+ F (t). (8) From experimets, we fid that the choice of α is very importat to the algorithm. We suggest choosig α = ᾱ T, where T is the average throughput up to the -th iteratio, i= i.e., T = ˆT (Δ i). The, (7) become α ˆT (Δ+ɛ) Δ += Δ ˆT ] S U (Δ ) +, (9) T ɛ where ] A deotes the mappig oto set A, which is used to reduce the effect of the radom variatio of the throughput o the CQI offset update algorithm. For example, we could choose S= 0.,0.] ad U = 2,2] to limit the absolute CQI offset icremet withi 0. ad the absolute CQI offset withi 2. As i (6), the mappig ] A is used to avoid udesired large CQI offset. The algorithm (9) differs from the Kiefer- Wolfowitz algorithm 4] i that (9) uses a fixed α ad ɛ, while the oe i 4] uses decreasig α ad ɛ. I fadig chaels, the throughput may chage over time. To track the throughput variatio, we suggest adaptig α ad ɛ i (9). Let ˆT (Δ )= ˆT (Δ +ɛ) ˆT (Δ ) ɛ, F = Δ α, ad P = Δ ɛ. Igorig the mappig i (9), motivated by 7], we take the partial derivative of both sides with respect to α, which gives F 0 =0 ad ( F += + ᾱ ) ˆT (Δ ) F + T ˆT (Δ ). (20) T Let α deote α i the -th iteratio. We update α usig aother gradiet method as ] V T(Δ ] ) α += α +μ α +μ ˆT V (Δ )F, (2) α dt (Δ) where we have replaced by its short term average ˆT (Δ ), μ >0 is a stepsize ad V is a costrait set, e.g., V =0.000,0.5]. Similarly, takig the partial derivative of both sides of (9) with respect to ɛ, we obtai P 0 =0 ad P ᾱ +=P + (+P ) T ˆT (Δ ). (22) Let ɛ deote the ɛ i the -th iteratio. By usig the gradiet method, we update ɛ as ] V2 T(Δ ] ) ɛ += ɛ +μ 2 ɛ +μ 2 ˆT V2 (Δ )P, (23) α where μ 2 >0 is a stepsize ad V 2 is a costrait set, e.g., V 2 = 0.2,0.2]. I both (20) ad (22), we approximate the secod order derivative ˆT (Δ ) by ˆT (Δ ˆT (Δ ) ˆT (Δ k ) )=, (24) Δ Δ k where k< is the first iteger such that Δ Δ k >ɛ. Equatios (9), (2) ad (23) costitute the olie throughput maximizatio algorithm. To implemet the olie algorithm, we iitially choose Δ =0 ad ɛ to be a small value to iduce the gradiet method, e.g., ɛ = 0.. Ithe-th iteratio, we first evaluate ˆT (Δ ) usig a widow of size w ad the switch to Δ +ɛ to compute ˆT (Δ +ɛ ) usig aother widow. α, ɛ, Δ, F ad P are the updated usig (9)-(23), respectively. To prevet throughput degradatio, we may give a sig of ɛ i the directio of the gradiet ˆT (Δ ). Similar to the offlie algorithm i Sectio III-C, the olie algorithm ca also be implemeted at both Node B ad UE. We have used a simple approximatio of the gradiet. There are other alteratives such as parabolic iterpolatio, see e.g., 3], which may improve the gradiet estimatio but at the expese of higher complexity. The olie algorithm ca also be readily exteded to maximize the average throughput per OVSF code. B. Covergece Aalysis If we allow α ad ɛ decrease as ad α ad ɛ satisfy ɛ 0, α=, αɛ <, α 2 ɛ 2 <, (25) e.g., α = ad ɛ = /3, from 4], Δ coverges to Δ stochastically as. I practical implemetatio, a costat stepsize is typically desired. Existig literature lacks the aalysis of costat stepsize Kiefer-Wolfowitz algorithm. Assumig that 0<m T (Δ) M, i.e., T (Δ) is a cocave fuctio with bouded curvature, we ca show that with fixed α ad ɛ EΔ } coverges to a small eighbor aroud Δ determied by ɛ. A smaller ɛ meas that EΔ } is closer to Δ. However, this is at the expese of a larger MSE. Let N(w) be the upperboud o the variace of ˆT (Δ ), which is a decreasig fuctio i w. We obtai E Δ Δ 2} ɛ2 M 2 ( αm) 2m 2 (2 αm) + 2αN(w) ɛ 2 m(2 αm) + αɛ2 M 2 4m(2 αm). (26)

5 TABLE I ACHIEVED FIRST BLER USING THE OFFLINE ALGORITHM IN DIFFERENT CHANNELS AND GEOMETRIES, WHEN THE TARGET BLER IS SET TO 5%. Geometry (db) AWGN PA3 PB3 VA30 VA % 5.04% 5.02% 5.3% 5.0% 5 5.2% 5.03% 5.02% 5.% 5.09% % 5.03% 5.02% 5.09% 5.05% % 5.02% 5.0% 5.2% 5.07% TABLE II GAIN IN THROUGHPUT AT OPTIMAL BLER RELATIVE TO THROUGHPUT AT 0% BLER IN DIFFERENT CHANNELS AT GEOMETRY=0DB. AWGN PA3 PB3 VA30 VA20 Optimal BLER 5% 5% 20% 90% 90% Throughput gai 0.3% 4.4% 6.6% 32.6% 6.7% From (26), we ca see that for fixed α, E Δ Δ 2} may icrease dramatically by decreasig ɛ. If the first order derivative T (Δ) is also bouded by P, i.e., T (Δ) <P for all Δ, which is the case for HSDPA throughput optimizatio, we ca further show that ET (Δ )} T (Δ ) ɛ MP 2m. (27) Therefore, the average throughput also coverges to a eighborhood of the local maximum throughput T (Δ ). V. SIMULATION RESULTS We evaluate the performace of proposed throughput optimizatio algorithms usig a C++ HSDPA system simulator. The CQI mappig table for UE category 0 specified i 2] is used i all tests. A parameter, geometry G, isusedithe simulatios, which is defied as G= Ior I, where I oc+n or is the sigal power from Node B i the target cell, I oc represets the sigal power received from adjacet cells ad N is the thermal oise power. We cosider AWGN, Pedestria 3kmph (PA3, PB3), ad Vehicular A 30kmph ad 20kmph (VA30, VA20) chaels 8]. The offlie algorithm depeds o two parameters: widow size w ad stepsize α. For a fixed widow size w, the larger the α, the larger the deviatio of the CQI offset, the lower the throughput, ad the faster the covergece. There is a tradeoff betwee the achievable throughput ad the covergece rate, as expected from the aalysis i Sectio III-D. We used w=20 ad α= to geerate the results below. Table I compares the BLER achieved usig the proposed offlie algorithm i differet chaels whe the target BLER is 5%. We ca see that the proposed offlie algorithm ca achieve the target BLER with a high accuracy. I the olie algorithm, the widow size should be chose to be small so as to track the chael variatio whe chael varies fast. A target BLER of 0% is ot always throughput optimal. The throughput optimal BLER depeds o the chael profile. I Table II, we see that the throughput optimal BLER varies from 0% to 90% for differet chael profiles. At 0dB geometry, the throughput optimal BLER is 5% i AWGN ad PA3, 20% i PB3, ad 90% i VA30 ad VA20. I some chaels, the throughput at the optimal BLER is about 30% higher tha the throughput at 0% BLER. The 30% throughput gai is sigificat i practical systems ad it is achieved with low complexity. The etwork should base its choice of target BLER o chael variatios. A ote o our assumptios: i our simulatios, we sed retrasmissios at the same trasmit power as first trasmissios - i.e. there is o power back-off i retrasmissios. The presece of a power back-off i retrasmissios iflueces the throughput optimal BLER i differet path profiles, as explaied below. Retrasmissios at lesser power are sufficiet to recover the data i semi-static chaels. I dyamic chaels, the chael state may chage sigificatly i the time betwee CQI measuremet ad the correspodig dowlik trasmissio. For example, i Vehicular A 30km/h, the chael fadig chages to a idepedet state i roughly 4ms, i.e. 2 subframes. Whe the etwork receives a CQI report ad schedules data accordigly, the chael coditio i which this data is received may be very differet from that suggested by the CQI, ad retrasmissios may be ecessary. Depedig o the power level of retrasmissios relative to the first trasmissio, relyig more o retrasmissios by operatig at a higher first BLER may improve throughput. VI. CONCLUSION We have ivestigated throughput optimizatio i HSDPA usig two adaptive outer loop algorithms. Both algorithms adjust the CQI offset to maximize the throughput. The offlie algorithm used a adaptive algorithm to achieve a give target BLER usig the stochastic gradiet descet method based o the history of ACK/NACK. By searchig through differet target BLERs, the throughput optimal BLER ca be foud offlie. The olie algorithm used a variatio of the Kiefer- Wolfowitz algorithm without specifyig a target BLER. A adaptive stepsize mechaism was also proposed to make the algorithm robust to o-statioary coditio. We have show the covergece of both algorithms with a costat stepsize. Simulatio results show that the proposed algorithms ca achieve up to 30% throughput improvemet over that with 0% target BLER. Iterplay betwee the algorithms proposed here ad other system level optimizatios such as opportuistic schedulig may be a area for further study. REFERENCES ] 3GPP, TR versio 5.0.0, Physical Layer Aspects of UTRA High Speed Dowlik Packet Access, Mar. 29, ] 3GPP, TR versio 7.0.0, Physical layer procedures (FDD), Mar. 22, ] H. J. Kusher ad G. G. Yi, Stochastic Approximatio ad Recursive Algorithms ad Applicatios, 2d ed. Spriger-Verlag, ] J. Kiefer ad J. Wolfowitz, Stochastic estimatio of the maximum of a regressio fuctio, Aals of Mathematical Statistics, vol. 23, o. 3, pp , Sept ] J. Derkse, R. Jase, M. Maijala, ad E. Westerberg, HSDPA performace ad evolutio, Ericsso Review, vol. 3, pp. 7 20, ] L. Breima, Probability Theory. Addiso-Wesley, 968. Republished (99) i Classics of Mathematics. SIAM, Philadelphia. 7] H. J. Kusher ad J. Yag, Aalysis of adaptive step size SA algorithms for parameter trackig, IEEE Tras. Automat. Cotr., vol. 40, o. 8, pp , Aug ] 3GPP, TR 25.0 versio 8.3.0, Techical specificatio group radio access etwork; user equipmet (UE) radio trasmissio ad receptio (FDD), Jue 6, 2008.