PARTICLE SWARM OPTIMIZATION FOR SAGE MAXIMIZATION STEP IN CHANNEL
PARAMETER
ESTIMATION
Harun Bodur, Celal Alp Tunc, Defne Aktas, Vakur B. Erturk, and Ayhan Altintas
Dept. ofElectrical and ElectronicsEngineering, BilkentUniversity, Ankara, Turkey, e-mail: harun@ee.bilkent. edu.tr
Keywords: MIMO; channel estimation; SAGE; particle swarmoptimization.
Abstract
Thispaper presents anapplication of Particle Swarm Op-timization (PSO) in Space Alternating Generalized Ex-pectation Maximization (SAGE)algorithm. SAGE algo-rithm is apowerful tool for estimating channel parame-terslikedelay, angles (azimuth and elevation) of arrival and departure, Doppler frequency and polarization. To demonstrate theimprovementinprocessing time by uti-lizing PSOin SAGE algorithm, the channel parameters areestimated fromasynthetic data and thecomputational expenseof SAGEalgorithm with PSO is discussed. 1 Introduction
Recentworks have shown thatappropriate coding using multiple antennas atthe transmitter and receivercan in-crease the capacity of mobile systems
[5].
Such sys-tems arecalledMultiple
InputMultiple
Output (MIMO) systems. Design and optimization of MIMO systems require realistic model of the propagation channel. In otherwords, model needs tocharacterize the properties of eachpropagating path such asdelay, angles
(azimuth andelevation)
of arrival (AoA) and departure (AoD), Doppler frequency and polarization [6]. To generate ac-curatechannelmodels,
extensivechannelmeasurements andhigh resolution estimation toolsarerequired. Various estimation tools have been usedto estimate the channel parameters such asMultiple Signal
Classifica-tion(MUSIC), Estimation ofSignal
Parametervia Rota-tional InvarianceTechniques (ESPRIT)
and maximum-likelihood (ML) methods like Expectation Maximiza-tion (EM) and SpaceAlternating
Generalized Expecta-tion MaximizaExpecta-tion (SAGE)algorithms
[2]. MLmethodsyield
more accurateresults andprovide higher resolution than othermethods, butcomputational complexity
ishigh dueto the brute force search requiredto find the likeli-hoodmaximizingparameters.The SAGE algorithm is apopular technique for
param-eter estimation. It
updates
eachparametersequentially
and
successively,
soit hasalowercomplexity
andafaster convergence[1].
Thealgorithm
hastwosteps:expecta-tion(E) and maximization (M)steps.
The mostcomputationallyintensive part of the SAGE al-gorithm is inthe M-step, and hence, fast search proce-dures arerequiredtoreduce thecomputational complex-ity. In this paper, we intend to acceleratethe SAGE algo-rithm using PSO for the search procedure. PSO is a pop-ularevolutionary computation technique which is based onintelligence andmovementofswarms [4].Itis shown that PSO isaneffectivealgorithm and hasalow computa-tionalcomplexity for solving optimization problems. By using PSO, SAGE algorithm converges rapidly and this convergence will be showninSection6.
The signal model for MIMO channels is summarizedin Section 2. Section 3 provides a brief introduction to SAGE algorithm. Section 4 explains how PSO can be utilized in SAGEalgorithm. Section 5presents the per-formance of theproposedalgorithm forasynthetic data. Weconclude withsomeremarksinSection7.
2 MIMOSignal Model
Inatypical MIMO channel environment showninFigure 1 there existsatransmitterantennaarray(Tx),areceiver antennaarray(Rx) and the propagation paths of the trans-mitted signal. Obstacles thatarelocatedinthe environ-ment causereflection, diffraction and refraction. There-fore, the transmitted signal propagates to the receiver through a certain number of paths (1,2...L). When the transmitted signal travels from Tx to Rx on a path, it is delayed, its amplitude is attenuated and its phase is changed. With thisknowledgeathand the receivedsignal vectorcoming from any oneof thepathscanbe written as;
s(t;
e1)=
exp{j2wvlt}CRj
(QRX,l,)Al
CT.
(QTX
r1)u(t-
TO)
whereu(t) denotes the transmitted signalvector,e) given by
(1)
f--N
Path#1
Pat utX
Figure1. MIMO channel environment.
is a vectorthat consists of channel parameters such as
angles (azimuth andelevation)of arrival
(QR,
,il),
angles (azimuth andelevation)ofdeparture(QT
,i), delay (Ti), Dopplerfrequency (vi) and thepolarizationmatrix(Al).
Qisaunitvectorthat describesthe direction froma ref-erencepoint and givenas
Q =
[cosQ()
sin(0),sin(X)
sin(0),cos(0)IT,
(3)with 0 and X denoting the elevation and azimuth
an-gles, respectively. Finallyinequation (1),
CTX,P(Q)
andCRX,P(Q), the steering vectors for the transmit and
re-ceivearrays,aregiven by
CTX
,P(Q)=
LfT ,m,p(Q)exp{j
2wAo
m = 1,....M]T(p:
1(Q.rT.m)}
=v,h).
CRX,P(Q)=
fRT
,n,p(Q)exp{j2Ao
1(Q.rR
n = 1, ...,N]T(p
=v,h).
wherepdenotes the
polarization
type(either
vertical(v)
orhorizontal(h)),f
is thefieldpatternofantennas,risan-tenna
location,
M is number transmitterantennas and N is number receiverantennas.Equation
(1)
modelssignal
ofonepath foran antennaandtotal
signal
atthe receiver isasfollows:Figure2. MIMOchannelenvironment.
Figure2is theflowgraph of the SAGE
algorithm.
The al-gorithm hastwosteps,namely, expectation
(E)
andmax-imization (M) steps. In anE-step, the
expected
value of onepathat areceiverantennaoutputiscomputed.
InanM-step, thelikelihood function is maximized for param-eters ofone
path
sequentially.
For eachiteration,
these steps are performed for allpaths
and SAGEalgorithm
stopsuntil allparametersconverge.(4)
The formulas of SAGE
algorithm depend
on twocon-cepts: unobservable admissible data and observable
in-complete
data.Y(t)
describedinequation
(6)
is thein-complete
data and theadmissible data foronepath
isL,,,,n)
(;(5)
Xl(t)
=s(t;
E1)
+W(t).
(7)3.1 E-step
Since
Xl
is an unobservablefunction,
estimate of this function is basedontheobservable datavectorY(t)
by
L
Y(t)
=E
s(t;
El)
+W(t).
1=1(6)
Xl(t)
=Y(t)
L Es(t;
E,,)
i'=1,1'7zlHere W is the additive white Gaussian noise which is
assumedtobeindependent identically distributed
Gaus-sian.
3 SAGEAlgorithm
SAGE algorithm is an ML estimation method that has
beenused for differentapplicationslikechannel
parame-terestimation.
where
E1,
is the estimated parameters for l'thpathfromprevious
iteration(s)[2].
Thus, estimation of admissible data ofonepath (Xi (t)) canbe foundby subtractingestimatedsignalsof all other
pathsfrom observed dataY(t). NO K
No
SAGE +
outpu.i
a4 Particle Swarm Algorithmfor SAGE The computation ofparameters is done from the
like-lihood function and this function is maximized for the parametersindividually. The SAGE coordinate-wise up-dates for theparameters are derivedin [6] and given as follows,
Ti
= argmaxT,
Z(q5T.
1,T
,IqR,,
ORX
X,I,
T1,
V1
;
xi)
(9)
Vi=
argmax,j
IZ(q5T.,l,O:T.,l,q$R,,l,ORX,ll,Vl
;xi)
(10)
O¢)i
I=
arg
maxoT,,I
Z()T.,l,
OT.,l,
¢)RX,l
OR,,,l
-l,
I,)
;
:x)
(11)
OT,I=
arg
maxoTx I
|Z(OTl,I
OTx,l,
ORx,I:
,l,-l,T, VI)
;
xi)
(12)ORxll=
arg
maxx,
I
|Z(OTx,l
Txl,
OR,lR
(13,
)l,
I)l
;
--h)
(13)
As we see from theformulas, thealgorithm should search for theparametervalues maximizing the likelihood func-tion. Since the search domain is continuous this is a te-dious operation requiring efficient searchprocedures. To this endwepropose to usePSOtoperform the optimiza-tion.
PSO is one of the evolutionary computation techniques developed for non-linear optimization problems with continuousvaluedparametersthough itcanalso be used with discrete variables [4]. The procedure is based on researches on swarmslike bees and birdflocking and in-spired by social behavior ofswarms [4]. This procedure is simple and its computation time is short.
According to thebiological research, swarms find their food collectively notindividually; information is shared within the members ofswarm. Intheswarm,each mem-ber's position and its information are known, so each member'sposition andvelocityaremodified.
Swarmmovementoptimizesacertainobjective function. Each member knows its bestvalue
(pbest)
and its posi-tion. Moreover, each member knows the best value in theswarm(gbest)
amongpbests.
Each memberwants to change its position accordingto itsvelocity
and the dis-tance topbest
andgbest. Velocity
andposition of each membercanbecalculated by thefollowing equations:v= w*v+cl* rand* (pbest
-present)
+c2* rand* (gbest -present).present =present+v *t.
Al=
(|CTX,P
(Q)|CRX,P(Q) PTSC)Z(Tl,I,STl:XRX),l SRl,-Tl,Vl;--h) (15)
whereargmaxstands for theargumentof themaximum,
Pdenotes signal power,
T,c
denotes the sensing periodofareceiverantennaand zis likelihood function which
isgivenby
Z(ei; Xi) =
CRX
(QRXm,i)HXl (t;Ti,vli)CTX (QTX,l)(16) InEquation (16) (.)* denotes theconjugate, (.)H denotes the Hermitianoperators and(.)denotes normalization.
For the initialization procedure the successive
inter-ference cancelation is used and for the details of the derivation readermayreferto[6].
Here v is the member velocity, w is a scaling factor,
presentis thecurrentmember'sposition,rand isa
ran-dom number between (0,1). cl,c2 are learning factors,
usuallytakenas cl = c2= 2.
6 Performance of SAGEwith PSO for aSynthetic
Channel Data
To investigate the performance of SAGE algorithm
to-gether with PSO, asynthetic datais created. In the
cre-atedscenario, double polarizedsixantennasareused both
Txand Rx. L = 5waves arepropagating betweenRxand
Tx, and theparameters arechosenasinTable 1. The
an-tennainput signal consists of 55 rectangular pulses with
10 nanoseconds duration, the received signal (y(t)) is
sampledat1 GHz andsignaltonoise ratio(SNR)is taken
as40 dB.
The channel parameters are estimated by SAGE
algo-rithmusing three different search procedures: PSO,
ran-dom search anddogleg. Random Searchmeanssampling
the cost function within the reference domain and then
;xi)
(14)(17)
(18) 3.2
M-step
2.5 a) 1.5 a)C' X 1 C,) 0u Cn 0.5-_ PSO with Nb=5
-Trust Region Dogleg Algorithm
+ Random Search
2 3 4 5 6 7 8
Initialization(0) and SAGE iterations (1-10)
O0
0 9 10Figure3. MSE ofanangleofdeparture
Table]1. Createdsyntheticdataparameters.
choosing maximum of that function and dogleg isawell
knownalgorithm [3].
We observed that the SAGEalgorithm estimated the pa-rameters correctly and rapidly. Figure 3 shows MSE of
anangle of departureversusinitialization of SAGE
algo-rithm(0) and SAGE iteration numbers (1-10). From
Fig-ure 3, SAGE algorithm converges in 3 iterations for all
three algorithms. However, computation time for three approachesaresignificantly different. Table 2 shows the
computation time of search procedures both at initializa-tion and SAGE iterainitializa-tions. Table2illustrates that PSO's computation time for each iteration is significantly better than other methods.
7 Conclusion
In this paper we presented an application of particle swarmoptimization in SAGEalgorithm for channel pa-rameterestimation. PSOprovidesanefficientwayto
per-form optimization of the likelihood function for the
max-imizationstepof the SAGEalgorithm.
Table 2. Computational expense ofthe algorithm
com-pared with other methods. The algorithm works in Mat-labenvironmenton anAMDAthlon 3800processor
rate in terms of number of iterations, the computation time for each iteration is significantly improved by PSO comparedtoother searchprocedures. Therefore PSOcan
also beappliedtochannelparameterestimationproblem
toprovideanefficientoptimization.
References
[1] J. A. Fessler and A. 0. Hero, "Space-alternating
generalized expectation maximization algorithm," IEEE Trans. on Signal Processing, vol. 42, no. 10, pp.
26642677, Oct. 1994.
[2] B. H. Fleury, D. Dahlhaus, R. Heddergott, and M.
Tschudin,
"Wideband angle of arrival estimation using the SAGEalgorithm," in Proc. IEEE Fourth Int. Symp. Spread Spectrum Techniques and Applications (ISSSTA '96),pp. 79-85, Mainz, Germany, Sept. 1996.[3] M.J.D. Powell, "A Fortran subroutine for solving sys-temsof nonlinearalgebraic equations," Numerical Meth-ods for Nonlinear Algebraic Equations, (P. Rabinowitz, ed.), Ch.7, 1970.
[4] J. Robinson and Y. Rahmat-Samii, "Particle swarm
optimizationinelectromagnetics,"IEEETrans. Antennas Propag., vol. 52,no. 2,pp. 397-407,Feb. 2004.
[5] I. E, Telatar. "Capacity of multi-antenna gaussian
channels," European Transactions on
Telecommunica-tion, vol. 10,pp. 585595, Nov/Dec. 1999.
[6] X. Yin, B. H. Fleury, P. Jourdan, and A. Stucki, "Po-larization estimation of individual propagation paths
us-ing the SAGE algorithm," in Proceedus-ings of the IEEE In-ternational Symposium onPersonal, Indoor and Mobile
Radio Communications (PIMRC), Beijing, China, Sept. 2003.
To demonstrate thegains in computational complexity by utilizing PSOinSAGEalgorithm,weestimated the
chan-nel parameters from a synthetically generated data. We
observed that although there is no gain in convergence
Computation time of one
Estimation of iteration
onepath with (seconds)
Initialization 120 Random Search Sage iteration 1.67 Initialization 1.9 Dogleg Sage iteration 1.5 Initialization 1.75 PSO Sage iteration 0.025 channel parameters 1 =2 f= 3 £=4 f =5 Ti (nsecs) 25 35 45 55 65 AoA azimuth(°) 100 120 140 160 200 AoA elevation(°) 20 40 60 80 100 AoD azimuth(°) 300 330 20 50 70 AoD elevation