Beruflich Dokumente
Kultur Dokumente
y
ks+vT+2
.
.
.
y
(k+1)s
0
(1)
h . . . 0
.
.
.
.
.
.
0 . . . h
0
(2)
PQ
N
0 0
0 PQ
N
0
0 0 PQ
N
X
. .. .
X
(k1)
1:N
X
(k)
1:N
X
(k+1)
1:N
+
n
. .. .
n
ks+vT+2
.
.
.
.
.
.
n
(k+1)s
= HX + n (1)
Perfect synchronization at the receiver is assumed in this
paper. Van Acker et al. [6] have proposed an alternate equal-
ization structure, called per tone equalization, which accom-
plishes the same task as the TEQ/FEQ, but with improved
performance and comparable complexity. The full details of
the per tone structure can be found in [6]. briey, demodula-
tion is accomplished by an FFT of size N, which is done by
pre-multiplying y by F
N
. Per tone equalization of bin i is ac-
complished by forming a linear combination of the i th FFT
output and T 1 difference terms of the pre-FFT signal, y
z
i
= v
T
i
I
T 1
0 I
T 1
0 F
N
(i, :)
F
i
y
. (2)
where v
T
i
= [v
i,T1
. . . v
i,0
], F
N
an N N FFT-matrix and
F
N
(i, :) the i th row of F
N
. The linear combiner (not a tapped
delay line) v
i
is the equalizer for tone i ; v
i
is the time-reversal
of v
i
, dened for convenience; and z
i
is the equalized data
for tone i. The notation on (1) and (2) was introduced in [6]
with full details, but is repeated here for our reference. In
this paper, we propose a new method (SVM) for determina-
tion of the per tone equalizer coefcients and compare its
performance with constant modulus algorithm (CMA) and
multi-modulus algorithm (MMA) from view point of aver-
age bit error rate (BER).
3. Problem formulation
3.1. Multilevel SVM-based blind per tone
equalization the blind algorithm
The proposed algorithm [2,4,5] minimizes the following
SVM-based cost function for tone i
L
p
( v
i
) =
1
2
v
i
2
+ c
M
k=1
L
c
(u
i
(k)) (3)
where
L
c
(u
i
) =
0, u
i
<c
u
2
i
2u
i
c + c
2
, u
i
c
(4)
is a c-insensitive quadratic loss function modied to guar-
antee a continuous derivative. Continuity of the derivative
is necessary for the numerical stability of the algorithm. We
select a suitable penalization term called u
i
(k) in order to
apply aforementioned cost function for blind equalization.
Here, we propose to use u
i
(k) = |e
i
(k)| with the error term
e
i
(k) being
z
i
(k) = v
T
i
F
i
y(k) (5)
and
e
i
(k) = |z
i
(k)|
2
R
2,i
= z
i
(k)z
i
(k) R
2,i
(6)
R
2,i
is the Godard constant, for tone i and superindex de-
notes the complex conjugate. The Godard algorithms adapt
the equalizer to minimize the following cost function:
J
G
( v
i
) = E[(|z
i
(k)|
p
R
p,i
)
2
] (7)
the ratio R
p,i
contains the a priori knowledge about the
current modulation
R
p,i
=
E[|X
i
|
2p
]
E[|X
i
|
p
]
(8)
CMA is the Goadard algorithm for p = 2. The proposed
method introduces a penalty term inspired by the CMA cost
function. For optimality reasons, iterative re-weighted least
squares algorithm (IRWLS) is used. This procedure has been
successfully applied to solve SVMs [7] and it has recently
proven to converge to the SVM solution [8]. A rst order
Taylor series expansion of L
c
(u
i
) is used to obtain the cost
function that produces the IRWLS algorithm
L
p
( v
i
) =
1
2
v
i
2
+ C
k=1
L
c
(u
i
( j, k)) +
dL
c
(u
i
)
du
i
|
u
i
( j,k)
[u
i
(k)
u
i
( j, k)]
(9)
188 B.H.B. Naeeni et al. / Int. J. Electron. Commun. (AE) 64 (2010) 186190
where u
i
( j, k)=|e
i
( j, k)| and e
i
( j, k)=| v
T
i
( j )F
i
y(k)|
2
R
2,i
is the error term after the j -th iteration. Then, a quadratic
approximation is constructed as follows:
L
p
( v
i
) =
1
2
v
i
2
+ C
k=1
L
c
(u
i
( j, k))
+
dL
c
(u
i
)
du
i
u
i
( j,k)
(u
i
(k))
2
(u
i
( j, k))
2
2u
i
( j, k)
=
1
2
v
i
2
+
1
2
M
k=1
a
i
(k)|e
i
(k)|
2
+ CTE (10)
CTE represents constant terms that do not depend on v
i
, and
the weights a
i
(k) are
a
i
(k) =
C
u
i
( j, k)
dL
c
(u
i
)
du
i
u
i
( j,k)
(11)
=
0, u
i
( j, k) <c
2C(u
i
( j,k)c)
u
i
( j,k)
, u
i
( j, k) c
(12)
L
p
( v
i
) is a quadratic functional for L
p
( v
i
) in (3) that
presents the same value L
p
( v
i
( j )) = L
p
( v
i
( j )) and gradient
for
v
i
L
p
( v
i
( j )) =
v
i
L
p
( v
i
( j )) for v
i
= v
i
( j ). Therefore,
we can dene p
i
( j ) = v
i
(s) v
i
( j ) as a descending direc-
tion for L
p
( v
i
), where v
i
(s) is the least square solution to
(10), and we use it to construct a line search method [9],
i.e v
i
( j + 1) = v
i
( j ) + n
i
( j )p
i
( j ). The value to n
i
( j ) is
computed using a backtracking line search [9], in which
n
i
( j ) is initially set to 1 and if L
p
( v
i
( j + 1))L
p
( v
i
( j )),
it is iteratively reduced until a strict decrease in the func-
tional in (3) is observed. To obtain the solution to L
p
( v
i
),
its gradient is set to zero
v
i
L
p
( v
i
) = v
i
+ 2
M
k=1
a
i
(k)(|z
i
|
2
R
2,i
).
i
= 0 (13)
where z
i
= v
T
i
F
i
y(k), and .
i
= v
T
i
F
i
y(k)F
i
y
v
i
L
p
( v
i
) = v
i
+ 2
M
k=1
a
i
(k)([
i
R
2,i
)
i
= 0 (14)
where [
i
= v
T
i
F
i
y(k)z
i
(k), and
i
= z
i
(k)F
i
y
R
2,i
, . . . , e
i
(1, M) = | v
T
i
(1)F
i
y(M)|
2
R
2,i
. amounts of
y(k) for k = 1, 2, . . . , M are constant for other iterations
until complete convergence and change with changing the
tap coefcients of the channel and this method repeats for
one thousand realization of channel. Finally, the IRWLS
procedure is summarized in the following steps:
1. Initialization: initialize v
i
(1), obtain z
i
(k) by (5), e
i
(k)
by(6),calculate u
i
(k) = |e
i
(k)| and compute a
i
(k) from
(11). Set j = 1.
2. Compute v
i
(s) by solving (15) and set n
i
( j ) = 1.
3. Set v
i
( j +1) = v
i
( j ) +n
i
( j )[ v
i
(s) v
i
( j )]. If L
p
( v
i
( j +
1)) < L
p
( v
i
( j )) go to step 5.
4. Set n
i
( j ) = jn
i
( j ) with 0 <j <1 and go to step3.
5. Recompute e
i
(k), u
i
(k) and a
i
(k), set j = j + 1 and go
to step 2 until convergence.
3.2. CMA-based blind adaptive per tone
equalization
The CMA is a popular alternative in decision-directed al-
gorithms. A detailed review of its convergence behavior in
single-carrier systems can be found in [10]. CMA attempts
to minimize the dispersion of the equalized symbols by per-
forming a stochastic gradient descent of
J
CM,i
= E[(|z
i
(k)|
2
i
)
2
] (16)
for each bin i, where z
i
(k) is the per tone equalizer output
and
i
= E[|X
i
|
4
]/E[|X
i
|
2
]. The resulting algorithm, i.e.,
CMA-based blind adaptive per tone equalization, for i =
1, . . . , N and k = 1, 2, 3, . . . , is
z
i
(k) = v
T
i
(k)F
i
y(k)
v
i
(k + 1) = v
i
(k) jz
i
(k)(|z
i
(k)|
2
i
)F
i
y
(k)
(17)
3.3. MMA-based blind adaptive per tone
equalization
Yang et al. [11] proposed a modied CMA called the
MMA, whose cost function is
J
MM,i
=J
R,i
+J
I,i
= E[(z
2
R,i
(k)
R,i
)
2
]
+ E[(z
2
I,i
(k)
I,i
)
2
] (18)
B.H.B. Naeeni et al. / Int. J. Electron. Commun. (AE) 64 (2010) 186190 189
for each bin i, where z
R,i
(k) and z
I,i
(k) are the real and
imaginary parts of the per tone equalizer output, respectively,
R,i
and
I,i
are given by
R,i
=
E[X
4
R,i
]
E[X
2
R,i
]
and
I,i
=
E[X
4
I,i
]
E[X
2
I,i
]
(19)
in which X
R,i
and X
I,i
denote the real and imaginary parts of
X
i
, respectively. The resulting algorithm; i.e., MMA-based
blind adaptive per tone equalization, for i = 1, . . . , N and
k = 1, 2, 3, . . . , is
z
i
(k) = v
T
i
(k)F
i
y(k)
v
i
(k + 1) = v
i
(k) je
i
(k)F
i
y
(k)
(20)
where e
i
(k) = e
R,i
(k) + j e
I,i
(k), in which e
R,i
(k) =
z
R,i
(k)(z
2
R,i
(k)
R,i
) and e
I,i
(k) = z
I,i
(k)(z
2
I,i
(k)
I,i
).
4. Channel model and simulation results
Amultipath fading channel in [12] with L=(J 1)L
1
+L
2
taps is used based on the following impulse response model:
h(l) =
J1
j =0
e
[
1
j
j L
1
+L
2
1
m=j L
1
:
m
e
[
2
(mj L
1
)
o(l m) (21)
where a
m
is a zero mean complex, circularly symmetric,
Gaussian random process such that E[a
m
a
j
] =o(m j ), [
1
and [
2
are exponential decay factors. In simulations, J =3,
L1=L2 =7 for L =21 and J =4, L
1
=L
2
=6 for L =24,
[
1
= 0, and [
2
= 0.25 are chosen. The CIR intervals are
L = 21 and L = 24 which are longer than the CP interval
and the additive Gaussian noise, n(k), is a white process.
One thousand independent realizations of h(l) based on (21)
have been used in simulations for a 16-QAM OFDM system
with N =64 subcarriers and a CP interval v=16. Number of
taps for per tone equalization is T =v +1. A burst of 2500
OFDM symbols is assumed to be transmitted (M = 2500).
Hence the CIR is assumed constant for each burst. Each data
point in the simulation results (Figs. 1 and 2) is obtained by
averaging over 1000 such bursts. In these gures the value
of j is set to 1e-4 for CMA and MMA and C =10, c =0.01
and j=0.05 for blind SVM. Without loss of generality, 16-
QAM mapping for all subchannels has been employed and
all subchannels are used. The simulation results for per tone
equalization using CMA and MMA and blind SVM have
been shown in Figs. 1 and 2, they show the average BER
as a function of SNR for 16-QAM and L =21 and L =24,
respectively. These gures show that the performance of
blind adaptive per tone equalization using SVMis better than
blind adaptive per tone equalization using CMA and MMA
fromthe viewpoint of average BERfor 16-QAM. Though,in
terms of accuracy blind SVMis superior, however, it requires
more computations than CMA and MMA.
0 5 10 15 20 25 30
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e
B
E
R
One tap FEQ
CMA
MMA
blindSVM
Fig. 1. The average bit-error rate as a function of SNR, CIR
interval is L = 21,16-QAM.
0 5 10 15 20 25 30
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e
B
E
R
One tap FEQ
CMA
MMA
blindSVM
Fig. 2. The average bit-error rate as a function of SNR, CIR
interval is L = 24,16-QAM.
5. Conclusion
In this paper, we propose a new blind OFDM channel
equalization method based on SVM for multilevel signals.
In order to determine the equalizer coefcients, we use the
difference of the per tone equalizer output and the R
2,i
pa-
rameter of the CMA algorithm. Our simulations show that
the average BER for per tone equalization using blind-SVM
is better than per tone equalization using CMA and MMA
for 16-QAM.
Acknowledgment
I would like to thank the Islamic Republic of Iran Broad-
casting University for their nancial support.
190 B.H.B. Naeeni et al. / Int. J. Electron. Commun. (AE) 64 (2010) 186190
References
[1] Scholkopft B, Smola A. Learning with kernels. Cambridge,
MA: MIT Press; 2002.
[2] Santamaria I, Ibanez J, Vielva L, Pantaleon C. Blind
equalization of constant modulus signals via support vector
regression. In: ICASSP, vol. II; April 2003. p. 73740.
[3] Sebald DJ, Bucklew JA. Support vector machine techniques
for nonlinear equalization. IEEE Transactions on Signal
Processing 2000;48(11):321726.
[4] Santamaria I, Pantaleon C, Vielva L, Ibanez J. Blind
equalization of constant modulus signals using support
vector machines. IEEE Transactions on Signal Processing
2004;52(6):177382.
[5] Lazaro M, Santamaria I, Via J, Erdogmus D. Blind
equalization of multilevel signals using support vector
machines. In: Proceedings of the European signal processing
conference. September 2004.
[6] Van Acker K, Leus G, Moonen M, van de Weil O, Pollet T. Per
tone equalization for DMT-based systems. IEEE Transactions
on Communications 2001;49(1):10919.
[7] Perez-Cruz F, Navia-Vazquez A, Alarcon-Diana P, Artes-
Rodriguez A. An IRWLS procedure for SVR. In: Proceedings
of the EUSIPCO. September 2000.
[8] Perez-Cruz F, Bousono-Calzon C, Lin C-H, Artes-Rodriguez
A. Convergence of the IRWLS procedure to the support vector
machine solution. IEEE Transactions on Neural Networks
2003;14(2):296303.
[9] Nocedal J, Wright SJ. Numerical optimization. Berlin:
Springer; 1999.
[10] Johnson CR, Schniter Jr P, Endres TJ, Brown DR, Casas RA.
Blind equalization using the constant modulus criterion: a
review. Proceedings of the IEEE 1998;86:192750.
[11] Yang J, Werner J-J, Dumont GA. The multimodulus blind
equalization and its generalized algorithms. IEEE Journal of
Selected Areas in Communications 2002;20(6):9971015.
[12] Zamiri-Jafarian H, Khoshbin H, Pasupathy S. Time-domain
equalizer for OFDM systems based on SINR maximization.
IEEE Transactions on Communications 2005;53(6).
Babak Haji Bagher Naeeni received
the B.S and M.S in electrical engi-
neering from central Tehran branch,
Islamic Azad University, Tehran, Iran,
and from south Tehran branch, Is-
lamic Azad University, Tehran, Iran,
in 1993 and 1999, respectively. He is
currently pursuing the Ph.D. degree
in electrical engineering in science
and research branch, Islamic Azad
University, Tehran, Iran. His research interests include multiuser-
detection, channel equalization, articial intelligence.
Hamidreza Amindavar is a faculty
member in electrical engineering de-
partment of Amirkabir University of
Technology since 1993. He obtained
his B.S.E.E. in 1985, M.S.E.E. in 1987,
M.Sc. in applied mathematics in 1991,
Ph.D. in electrical engineering in 1991.
His research interests include statistical
image processing, signal processing,
multiresolution signal analysis, and
multiuser detection.
Hamidreza Bakhshi was born in
Tehran, Iran on May 11, 1971. He re-
ceived the B.Sc. degree in Electrical
Engineering from Tehran University,
Iran in 1992, the M.Sc. and Ph.D. de-
gree in Electrical Engineering from
Tarbiat Modarres University, Iran, in
1995 and 2001, respectively. Since
2001, he has been as an Assistant
Professor of Electrical Engineering at
Shahed University, Tehran, Iran. His research interests include
wireless communications, multiuser detection, and smart antennas.