Channel Estimation of OFDM System Using

Adaptive Boosting Algorithm

Prasanta Kumar Pradhan∗, Oliver Faust†, Sarat Kumar Patra∗ and Beng Koon Chua†
∗ Dept. of Electrons & Comm. Engg.
National Institute of Technology, Rourkela, India 769008
Email: pradhanp, [email protected]
† DSP Center, School of Engineering

Ngee Ann Polytechnic, Singapore 599489

Email: fol2, [email protected]

Abstract—In this work we have applied Adapive Boosting mobile channels, pilot-based signal correction schemes are
(AdaBoost) for channel estimation of Orthogonal Frequency feasible method for OFDM systems [2]. Most channel esti-
Division Multiplexing (OFDM) systems. The result of the Ad- mation methods for OFDM transmission systems have been
aBoost algorithm was compared with other algorithms such
Least Square (LS), Best Linear Unbiased Estimator (BLUE) developed under the assumption of a slow fading channel,
and Minimum Mean Square Error (MMSE). AdaBoost not only where the channel transfer function is assumed stationary
increases the performance of the OFDM systems it also renders within one OFDM data block. In addition, the channel transfer
the QAM mapping obsolete and thereby reducing the complexity function for the previous OFDM data block is used as the
of receiver designs. transfer function for the present data block. In practice, the
Index Terms—OFDM, Channel Estimation, LS, MMSE, Ad-
aBoost. channel transfer function of a wideband radio channel may
have significant changes even within one OFDM data block.
Therefore, it is preferable to estimate channel characteristic
I. I NTRODUCTION
based on the pilot signals in each individual OFDM data block
In communication system many techniques, like Frequency [3].
Division Multiplexing Access (FDMA), Time Division mul- Recently, an elegant channel estimation method for OFDM
tiplexing Access (TDMA) and Code Division Multiplexing mobile communication systems has been proposed by [4] in
Access (CDMA), are used for transmission of signal. Where which a semi-blind low complexity frequency domain based
FDMA has very bad spectrum usage and TDMA performance channel estimation algorithm for multi-access OFDM systems
degrades by multipath delay spread causing Inter Symbol was proposed [4]. Many researchers have pursued channel
Interference (ISI). In contrast OFDM enables high bit-rate estimation in the time domain. A joint carrier frequency
wireless applications in a multipath radio environment the need synchronization and channel estimation scheme, using the
for complex receivers. OFDM is a multi-channel modulation expectation-maximization (EM) approach, is presented in [5]
system employing Frequency Division Multiplexing (FDM) while [6] used subspace tracking. In [7], a joint and data
of orthogonal sub-carriers, each modulating a low bit-rate estimation algorithm is presented which makes a collective use
digital stream. OFDM uses N overlapping (but orthogonal) of data and channel constraints. A joint frequency-offset and
sub bands, each carrying a baud rate of 1/T and they are channel estimation technique for multi-symbol encapsulated
spaced 1/T apart. Because of the selected frequency spacing, (MSE) OFDM system is proposed in [8], while the authors of
the sub-carriers are all mathematically orthogonal to each [9] presented a sequential method based on carrier frequency
other. This permits proper demodulation of symbol streams offset and symbol timing estimation. The authors of [10]
without the requirement of non overlapping spectra. estimated the channel based on power spectral density (PSD)
Currently, there is increasing interest in OFDM as it com- and least squares (LS) estimation for OFDM systems with
bines the advantages of high data rates and easy implementa- timing offsets. A pilot aided channel estimation algorithm
tion. This is reflected by the many standards that considered in the presence of synchronous noise by exploiting the a
and adopted OFDM those for Digital Audio Broadcast (DAB) priori available information about the interference structure
and Digital video Broadcast (DVB), high speed modems over was presented in [11] while [12] used implicit pilots for joint
digital subscriber lines, and Wireless Local Area Network detection and channel estimation. A joint time domain tracking
(WLAN) broadband systems as of ieee 802.11a, 802.11b and of channel frequency offset for OFDM systems is suggested
802.11g. in [13] while a time domain carrier frequency offset (CFO)
The accuracy of channel state information greatly influences tracking method based on Particle filtering is presented in [14].
the overall system performance [1]. The main challenges asso- In this work we have implemented AdaBoost Algorithm,
ciated with OFDM systems today are- channel identification a pattern recgnition algorithm, for estimation of the pilot
and tracking, channel coding and equalization. In wideband aided channel estimation. We have compared the result of
 (a) Block Type (b) Comb Type
Fig. 2. Pilot Arrangements

circular convolution. The channel impulse response h(n) can

Fig. 1. Block Diagram of the OFDM systems be expressed as [15]
r−1
n
X
h(n) = hi ej2πfDi T N δ(λ − τi ) (4)
the AdaBoost with other well known estimation algorithms. i=0
AdaBoost outperformed the other algorithms and that it
Where r the total number of propagation paths, hi is the
renders QAM mapping redundant. This can lead to more
complex impulse response of the ith path, fDi is the ith
performant receiver systems which have a lower computational
path’s Doppler frequency shift which causes Inter channel
complexity. Both factors aid power efficient receivers which
Interference (ICI) of the received signals, λ is the delay spread
are used in wireless mobile devices.
index, and τi is the ith path delay time normalized by the
The rest of the article is arranged as follows: Section. 2
sampling time. After removing the guard interval from yg (n),
describes the mathematical model of OFDM systems and its
the received samples y(n) are sent to a Fast Fourier Transform
channel model along with two types of pilot arrangement. The
(FFT) block to demultiplex the multi-carrier signals.
implementation of AdaBoost and other algorithms (like LS, PN −1
MMSE, and BLUE) is discussed in Section. 3 . Section. 4 Y (k) = F F T {y(n)} = N1 n=0 y(n)e−j2πkn/N (5)
describes the performance analysis of different algorithms and f or k = 0, 1, . . . , N − 1
finally Section. 5 draws up the conclusions.
If we assume that the guard interval is longer than the length
II. S YSTEM D ESCRIPTION of channe1 impulse response- there is no inter-symbol inter-
ference between OFDM symbols- the demultiplexed samples
Y (k) can be represented by
The block diagram of an OFDM system is given in Figure.1.
The binary information data are grouped and mapped into Y (k) = X(k)H(k) + W (k), k = 0, 1, . . . , N − 1 (6)
multi-amplitude-multi-phase signals. In this paper, we consider sin(πfDi T ) 2πτi
j2πf T
the 16-QAM modulation. After pilot insertion, the modulated Where H(k) = hi e Di πfDi T
e−j N k and W (k) is
data X(k) are sent to an IDFT and are transformed and the Fourier transform of the AWGN w(k).
multiplexed into x(n) as After that, the received pilot signals {Yp (k)} are extracted
PN −1 from {Y (k)}, the channel transfer function {H(k)} can be
j2πkn/N
x(n) = IF F T {X(k)} = k=0 X(k)e (1) obtained from the information carried by {Hp (k)}. With the
f or n = 0, 1, ..., N − 1 knowledge of the channel responses {H(k)}, the transmitted
data samples {X(k)} can be recovered by simply dividing the
where N is the number of subcarriers. The guard interval received signal by the channel response:
Ng is inserted to prevent possible inter-symbol interference
in OFDM systems, and the resultant samples xg (n) are y(k)
x̂(k) = (7)
 Ĥ(k)
x(N + n) n = Ng , Ng − 1, . . . , −1
xg (n) = (2)
x(n) n = 0, 1, . . . , N − 1 where Ĥ(k) is an estimate of H(k) . After signal demapping,
where Ng is the number of samples in the guard interval. The the source binary information data are reconstructed at the
transmitted signal is then sent to a frequency selective multi- receiver output.
path fading channel. The received signal can be represented The OFDM transmission scheme makes it easy to assign
by pilots in both time and frequency domain. Figure.2 shows
yg (n) = xg (n) ⊗ h(n) + w(n) (3) two major types of pilot arrangement. The first kind of pilot
arrangement shown in Figure.2a is denoted as block-type pilot
Where h(n) is the channel impulse response (CIR) and w(n) arrangement. The pilot signal is assigned to a particular OFDM
is the Additive White Gaussian Noise (AWGN) and ⊗ is the block, which is sent periodically in time domain. This type
of pilot arrangement is especially suitable for slow-fading The LS estimate of Hp is susceptible to AWGN and Inter-
radio channels. The estimation of channel response is usually Carrier Interference (ICI). Because the channel responses of
obtained by either LS or MMSE estimates of training pilots data subcarriers are obtained by interpolating the channel
[5]. characterstics of pilot subcarriers, the performance of OFDM
The second kind of pilot arrangement, shown in Figure. 2b, systems which are based on comb-type pilot arrangement is
is denoted as comb-type pilot arrangement. The pilot signals highly dependent on the rigorousness of estimate of pilot sig-
are uniformly distributed within each OFDM block. Assuming nals. Thus, a estimate better than the LS estimate is required.
that the payloads of pilot signals of the two arrangements The MMSE estimate has been shown to be better than the LS
are the same, the comb-type pilot assignment has a higher estimate for channel estimation in OFDM systems based on
retransmission rate. Thus, the comb-type pilot arrangement block-type pilot arrangement. The Mean Square Error (MSE)
system is provides better resistance to fast-fading channels. estimation, given in [15], shows that MMSE estimate has about
Since only some sub-carriers contain the pilot signal, the 10-15 dB gain in SNR over the LS estimate for the same MSE
channel response of nonpilot subcarriers will be estimated by values. The major drawback of the MMSE estimate is its high
interpolating neighboring pilot sub-channels. Thus, the comb- complexity, which grows exponentially with the observation
type pilot arrangement is sensitive to frequency selectivity samples.
when comparing to the block-type pilot arrangement system.
B. MMSE Estimation
That is, the pilot spacing (∆f )p , must be much smaller than
the coherence bandwidth of the channel (∆f )c [3]. The MMSE estimator employs the second order statistics
of the channel conditions in order to minimizes Mean Square
III. C HANNEL E STIMATION Error (MSE). Let Rhh , RHH , and RY Y be the autocovariance
For comb-type pilot sub-carrier arrangement, the Np pilot matrix of h, H, and Y , respectively and RhY be the cross
signals Xp (m) , m = 0, 1, ..., Np − 1, are uniformly inserted 2
covariance matrix between h and Y . Also σw denotes the
into X(k). That is, the total N sub-carriers are divided into Np 2
AWGN variance E{|w |}. Assume the channel vector h, and
groups, each with L = N/Np adjacent sub-carriers. In each the AWGN w are uncorrelated, it can be derived that [16]
group, the first sub-carrier is used to transmit pilot signal.
RHH = E{HH H } = E{(F h)(F h)H } = F Rhh F H (13)
The OFDM signal modulated on the kth sub-carrier can be
expressed as where
 
0(N −1)
X(k) = X(mL + l) T WN00 ... T WN
Xp (m), l = 0,

F = .. .. .. 
(14)
=  . . .


inf ormation data, l = 1, 2, . . . , L − 1 (N −1)0 (N −1)(N −1)
T WN ... T WN
(8)
and T WNi,k = √1N e−j2πi N is called as the twidle factor
k
Let
T
HP = [Hp (0) Hp (1) . . . Hp (Np − 1)] matrix. Assuming Rhh (thus RHH ) and σw 2
are known at the
= [H(0) H(L − 1) . . . H((Np − 1)(L − 1))]T receiver in advance, the MMSE estimator of the h is given by
(9) ĥMMSE = RhY RY Y Y Y H . At last it can be estimated that
be the channel response of pilot carriers, and
2
ĤMMSE = F ĥMMSE = RHH [RHH +σw (XX Ht )−1 ]−1 ĤLS
Yp = [Yp (0) Yp (1) . . . Yp (Np − 1)] (10) (15)
be a vector of received pilot signals. The received pilot signal MMSE employs the second order statistics of the channel
vector Yp can be expressed as for estimation. Some times the channel statistics are not avail-
able, so it is difficult to estimate the channel in this situation.
Yp = Xp Hp + Wp (11) However, in OFDM systems the signal can be available at the
where   receiver by means of pilot carriers.
Xp (0) 0 ... 0
 .. .. C. BLUE
Xp =  .

Xp (1) . . . . 
If we restrict the estimator to be linear in data and and find
0 0 . . . Xp (Np − 1) the linear estimator that is unbiased and minimum variance
where Wp is the vector of Gaussian noise in pilot sub-carriers.
then the estimator is called Best Linear Unbiased Estimator
A. LS Estimation (BLUE). BLUE can be determined with knowledge of only
In conventional comb-type pilot based channel estimation the first and second moment of the PDF. Since complete
methods, the estimation of pilot signals, is based on the LS knowledge of the PDF is not necessary, the BLUE is more
method is given by suitable for practical implementation [17].
Ĥp,ls = [Hp,ls (0) Hp,ls (1) . . . Hp,ls (Np − 1)] Gauss-Markov Theorem 1. If the data can be modeled in
= Xp−1 Yp (12) the following linear form
Y (0) Y (1) Y (N −1)
= [ Xpp (0) Xpp (1) . . . Xpp (Npp −1) ] Y = XH + W (16)
 TABLE I
where X is a known N × p matrix and H is a p × 1 vector of OFDM PARAMETERS
parameters to be estimated, and W is a N × 1 noise vector
with zero mean and covariance C (the PDF of W is otherwise No of Subcarriers 256
No. of Pilot Carriers 32
arbitrary ), then BLUE of H is given by Guard Interval 64
Guard Type Cyclic Extension
HBLUE = (H Ht C −1 H)−1 H Ht C −1 X (17)

where C = (X − E{x})(X − E{x})Ht and Ht is the con-

jugate transpose or Hermitian Transpose. and the minimum
variance of Ĥi is

var(Ĥi ) = [(H Ht C −1 H)−1 ]ii (18)

D. AdaBoost
Boosting is a general method for improving the accuracy of
any given learning algorithm. AdaBoost was originally defined
for two class problems but it can be extended boosting to
multi-class and regression problems. The AdaBoost algorithm,
introduced in 1995 by Freund and Schapire, solved many of
the practical problems of the earlier boosting algorithms [18].
The AdaBoost algorithm for multiclass problem is described
as below.
Suppose we are given a set of training data
(x1 , c1 ), (x2 , c2 ), . . . , (xn , cn ), where the input (prediction
variable) xi ∈ Rp and the output (response variable) Fig. 3. Performance Comparison of Different Algorithms
ci is quantitative values in afinite set, e.g. 1, 2, . . . , K.
Where K is the number of classes. Usually it is assumed
that the training data are composed from independently 1) Initialize the observation weights wi = 1/n, i =
and identically distributed (iid) samples from a unknown 1, 2, . . . , n.
probability distribution P rob(X, C). The gaol is to find out a 2) For m = 1 to M:
classification rule C(x) from the training data, so that when a) Fit a classifier T (m) (x) to the training data using
given a new input x, we can assign a class label c from weights wi .
1, 2, . . . , K. The misclassification error rate of a classifier b) Compute
C(x) is given by 1 − K
P
k=1 E X [IC(X)=k P rob(C = k|X)].
n n
So from [18] (m)
X
(m)
X
err = wi I(ci 6= T (xi ))/ wi
C ⋆ (x) = arg maxk P rob(C = k|X = x) (19) i=1 i=1
(20)
will minimize this quantity with the misclassification error rate c) Compute
equal to 1 − EX maxk P rob(C = k|X). This classifier is 1 − err(m)
known as Bayes Classifier and its rate is Bayes Error Rate. α(m) = log + log(K − 1) (21)
err(m)
The multiclass AdaBoost algorithm tries to approximate the
Bayes Classifier C ⋆ (x) by combining many weak classifiers. d) Set
Starting with an unweighted training sample, the AdaBoost wi ← wi exp(α(m) . I(ci 6= T (m) (xi ))) (22)
builds a classifier that produces class labels. If a training data
point is misclassified, the weight of that training data point for i=1, 2, . . . , n.
is increased (boosted). A second classifier is built using these e) Re-normalize wi .
new weights, which are no longer equal. Again, misclassified 3) Output
training data have their weights boosted and the procedure M
is repeated. Typically, one may build 500 or 1000 classifiers C(x) = arg maxk
X
α(m) I (T (m) (x) = k)
this way. A score is assigned to each classifier, and the final m=1
classifier is defined as the score weighted linear combination (23)
of the classifiers from each stage. Specifically, let T (x) denote
a weak multi-class classifier that assigns a class label to x, then IV. S IMULATION R ESULTS
the AdaBoost algorithm proceeds as follows: [19]
Multiclass AdaBoost Algorithm: 1. The Algorithm is as The OFDM system channel estimation was simulated with LS,
follows: MMSE, BLUE and AdaBoost methods. In all simulations we
have used QAM16 as the modulation scheme for the individual [13] T.Roman, M.Enescu, and V.Koivunen, “Joint time-domain tracking of
carriers. Other parameters of the simulation are given in the channel and frequency offsets for mimo ofdm systems,” in Wireless
Personal Communications, vol. 31, no. 3-4, 2004, pp. 181–200.
Table.I. Figure.3 shows comparison of Bit Error Rate (BER) [14] T.Nyblom, T.Roman, M.Enescu, and V.Koivunen, “Time-varying carrier
for LS, MMSE, BLUE and AdaBoost. Figure.3 shows that offset tracking in ofdm systems using particle filtering,” in Proceedings
the AdaBoost algorithm improves the performance specially of the Fourth IEEE International Symposium on Signal Processing and
Information Technology, 2004, pp. 217–220.
in low SNR values. However, at high SNR both MMSE [15] J. V. de Beek, O. Edfors, M. Sandel, S. Wilson, and P. Borjesson, “On
and AdaBoost show a similar performance. Furtheremore, channel estimation in ofdm systems,” in IEEE 45th Vehicular Technology
AdaBoost gives better performance when compared to LS and Conference, 1995.
[16] Y. Shen and E. Martinez, “Channel estimation in ofdm systems,” in
BLUE. Nevertheless, BLUE and LS are performing same or Frescale Semiconductor Application Note, 2006.
we can tell BLUE gives very marginal improvement to LS. [17] S. M. Kay, Fundamentals of Statistical Signal Processing- Estimation
The reason for this performance increase is because of the Theory. Prentice Hall PTR, 1993, vol. 1.
[18] Y. Freund and R. E. Scapire, “A short introduction to boosting,” Journal
covariance matrix used in the BLUE. As our noise is ASWGN of Japaneese Society for Artificial Intelligence, vol. 14, no. 5, pp. 771–
and it has variance of 1 so the BLUE’s performance is all most 780, 1999.
that of LS algorithm. [19] J. Zhu, H. Z. S. Rosset, and T. Hastie, “Multiclass adaboost,” Journal
of Statistics and it’s Interface, vol. 2, pp. 349–360, 2009.
V. C ONCLUSSION
AdaBoost outperforms the other three algorithms which
were discussed. The computational complexity of the MMSE
increases exponentially as the number of carrier increases.
Whereas the computatinal complexity is not exponential in the
case of AdaBoost. Hence, AdaBoost can be employed when
a high number of carriers is required. Furthermore, as it is a
classification algorithm so in the receiver side we will require
a separate demapper (or decoder) to get the desired data bits.
R EFERENCES
[1] Y. Li and G. Stber, Orthogonal Frequency Division Mutiplexing for
Wireless Communication, Y. Li and G. Stber, Eds. Springer, 2006.
[2] F. Tufvesson and T. Masseng, “Pilot aided channel estimation for ofdm
in mobile cellular system,” in in Proceeding IEEE 47th Vehicular
Technology Conference. Phoneix, USA: IEEE, May 1997, pp. 11 639–
1643.
[3] M.-H. Hsieh and C.-H. Wei, “Channel estimation for ofdm systems
based on comb type pilot arrangement in frequency selective fading
channels,” IEEE Transactions on Consumer Electronics, vol. 44, no. 1,
pp. 217–225, February 1998.
[4] M. Sohail and T.Y.Al-Naffouri, “An em based frequency domain channel
estimation algorithm for multi access ofdm systems,” Elsevier Journal
of Signal Processing, vol. 10, pp. 1562–1572, 2010.
[5] J. Lee, J. Han, and S. Kim, “Joint carrier frequency synchronization
and channel estimation for ofdm systems via em algorithm,” IEEE
Transactions on Vehicular Technology, vol. 55, no. 1, pp. 167–172, 2006.
[6] X. Hou, Z. Zhang, and H. Kayama, “Time domain mmse channel
estimation based on sbspace tracking for ofdm systems,” in IEEE 63rd
Vehicular Technology Conference, Melbourne, Victoria, May 2006, pp.
1565–1569.
[7] T. Al-Naffouri, “An em based forward backward kalman filter for
estimation time variant channels in ofdm,” IEEE Transaction on Signal
Processing, vol. 1, no. 11, pp. 1–7, 2007.
[8] X. Wang, Y. Wu, J. Chouinard, and H. Wu, “On the design and
performance analysis of multisymbol encapsulated ofdm systems,” IEEE
Transactions on Vehicular Technology, vol. 55, no. 3, pp. 990–1002,
2006.
[9] J. Liu and J. Li, “Parameter estimation and error reduction for ofdm
based wlans,” IEEE Transactions on Mobile Computing, vol. 3, no. 2,
pp. 152–163, 2004.
[10] S. Qin, P. Liu, L. Zheng, and D. Wang, “Channel estimation with timing
offset based on psd & ls estimation for wireless ofdm systems,” in
International Symposium on Intelligent Signal Processing and Commu-
nication Systems, November-December 2007, pp. 248–251.
[11] A. Jeremic, T. Thomas, and A. Nehorai, “Ofdm channel estimation in
the presence of interference,” IEEE Transactions on Signal Processing,
vol. 52, no. 12, pp. 3429–3439, 2004.
[12] R.Dinis, N.Souto, J.Silva, A.Kumar, and A.Correia, “Joint detection and
channel estimation for ofdm signals with implicit pilots,” in IEEE Mobile
and Wireless Communications Summit, July 2007, pp. 1–5.