Wireless Pers Commun (2009) 51:221–229

DOI 10.1007/s11277-008-9639-2

Artificial Neural Network Channel Estimation Based

on Levenberg-Marquardt for OFDM Systems

Cebrail Çiflikli · A. Tuncay Özşahin · A. Çağrı Yapici

Published online: 14 November 2008

© Springer Science+Business Media, LLC. 2008

Abstract The many advantages responsible for the widespread application of orthogonal
frequency division multiplexing (OFDM) systems are limited by the multipath fading. In
OFDM systems, channel estimation is carried out by transmitting pilot symbols generally.
In this paper, we propose an artificial neural network (ANN) channel estimation technique
based on levenberg-marquardt training algorithm as an alternative to pilot based channel esti-
mation technique for OFDM systems over Rayleigh fading channels. In proposed technique,
there are no pilot symbols which added to OFDM. Therefore, this technique is more band-
width efficient compared to pilot-based channel estimation techniques. Also, this technique
is making full use of the learning property of neural network. By using this feature, there
is no need of any matrix computation and the proposed technique is less complex than the
pilot based techniques. Simulation results show that ANN based channel estimator gives bet-
ter results compared to the pilot based channel estimator for OFDM systems over Rayleigh
fading channel.

Keywords OFDM · Channel estimation · Artificial neural network · Cosimulation

1 Introduction

The increasing demand of higher data rates and requirements of broadband in communica-
tion systems leads the interest to multicarrier modulation schemes. Among these multicarrier

C. Çiflikli · A. T. Özşahin (B)

Kayseri Vocational College, Erciyes University, 38039 Melikgazi, Kayseri, Turkey
e-mail: atozsahin@erciyes.edu.tr
C. Çiflikli
e-mail: cebrailc@erciyes.edu.tr

A. Ç. Yapici
Faculty of Engineering, Department of Electrical & Electronics Engineering, Atilim University,
06836 Golbasi, Ankara, Turkey
e-mail: cyapici@atilim.edu.tr

123
222 C. Çiflikli et al.

schemes, OFDM has received considerable attention in the recent past, because of its resis-
tance of multipath fading.
Frequency selective fading caused by multipath could lead to carriers used, being heav-
ily attenuated due to destructive interference at the receiver. This can result in the carriers
being lost in the noise [1]. To increase OFDM system performance under frequency selective
channels, a channel estimation is required before demodulation of OFDM signals [2]. The
channel estimation is a process of characterizing the effect of the transmission medium on
the input signal.
There have been proposed several techniques for channel estimation in OFDM system
[2–14]. Among these techniques, both pilot-based channel estimation and blind channel esti-
mation techniques are most popular. When bandwidth is limited, the pilot-based channel
estimation is not appropriate since inserting pilot signals into OFDM symbol causes waste
of bandwidth [1]. Pilot-based estimation techniques can be based on least-squares (LS),
minimum mean-square error (MMSE) and linear minimum mean-square error (LMMSE)
estimators [2,6]. LS estimators have low complexity, but they suffer from many artifacts
such as a high mean-square error (MSE), especially if the system operates with low sig-
nal-to-noise ratios (SNR). On the other hand, MMSE estimators, based on time-domain
channel statistics, are high-complexity estimators. They have satisfactory performance for
sample-spaced channel environments, but they have limited performance for non-sample-
spaced channels and high SNR’s. Finally, LMMSE estimators have good performance for
both sample-spaced and non-sample-spaced channels [3]. Particularly, a low-complexity fre-
quency-domain LMMSE estimator had been studied by Tolochko and Faulkner [4]. However,
its performance can be further improved based on real time knowledge of the time-domain
channel parameters, especially in the cases; OFDM training symbols used require interpo-
lation at the receiver end, the system operates in the changing channel environments and
system has high SNR’s.
In blind techniques, no training sequence is required. It uses certain underlying mathe-
matical properties of the data being sent. This technique has a drawback of being extremely
computationally intensive. This is the disadvantage of the blind techniques. Because of no
training sequence required, it is appropriate for applications where bandwidth is scarce. This
estimation technique can be based on recursive least square (RLS) or LMS [7,8].
In this paper, we propose an artificial neural network (ANN) based channel estimation
technique as an alternative to pilot based channel estimation technique for OFDM systems
over Rayleigh fading channels. In proposed technique, there are no pilot symbols which
added to OFDM. Therefore, this technique is more bandwidth efficient compared to pilot-
based channel estimation techniques. Moreover, this technique is making full use of the
learning property of neural network. By using this feature, there is no need of any matrix
computation and the proposed technique is less complex than the pilot based techniques.
In addition, we designed a cosimulation environment by using ADS and MATLAB for
simulations in this paper. It provides a powerful simulation environment with using two
different software programs. This powerful simulation environment offers some significant
advantages. We compare the performance of this technique with pilot based channel estimator
for OFDM systems by applying 16QAM and BPSK as modulation schemes over Rayleigh
fading channels. Simulation results show that ANN based channel estimator gives better
results compared to pilot based channel estimator for OFDM systems over Rayleigh fading
channel.
The organization of this paper is as follows. In Sect. 2, description of OFDM system
model is given. ANN based channel estimator is described in Sect. 3. Simulation results are
offered in Sect. 4 and finally, Sect. 5 concludes the paper.

123
ANN Channel Estimation Based on LM for OFDM Systems 223

2 Related Works

ANN’s can perform complex mapping between its input and output space and are capable of
forming complex decision regions with nonlinear decision boundaries [9]. Further, because
of nonlinear characteristics of the ANN’s, these networks of different architecture have found
successful application in channel estimation problem [9–13]. One of the earliest applications
of the ANN in digital communication channel equalization is reported by Siu et al. [10]. ANN
is proposed as a channel estimator for QAM and QPSK constellation, respectively [10] and
[11]. In [12], a three-layer ANN was used to predict channel for MIMO systems. The most
closely related work to our work was proposed for pilot-symbol-aided OFDM systems by
Sun and Yuan [13]. Differences between our study and [13] are that we do not need any
pilot-symbol-aid and we also use levenberg-marquardt (LM) training algorithm to increase
the convergence rate instead of the least mean error (LSE).

3 System Outline

3.1 OFDM System Model

A block diagram of the baseband model of OFDM system is shown in Fig. 1. The binary infor-
mation is first mapped using baseband modulation schemes such as QAM or PSK. Then the
serial-to-parallel conversion is applied to baseband modulated signals. The serial-to-parallel
converted data is modulated using inverse fast Fourier transform (IFFT). After IFFT, the time
domain signal is given by following equation:
s(n) = IFFT(S(k)), n = 0, 1, 2 . . . , N − 1


N −1
= S(k)e j2π kn/N (1)
k=0

where N is the length of FFT, S(k) is baseband data sequence. After IFFT, cyclic prefix is
inserted to prevent inter-symbol interference (ISI). This interval should be chosen to be larger
than the expected delay spread of the multipath channel. The guard time includes the cycli-
cally extended part of the OFDM symbol in order to eliminate the inter-carrier interference
(ICI). The symbol extended with the cyclic prefix is given as follows:

s(N + n), n = −Nc , −Nc + 1, . . . , −1
st (n) = (2)
s(n), n = 0, 1, . . . , N − 1
where Nc is the length of the cyclic prefix. The resultant signal st (n) will be transmitted
over frequency selective time varying fading channel with additive white Gaussian noise
(AWGN). The received signal is given by following equation:
yt (n) = st (n) ⊗ h(n) + w(n) (3)
where h(n) is the impulse response of the frequency selective channel and w(n) is AWGN.
Firstly, at the receiver, the cyclic prefix is removed. Then the signal y(n) without cyclic
prefix is applied to FFT block in order to obtain following equation:
Y (k) = FFT {y(n)} k = 0, 1, 2, . . . , N − 1
N −1
1

= y(n)e− j2π kn/N (4)
N
n=0

123
224 C. Çiflikli et al.

Fig. 1 Block diagram of the baseband model of general OFDM system

After FFT block, assuming there is no ISI, demodulated signal is given by following equation:

Y (k) = S(k)H (k) + W (K ) (5)

where H (k) is FFT[h(n)] and W (k) is FFT[w(n)].

It is obvious from Eq. 5 that before demodulation, the channel estimation should be done
at the receiver, in order to compensate effects of the channel on the received signal.

4 Artificial Neural Networks (ANNs)

Artificial neural networks (ANNs) are one of the popular branches of artificial intelligence
[15]. They have very simple neuron-like processing elements (called nodes or artificial neu-
rons) connected to each other by weighting. The weights on each connection can be dynam-
ically adjusted until the desired output is generated for a given input. An artificial neuron
model consists of a linear combination followed by an activation function. Different types of
activation functions can be utilized for the network; however, the common ones, which are
sufficient for most applications, are the sigmoidal and hyperbolic tangent functions.

Fig. 2 A feed-forward multilayer neural network

123
ANN Channel Estimation Based on LM for OFDM Systems 225

The general structure of an MLPNN is given in Fig. 2 and consists of three layers: an input
layer, with a number of neurons equal to the number of variables of the problem, an output
layer, where the Perceptron response is made available, with a number of neurons equal to the
desired number of quantities computed from the inputs, and an intermediate or hidden layer.
While an MLPNN consisting of only the input and the output layers provide satisfaction for
linear problems, additional intermediate layers are required in order to approximate nonlinear
problems. For example, all problems which can be solved by a perceptron can be solved with
only a hidden layer, but it is sometimes more efficient to use two (or more) hidden layers.

5 Application of ANN as Channel Estimator

Neural networks contain neurons with nonlinear activation functions in the input layer and
the hidden layer and neurons with linear activation functions in the output layer. ANN has
MLP structure, which uses forward propagation neural network. Various training algorithms
exist for MLP. To increase the convergence rate, levenberg-marquardt (LM) algorithm was
used in our work. MLPNN consisting of one input layer, two hidden layers and one output
layer was used to predict channel response.
Block diagram of the proposed ANN based technique is shown in Fig. 3.
The signal xi is OFDM signal which is affected by channel impact. Before demodulation,
a channel estimation technique should be applied to the signal. At the beginning of the esti-
mation process, this complex signal is split into two parts: real and imaginary. These parts
are normalized between −1 and +1 before training. As shown in Fig. 3, each part is trained
and estimated at different neural network. Real part of signal is trained and estimated at NNR
and imaginary part of signal is trained and estimated at NNI. After training and estimation,
these components are merged again and demodulated. The estimated output of the system is
given as

Fig. 3 Artificial neural network structure

123
226 C. Çiflikli et al.

Table 1 Parameters of artificial

Parameter Value
neural networks
Number of inputs 2
Number of hidden layers 2
Number of neurons 5,12
Epoch number 200
Training function Levenberg-Marquart
Transfer functions Tansig

Table 2 Parameters of OFDM

Parameter Value
system
Baseband modulation 16QAM, BPSK
FFT size 2048
Number of carriers 256
Cyclic prefix 32

ŷk = NNR(R[X k ]) + jNNI(I m[X k ]) (6)

here,


m
NNR(R[X k ]) = cr n f (wr n R[X k ] + br n ),
n=1

m
NNI(I m[X k ]) = cin f (win R[X k ] + bin ) (7)
n=1

The activation function f (·) is hyperbolic tangent sigmoid, {wr n , cr n , br n } and {win , cin , bin }
are the weights of subnetwork NNR and NNI, respectively. In order to minimize the error,
we used levenberg-marquardt algorithm.
In Table 1, the parameters of the proposed technique are given. These parameters with
values are obtained with iterative methods. In Table 2, the parameters of the OFDM system
designed in ADS are given.

6 Simulations and Results

In this paper, OFDM signal is generated using advanced design system (ADS) and the ANN
based channel estimator is designed using MATLAB. In the simulations, a cosimulator based
approach is proposed to get RF and digital data flow together.
Frequency selective fading caused by multipath could lead to carriers used, being heav-
ily attenuated due to destructive interference at the receiver. This can result in the carriers
being lost in the noise. In this paper, to increase OFDM system performance under frequency
selective channels, we propose an ANN based channel estimation technique and we design
a cosimulation environment for simulations.
We compare the performance of this technique by applying 16QAM and BPSK as modu-
lation schemes over multipath channel which has 20 reflected paths and modeled to Rayleigh
fading channels. For figure of merits, bit-error rate (BER) has been used. Simulations have
been realized by Agilent ADS and optimizations have been realized by MATLAB. The BER
results of the study have been figured out by MATLAB.
In Fig. 4, a comparison of pilot based channel estimation technique proposed before as
low-pass, second order and decision feedback [5,6] and the proposed ANN based technique

123
ANN Channel Estimation Based on LM for OFDM Systems 227

0
10

-1
10
BER

-2
10

-3 Without Estimation
10
Second Order
ANN
Decision Feedback
Low-Pass
-4
10
5 10 15 20
SNR

Fig. 4 16QAM Modulation with Rayleigh fading

0
10

-1
10
BER

-2
10

-3 Without Estimation
10
Second Order
ANN
Decision Feedback
Low-Pass
-4
10
5 10 15 20
SNR

Fig. 5 BPSK modulation with Rayleigh fading

utilizing 16QAM is shown. The performance among these techniques ranges from best to
the worst as follows: proposed ANN based technique, low-pass, second order and decision
feedback.
In Fig. 5, a comparison of pilot based channel estimation technique proposed before as
low-pass, second order and decision feedback [5,6] and the proposed ANN based technique

123
228 C. Çiflikli et al.

utilizing BPSK is shown. The best performance is also obtained with the proposed ANN
based technique for BPSK. As shown in Figs. 4 and 5, the best results are obtained with
ANN based estimation technique.

7 Conclusion

An artificial neural network channel estimation technique based on levenberg-marquardt

training algorithm has been proposed for OFDM systems over Rayleigh fading channel in
our work. At the end of the simulations using different baseband modulation schemes, BER
analysis of the ANN based estimator is obtained and compared with the pilot based chan-
nel estimation techniques. The proposed estimator is robust to multipath channel estimation
when signal-to-noise ratio is over 10 dB and more bandwidth efficient compared to the pilot
based estimator. We proposed this technique for OFDM systems because of the simple struc-
ture of this system as to MC-CDMA. We still work on adopting this technique as channel
estimator in MC-CDMA.

References

1. Lawrey, E. (1997). The suitability of OFDM as a modulation technique for wireless telecommunications,
with A CDMA comparison. MSc Thesis, James Cook University, October 1997.
2. Colieri, S., Ergen, M., Puri, A., & Bahai, A. (2002). A study of channel estimation in OFDM systems.
In Proceedings of the IEEE 56th Vehicular Technology Conference, 24–28 September 2002 (Vol. 2,
pp. 894–898).
3. Edfors, O., Sandell, M., Wilson, S. K., & Borjesson, P. O. (1998). OFDM channel estimation by singular
value decomposition. IEEE Transactions of Communications, 46, 931–939.
4. Tolochko, I., & Faulkner, M. (2002). Real time LMMSE channel estimation for wireless OFDM systems
with transmitter diversity. In Proceedings of the 56th IEEE VTC, Vancouver, Canada (pp. 1555–1559).
5. Gacanin, H., Takaoka, S., & Adachi, F. (2005). Pilot-assisted channel estimation for OFDM/TDM with
frequency-domain equalization. In Vehicular Technology Conference, 25–28 September 2005, USA.
6. Colieri, S., Ergen, M., Puri, A., & Bahai, A. (2002). Channel estimation techniques based on pilot arrange-
ment in OFDM systems. IEEE Transactions on Broadcasting, 48(3), 223–229.
7. Doukopoulos, X. G., & Moustakides, G. V. (2004). Adaptive algorithms for blind channel estimation in
OFDM systems. In 2004 IEEE International Conference on Communications, 20–24 June 2004 (Vol. 4,
pp. 2377–2381).
8. Roy, S., & Li, C. (2002). A subspace blind channel estimation method for OFDM systems without cyclic
prefix. IEEE Transactions on Wireless Communications, 1, 572–579.
9. Patra, J. C., Pal, R. N., Baliarsingh, R., & Panda, G. (1999). Nonlinear channel equalization for QAM sig-
nal constellation using artificial neural networks. IEEE Transactions on Systems, Man, and Cybernetics,
29(2), 254–262.
10. Siu, S., Gibson, G. J., & Cowan, C. F. N. (1990). Decision feedback equalization using neural net-
work structures and performance comparison with standard architecture. Proceedings of the Institution of
Electrical Engineers, 137(pt. 1), 221–225.
11. Naveed, A., Qureshi, I. M., Cheema, T. A., & Jalil, A. (2004). Blind equalization and estimation of channel
using artificial neural network. In 8th International Multitopic Conference, INMIC 2004 (pp. 184–190).
12. Zhang, L., & Zhang, X. (2007). MIMO channel estimation and equalization using three-layer neural
network with feedback. Tsinghua Science and Technology, 12(6), 658–662.
13. Sun, J., & Yuan, D. F. (2006). Neural network channel estimation based on least mean error algorithm in
the OFDM systems. Advances in Neural Networks. Berlin: Springer.
14. van de Beek, J. J., Edfors, O., Sandell, M., Wilson, S. K., & Borjesson, P. O. (1996). On channel estimation
in OFDM systems. In Proceedings of IEEE VTC’96, November 1996 (pp. 815—819).
15. Haykin, S. (1999). Neural networks: A comprehensive foundation (2nd ed.). Englewood Cliffs, NJ:
Prentice-Hall.

123
ANN Channel Estimation Based on LM for OFDM Systems 229

Author Biographies

Cebrail Çiflikli was born in K. Maras, Turkey, in 1961. He received

the B.S. degree in electrical and electronics engineering from the Black
Sea Technical University, Trabzon, in 1982, the M.S. degree in electrical
and electronics engineering from the Middle East Technical University,
Ankara, Turkey, in 1986, and the Ph.D. degree in electronics engineer-
ing from Erciyes University, Kayseri, Turkey, in 1990. In 1982 he joined
the Faculty of Engineering, Electronic Engineering Department, Erciyes
University, Kayseri, Turkey, as a Research Assistant, he became a Lec-
turer in 1986 and an Assistant Professor in 1991. From 1998 to 2004 he
was involved with lecturer as Associate Professor at the same department.
In 2004, he joined Erciyes University Kayseri Vocational College as Pro-
fessor where he is now Principal. Dr. Ciflikli’s current research interests
include spread-spectrum communications, wireless ATM/LAN, signal
processing, DS-CDMA system engineering, RF power amplifier linear-
ization for wireless communication systems.

A. Tuncay Özşahin was born in Kayseri, Turkey, in 1980. He received

the B.Sc. degree in electronics engineering from the Erciyes University,
Kayseri, in 2002, the M.Sc. degree in electrical and electronics engi-
neering from the Erciyes University, Kayseri, Turkey, in 2006, and is
currently working toward the Ph.D. degree in electronics engineering at
Erciyes University, Kayseri, Turkey. Since 2004 he has been working as
a Instructer at Kayseri Vocational College, Erciyes University, Kayseri,
Turkey. His current research interests are in the areas of wireless sensor
networks, multi-carrier communications and artificial neural network.

A. ÇağrıYapici was born in Adana, Turkey, in 1978. He received the

B.Sc. degree in electrical and electronics engineering from the Hacet-
tepe University, Ankara, in 2001, the M.Sc. degree in electrical and elec-
tronics engineering from the Middle East Technical University, Ankara,
Turkey, in 2004, and is currently working toward the Ph.D. degree in elec-
tronics engineering at Erciyes University, Kayseri, Turkey. From 2000 to
2004 he was with ASELSAN Electronics Industries Inc., Ankara, Turkey.
Since 2004, he has been working as a Research Assistant at Electrical &
Electronics Engineering Department, Erciyes University, Kayseri, Tur-
key. His current research interests are in the areas of RF power amplifier
design and RF power amplifier linearization for wireless communication
systems.

123