Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Int. J. Electron. Commun.

(AE) 59 (2005) 111117


www.elsevier.de/aeue
Anewmethod for adaptive IIRlter design based on tabu search algorithm
Adem Kalinli
a,
, Nurhan Karaboga
b
a
Department of Industrial Electronics, Vocational High School, Erciyes University, 38039 Melikgazi, Kayseri, Turkey
b
Department of Electronic Engineering, Faculty of Engineering, Erciyes University, 38039 Melikgazi, Kayseri, Turkey
Received 17 January 2004; received in revised form 30 March 2004
Abstract
Adaptive digital lters have proven their worth in a wide range of applications such as channel equalisation, noise reduction,
echo cancelling, and system identication. These lters can be broadly classied into two groups: nite impulseresponse
(FIR) and innite impulseresponse (IIR) lters. IIR lters have become the target of increasing interest because these lters
can reduce the lter order signicantly as compared to FIR lters. Tabu search is a heuristic optimisation algorithm which
has been originally developed for combinatorial optimisation problems. It simulates the general rules of intelligent problem
solving and has the ability of discovering the global minima in a multi-modal search space. In this work, a novel method
based on tabu search is described for the design of adaptive IIR lters.
2004 Elsevier GmbH. All rights reserved.
Keywords: Adaptive lters; IIR lter design; Global optimisation; Heuristics; Tabu search
1. Introduction
In order to transform and analyse signals that have been
sampled from analogue sources, digital signal processing
(DSP) algorithms are employed. Any DSP algorithm or
processor can be described as a lter. Digital lters can
be broadly classied into two groups: recursive and non-
recursive lters. The response of non-recursive, or nite
impulseresponse (FIR) lters is dependent only upon
present and previous values of the input signal. Recursive,
or innite impulseresponse (IIR) lters, however, depend
not only upon the input data but also upon one or more
previous output values. In the design of a digital lter, a
priori knowledge about the statistical characteristic of the
input data is required. If the statistical characteristic of
the input data varies with respect to time or the required
knowledge about input data is not satisfactory, adaptive l-
ters are needed. Adaptive lters have proven their worth in

Corresponding author. Tel.: +90 352 4374915; fax: +90 352 4375267.
E-mail addresses: kalinlia@erciyes.edu.tr (A. Kalinli),
nurhan_k@erciyes.edu.tr (N. Karaboga).
1434-8411/$ - see front matter 2004 Elsevier GmbH. All rights reserved.
doi:10.1016/j.aeue.2004.11.003
a wide range of applications such as channel equalisation,
noise reduction, echo cancelling, and system identication.
Much research has been conducted in adaptive FIR lters
since these lters have convenient mathematical properties.
However, FIR lters are not suitable for pole-zero models
because an FIR lter requires a high lter order to model
the poles within an acceptable error level. Because of this
drawback of adaptive FIR lters, adaptive IIR lters have
become the target of increasing interest because pole-zero
lters may reduce the lter order signicantly as compared
to FIR lters.
Although an adaptive IIR lter can provide a much better
performance than the adaptive FIR lter having the same
number of coefcients, some problems arise in the design
of adaptive IIR lters [13]. The fundamental problem is
that they might have a multi-modal error surface. A further
problem is the possibility of the lter becoming unstable
during the adaptation process. This second problem can be
easily handled by limiting the parameter space. In order to
avoid the rst problem, a design method which can achieve
the global minima in a multi-modal error surface is required.
Therefore, some researchers have attempted to develop this
112 A. Kalinli, N. Karaboga / Int. J. Electron. Commun. (AE) 59 (2005) 111117
type design methods such as those based on simulated an-
nealing (SA) [46] and genetic algorithms (GA) [710].
SA and GA employ probabilistic transition rules to nd
the global minima in a multi-modal error surface. GA is a
population based algorithm and evolves a population of so-
lutions to the problem as SA attempts to improve a single
solution using a neighbourhood search mechanism [11]. Al-
though SA algorithm is quite easy to be programmed and
good at local convergence, depending on the initial solution
it might often require too many cost function evaluations
to converge to the global minima. GA usually discovers the
promising regions of search space very quickly, however it
often needs too many computations to reach a local minima
since the probabilistic transition rules are employed [12] and
a neighbourhood search mechanism is not used. These dis-
advantages of both algorithms are not desired in the design
of adaptive IIR lters.
Tabu search (TS) algorithm which is based on the general
rules of intelligent problem solving has been originally de-
veloped for combinatorial optimisation problems [13]. It has
been also applied to solve numeric optimisation problems
[14]. Although TS algorithm uses neighbourhood search
principle as SA algorithm and employs the deterministic
transition rules rather than probabilistic ones, it also has
the ability of discovering global minima in a multi-modal
search space. TS uses exible memory storing information
about the past steps of search to avoid the evaluation of the
same solutions several times. This feature improves the con-
verging speed of search to global minima. Therefore, TS
seems to be as a promising algorithm for designing adap-
tive IIR lters. Although TS has been successfully used for
designing adaptive and non-adaptive FIR lters [1518], to
our best knowledge it has not been applied to design adap-
tive IIR lters, yet. In this work, a novel method based on
TS algorithm is proposed for the design of adaptive IIR l-
ters. The results obtained by the proposed novel method are
compared to those of adaptive simulated annealing (ASA)
algorithm [19], gradient lattice algorithm and genetic adap-
tive algorithm [20]. Section 2 briey describes the problem.
Principles of TS are presented in Section 3. In Section 4,
the application of TS to the problem is explained and the
simulation results are presented.
2. Denition of the problem
In this section, the problem of adaptive IIR lter de-
sign is dened. Consider the adaptive IIR lter with the
inputoutput relationship governed by
y(k) +
M

i=1
b
i
y(k i) =
L

i=0
a
i
x(k i), (1)
where x(k) and y(k) are the lters input and output, re-
spectively, and M ( L) is the lter order. The transfer
function of this IIR lter can be written in the following
general form:
H(z) =
A(z)
B(z)
=

L
i=0
a
i
z
i
1 +

M
i=1
b
i
z
i
. (2)
Hence, the design of an adaptive IIR lter can be consid-
ered as an optimisation problem of the cost function J(w)
stated as the following:
min
wW
J(w), (3)
where w =[a
0
a
1
...a
L
b
1
...b
M
]
T
is the lter adjustable coef-
cient vector.
The aimis to minimise the cost function J(w) by adjusting
w by means of an algorithm. The cost function is usually
expressed as the time-averaged cost function dened by
J(w) =
1
N
N

k=1
[d(k) y(k)]
2
, (4)
where d(k) and y(k) are the lters desired and actual re-
sponses, respectively, and N is the evaluation window size
used for the calculation of cost-function value.
3. Tabu search algorithm
Tabu search is a general heuristic search procedure de-
vised for nding a global minimum of a function f (x). The
problem of searching the optimum value of x which makes
f (x) minimum is called the optimisation problem of f (x)
and can be mathematically expressed by
Minimise f (x),
Subject to x X.
The function of f (x) may be linear or non-linear and the
condition x X describes the constraints on the solution x.
A step of the TS starts with a present solution x
now
. x
now

X, has an associated set of feasible solutions Q which can
be obtained by applying a simple modication to x
now
. This
modication is called a move. In order to be able to get rid
of a local minima, a move to the neighbour x

, is created
even if x

is worse than x
now
. This would cause the cycling
of the search. In order to avoid the cycling problem, a tabu
list T is introduced. The tabu list stores all the tabu moves
that cannot be applied to the present solution, x
now
. The
moves stored in the tabu list are those carried out most fre-
quently and recently according to some criteria called tabu
restrictions. The use of tabu list decreases the possibility of
cycling because it prevents returning in a certain number
of iterations to a solution visited recently. After a subset of
feasible solutions Q

is produced according to the tabu list


and evaluated for f (x) the next solution is selected from it.
The highest evaluation solution is selected as the next solu-
tion x
next
. This loop is repeated until a stopping criteria is
satised.
A. Kalinli, N. Karaboga / Int. J. Electron. Commun. (AE) 59 (2005) 111117 113
A tabu restriction is activated when the reverse of a move
recorded in the tabu list occurs within a predetermined num-
ber of iterations or with a certain frequency over longer
range of iterations. The former produces a recency-based
restriction and the latter a frequency-based restriction. Tabu
restrictions might sometimes prevent the search to nd the
solutions which have not been visited yet or even cause all
available moves to be classied as the tabu. Therefore, it
should be possible to forget the tabu constraints when a free-
dom is required for the search. A criterion called aspiration
criteria is employed to determine which moves should be
freed in such cases.
The owchart of a basic TS is presented in Fig. 1. In the
initialisation unit, a random feasible solution x
initial
X for
the problem is generated, and the tabu list and other param-
eters are initialised. In the neighbour production unit, a fea-
sible set of solutions is produced from the present solution
according to the tabu list and aspiration criteria. The eval-
uation unit evaluates each solution x

produced from the


present x
now
one. After the next solution x
next
is determined
by the selection unit, in the last unit the history record of the
search is modied. If the next solution determined is better
than the best solution found so-far x
best
, the next solution is
replaced with the present best solution.
In this work, a string of binary numbers is used to repre-
sent a possible solution to the problem and TS employs the
following two constraints which are based on recency and
frequency memories as the tabu constraints:
recency(x

)restriction_period
frequency_ratio(x

)frequency_limit
where restriction_period = recency factor length of
solution string and frequency_limit = frequency factor
average_frequency.
Final solution
Initial solution
Creation of the
solutions Q
*
Evaluation
Selection
Memory
modification
Fig. 1. Flowchart of a basic tabu search.
The recency of a move is the difference between the cur-
rent iteration count and the last iteration count at which that
move was created. The frequency measure is the frequency
ratio whose numerator represents the count of the number
of occurrences of a specic move and whose denominator
represents the average numerator value over all possible
moves.
In the algorithm used, the highest evaluation move is
selected as the next solution. The aspiration by default
is employed as the aspiration criteria. According to this
criterion the least tabu solution is selected as the next so-
lution. It is the solution which loses its tabu classication
by the least increase in the value of the present iteration
number.
4. Application of TS to the problem and
simulation results
Application of adaptive IIR lters in system iden-
tication have been widely studied since many prob-
lems encountered in signal processing can be char-
acterised as a system identication problem (Fig. 2).
Therefore, in the simulation study, adaptive IIR lters
are designed for system identication purpose. In this
case, the parameters of the adaptive lters are succes-
sively adjusted by TS algorithm until the error between
the outputs of the lter and the unknown system is
minimised. By constraining the range of the lter co-
efcients, the stability is guaranteed. The lter coef-
cients are encoded in the binary string form as shown
in Fig. 3 and 10 bits are used to represent each l-
ter parameter. In simulations, the following values were
used for the control parameters of TS algorithm. The
value of frequency and recency factors were 2 and 0.2,
respectively.
-
+
adaptive IIR filter
Tabu Search
unknown plant
x(k)
d(k)
y(k)
e(k)
Fig. 2. Block diagram of system identication process using adap-
tive IIR lter.
a
0
a
1
a
L
b
1
b
2
b
M
... ...
Fig. 3. Representation of the parameters in the string form.
114 A. Kalinli, N. Karaboga / Int. J. Electron. Commun. (AE) 59 (2005) 111117
The performance of the solution i is determined by using
the following formula
f it (i) =
1
1 +J
i
(w)
, (5)
where J
i
(w) is the cost function value calculated by using
the solution i.
A move producing a possible neighbour solution x

is
created by reversing the value of a binary bit of the present
solution x
now
. In order to calculate the performance of a
neighbour solution, a moving scheme is employed: The cost
function dened by Eq. (4) is calculated by using a block
of N samples (N =100) and the data block is shifted by 1
sample after every iteration.
Simulation studies have been carried out on widely used
three examples.
Example 1. In the rst example, the unknown plant and the
lter had the following transfer functions [19]:
H
p
(z) =
0.05 0.4z
1
1 1.1314z
1
+0.25z
2
,
H(z) =
a
0
1 +b
1
z
1
. (6)
The input x(k), to the system and the lter was a white
sequence. Since the lter order is smaller than the system or-
der, a local minima problem occurs. The MSE (mean-square
error) has a global minimum at w
global
=[0.311, 0.906]
T
with the value of the NMSE (normalised MSE) 0.2772 and
a local minimum at w
local
= [0.114, 0.519]
T
. Fig. 4 shows
the evolution of the normalised MSE averaged over 100 dif-
ferent runs of TS and ASA algorithm used in [19]. Each run
had a randomly chosen initial w. Fig. 5 demonstrates the tra-
jectories of the lter parameter vector for 4 different initial
points.
Example 2. In the second example, the plant was a
third-order system and the lter was a second-order IIR
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 100 200 300 400 500
Global minimum
ASA
TS
Number of Function Evaluations
A
v
e
r
a
g
e
d

C
o
s
t

F
u
n
c
t
i
o
n
Fig. 4. Normalised cost function value versus number of cost func-
tion evaluations averaged over 100 random runs of both algorithms
(Example 1).
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
a 0
b
1
.3
.8
.98
1.4
1.2
1.0
a
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
b
1
a 0
b
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
b
1
a 0
c
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
b
1
a 0
d
Fig. 5. Trajectories of parameter vector for 4 different initial points
when the proposed method is employed: (a) [0.114, 0.519]
T
(local
minima); (b) [0.97, 0.06]
T
; (c) [0.92, 0.78]
T
; (d) [0.94, 0.80]
T
.
A. Kalinli, N. Karaboga / Int. J. Electron. Commun. (AE) 59 (2005) 111117 115
lter with the following transfer functions [19]:
H
p
(z) =
0.3 +0.4z
1
0.5z
2
1 1.2z
1
+0.5z
2
0.1z
3
,
H(z) =
a
0
+a
1
z
1
1 +b
1
z
1
+b
2
z
2
. (7)
The system input was a uniform white sequence taking
values in [0.5, 0.5], and the SNR was 30 dB. The data
length used in calculating the MSE was N = 2000. Since
the reduced order lter is employed the MSE is multi-
modal. The simulation results showed that the MSE has a
global minimum of 0.059. Fig. 6 presents the normalised
cost function value versus number of cost function evalua-
tions averaged over 100 random runs of both algorithms for
the Example 2. Fig. 7 shows the behaviour of the standard
gradient algorithm when it started from two quite closed
initial solutions for the same example [19].
From Figs. 4 and 6 it is seen that the proposed method
reaches the global minima much quicker than ASA [19].
0.01
0.1
1
0 200 400 600 800 1000
TS
ASA
Global minimum
A
v
e
r
a
g
e
d

C
o
s
t

F
u
n
c
t
i
o
n
Number of Function Evaluations
Fig. 6. Normalised cost function value versus number of cost func-
tion evaluations averaged over 100 random runs of both algorithms
(Example 2).
0.01
0.1
1
0 50 100 150 200
(a)
(b)
Number of Iterations
C
o
s
t

F
u
n
c
t
i
o
n
Fig. 7. Convergence behaviours of the batch gradient algorithm,
started from the two initial conditions: (a) [0.0, 0.0, 0.3, 0.1]
T
and
(b) [0.0, 0.0, 0.3, 0.0]
T
, for Example 2.
For the rst example, TS algorithm needs about 100 evalu-
ations as ASA requires around 250 evaluations to reach the
global minima. For the second example, TS needs only 280
evaluations as ASA requires 440 evaluations.
Example 3. The plant was a sixth-order system and the
lter was a fth-order IIR lter with the following transfer
functions [20]:
H
p
(z) =
1.0 1.8z
2
+1.04z
4
+0.05z
5
+0.192z
6
1 0.8z
2
0.17z
4
0.56z
6
,
H(z) =
a
0
+a
1
z
1
+ +a
5
z
5
1 +b
1
z
1
+ +b
5
z
5
. (8)
The input to both the plant and adaptive lter was Gaus-
sian pseudo-noise sequence. The error surface for this ex-
ample is bimodal with a global minimum below the 18 dB
NMSE level.
Fig. 8 presents the results obtained with gradient lattice
algorithm. The gradient lattice algorithm converged to a
NMSE value of only 2 dB. This result shows that the gra-
dient lattice algorithm has failed to locate the global opti-
mum. The results of the simulations using the oating-point
genetic adaptive algorithms are presented in Fig. 9 [20]. The
graph in Fig. 9 shows the NMSE of the best model in the
population (averaged over 50 independent runs of the algo-
rithm) [20]. The genetic adaptive algorithmemployed in [20]
uses a population of 800. From Fig. 9, it is apparently seen
that the algorithm requires more than 400 000 (800 500
generation) evaluations to design a model able to produce
a NMSE below 11 dB. Fig. 10 shows the NMSE of the
model produced by TS algorithm (averaged over 50 inde-
pendent runs of the algorithm). From Fig. 10, it is clear that
TS algorithm requires less than 50 000 evaluations to design
a model able to produce a NMSE below 11 dB.
Fromthe results obtained for the rst and second examples
presented in Figs. 4 and 6, it can be concluded that the
proposed approach is more robust than ASA algorithm. The
results given in Figs. 810 also demonstrate that TS more
-10
0
10
20
0 20000 40000 60000 80000 100000
Time Samples
N
o
r
m
a
l
i
z
e
d

M
e
a
n

S
q
u
a
r
e
d

E
r
r
o
r

(
d
B
)
Fig. 8. Simulation results using the gradient lattice algorithm for
the sixth-order lter.
116 A. Kalinli, N. Karaboga / Int. J. Electron. Commun. (AE) 59 (2005) 111117
-15
-10
-5
0
0 20000 40000 60000 80000 100000
N
o
r
m
a
l
i
z
e
d

M
e
a
n

S
q
u
a
r
e
d

E
r
r
o
r

(
d
B
)
Generations
Fig. 9. Simulation results using genetic adaptive algorithm for the
sixth-order lter.
-15
-10
-5
0
5
10
0 50000 100000 150000
N
o
r
m
a
l
i
z
e
d

M
e
a
n

S
q
u
a
r
e
d

E
r
r
o
r

(
d
B
)
Number of Function Evaluations
Fig. 10. Normalised MSE value versus number of cost function
evaluations averaged over 50 random runs of TS algorithm (Ex-
ample 3).
rapidly convergences to the optimal solution than the genetic
adaptive algorithm and gradient lattice algorithm.
5. Conclusion
A novel method based on tabu search algorithm is pro-
posed for the design of digital adaptive IIR lters. The
method is able to discover the global minima of a multi-
modal error surface since a exible memory is used by the
tabu search algorithm. The new method requires less cost-
function evaluations than the methods employing probabilis-
tic rules such as genetic algorithm and simulated annealing
to converge to global minima. From the simulation results
it was concluded that the tabu search algorithm can be ef-
ciently used for adaptive digital IIR lter design.
Acknowledgements
This work has been supported by the Erciyes University
Research Fund project 02-012-3.
References
[1] Stearns SD. Error surface of recursive adaptive lters. IEEE
Trans Acoustics, Speech Signal Process 1981;29:7636.
[2] Shynk JJ. Adaptive IIR ltering. IEEEASSP Mag 1989;421.
[3] Radenkovic M, Bose T. Adaptive IIR ltering of nonstationary
signals. Signal Process 2001;81(1):18395.
[4] Nambiar R, Mars P. Genetic and annealing approaches to
adaptive digital ltering. IEEE 26th Asilomar Conference on
Signal Systems and Computers 1992; 8715.
[5] Radecki J, Konrad J, Dubois E. Design of multidimensional
nite-wordlength FIR and IIR lters by simulated annealing.
IEEE Trans on Circuits Systems II: Analog Digital Signal
Process 1995;42(6):42431.
[6] Chen S, Istepanian RH, Luk BL. Digital IIR lter design
using adaptive simulated annealing. Digital Signal Process
2001;11(3):24151.
[7] Etter DM, Hicks MJ, Cho KH. Recursive adaptive lter design
using an adaptive genetic algorithm. IEEE Int Conf ASSP
1982; 6358.
[8] Tang KS, Man KF, Kwong S, He Q. Genetic algorithms
and their applications. IEEE Signal Process Mag 1996;13(6):
2237.
[9] Ng SC, Leung SH, Chung CY, Luk A, Lau WH. The genetic
search approach: a new learning algorithm for IIR ltering.
IEEE Signal Process Mag 1996;13(6):3846.
[10] Thamvichai R, Bose T, Haupt RL. Design of 2-D
multiplierless IIR lters using the genetic algorithm.
IEEE Trans Circuits Systems-I: Fundamental Theory Appl
2002;49(6):87882.
[11] Kirkpatrick S, Gelatt Jr. CD, Vecchi Jr. MP. Optimization by
simulated annealing. Science 1983;220(4598):67180.
[12] Holland JJ. Adaptation in natural and articial systems. Ann
Arbor, MI: The University of Michigan Press; 1975.
[13] Glover F. Future paths for integer programming and links to
articial intelligence. Comput Oper Res 1986;5:53349.
[14] Pham DT, Karaboga D. Intelligent optimisation techniques:
genetic algorithms, tabu search, simulated annealing and
neural networks. London: Springer; 2000.
[15] Karaboga D, Horrocks DH, Karaboga N, Kalinli A.
Designing digital FIR lters using tabu search algorithm.
IEEE International Symposium on Circuits and Systems
(ISCAS97), Hong Kong 1997. 22369.
[16] Fanni A, Marchesi M, Pilo F, Serri A. Tabu search
metaheuristic for designing digital lters. COMPEL Inst J
Comput Math Electr Electron Eng 1998;17(5/6):78996.
[17] Belbachir AN, Belbachir MF, Fanni A, Bibbo S. Boulerial:
a new approach to digital FIR lter design using the
tabu search. IEEE Nordic Signal Processing Symposium
(NORSIG 2000), 2000. 2236.
[18] Traferro S, Uncini A. Power-of-two adaptive lters using tabu
search. IEEE Trans Circuits Systems-II: Analog and Digital
Signal Process 2000;47(6):5669.
[19] Chen S, Luk BL. Adaptive simulated annealing for
optimization in signal processing applications. Signal Process
1999;79:11728.
[20] White MS, Flockton SJ. Adaptive recursive ltering using
evolutionary algorithms. In: Dasgupta D, Michalewicz Z.,
A. Kalinli, N. Karaboga / Int. J. Electron. Commun. (AE) 59 (2005) 111117 117
editors. Evolutionary algorithms in engineering applications.
Berlin: Springer; 1997. 36176.
Adem Kalinli received the B.Sc.,
M.Sc. and Ph.D. degrees in Electron-
ics Engineering from Erciyes Univer-
sity, Turkey, in 1990, 1993 and 1996,
respectively. He is now an assistant
professor of Industrial Electronics Di-
vision of Vocational High School at
Erciyes University of Turkey. He also
holds various positions at Erciyes
University. His current research inter-
ests includes tabu search, ant colony
algorithm, genetic algorithms, articial immune algorithm, sim-
ulated annealing, neural networks and their engineering applica-
tions, particularly electronics engineering problems.
Nurhan Karaboga was born in Os-
maniye, Turkey, on December, 1965.
She received the B.Sc., M.Sc. and
Ph.D. degrees in Electronic Engineer-
ing from Erciyes University, Kayseri,
Turkey, in 1987, 1990 and 1995, respec-
tively. She was Research Assistant from
1987 to 1995 in Electronic Engineer-
ing Department at Erciyes University.
From 1992 to 1994 she also worked
as academic visitor in the University
of Wales College of Cardiff, UK. Since 1995 she has been
an assistant professor in the Department of Electronic Engi-
neering at Erciyes University. Her current research interests
includes genetic algorithms, ant colony algorithms, simulated
annealing algorithm, tabu search, digital lter design and data
communication.

You might also like