Bindiya 2016

JID: INS
ARTICLE IN PRESS [m3Gsc;January 5, 2016;17:54]
Information Sciences xxx (2016) xxx–xxx
Contents lists available at ScienceDirect
Information Sciences
journal homepage: www.elsevier.com/locate/ins
Meta-heuristic evolutionary algorithms for the design of

optimal multiplier-less recombination filter banks
Q1 T.S. Bindiya∗, Elizabeth Elias
Q2
Department of Electronics and Communication Engineering, National Institute of Technology, Calicut, Kerala 673601, India
a r t i c l e i n f o a b s t r a c t
Article history: This paper proposes a design for multiplier-less recombination non-uniform filter banks
Received 11 April 2014 (RNUFBs) optimized using meta-heuristic algorithms. The structure consists of an M-
Revised 8 December 2015
channel uniform filter bank, with some channels combined by the synthesis filters of a
Accepted 15 December 2015
transmultiplexer (TMUX), yielding non-uniform sub-bands. When any structure is realized
Available online xxx
in hardware, it is necessary to have low power consumption and a small chip area. These
Keywords: can be achieved by replacing the multipliers with shifters and adders. Once the continu-
Recombination non-uniform filter banks ous coefficient recombination non-uniform filter bank is designed, the coefficients are con-
Canonic signed digit (CSD) verted to the canonic-signed-digit (CSD) space to make the design multiplier-less, so as to
Meta-heuristic algorithms reduce the complexity of the hardware implementation. To reduce the number of adders
Harmony search algorithm (HSA) and shifters in the multiplier-less implementation, the filter coefficients are rounded with
Artificial bee colony algorithm (ABC) a restricted number of signed power-of-two (SPT) terms, which may cause degradation
Gravitational search algorithm (GSA) in the performance of the RNUFBs. To improve the performance of the CSD rounded filters
and filter bank, meta-heuristic algorithms such as the artificial bee colony (ABC) algorithm,
harmony search algorithm (HSA) and gravitational search algorithm (GSA) are deployed. Of
these meta-heuristic algorithms, GSA is found to give the best performance. The method
proposed in this paper results in non-uniform filter banks with rational sampling factors
which are multiplier-less and have linear-phase and near-perfect-reconstruction.
© 2016 Published by Elsevier Inc.
1 1. Introduction
2 In many applications such as image coding, non-uniform filter banks (NUFBs) are required, which divide the frequency
3 spectrum of the input signal into non-uniform sub-bands [61]. For example, in the design of filter banks that approximate
4 the time-frequency resolution of human ears, non-uniform spacing that matches the critical bands is preferred [57]. The
5 constraints of uniform filter banks, such as integer and uniform decimation in each sub-band and limited time frequency
6 resolution, increase the importance of NUFBs. Many design methods are proposed in the literature to implement NUFBs
7 [13,17,18,22,31,33,36,39,46,48,54,61,66–68,73].
8 Hoang and Vaidyanathan [22] have proposed a new structure for non-uniform quadrature mirror filter (QMF) banks with
9 conditions for perfect-reconstruction (PR) NUFBs. Another method, known as the indirect or recombination method, where
10 certain channels in an M-channel uniform filter bank are combined using the synthesis filters of another filter bank with a
11 smaller number of channels, is proposed in [13]. In [31], the problem of designing NUFBs with rational decimation factors
12 is converted into the design of an NUFB with integer decimators. The approach to designing near-perfect reconstruction
Q3 ∗
Corresponding author. Tel.: +91 9496818039.
E-mail addresses: bindiyajayakumar@nitc.ac.in (T.S. Bindiya), elizabeth@nitc.ac.in (E. Elias).
http://dx.doi.org/10.1016/j.ins.2015.12.018
0020-0255/© 2016 Published by Elsevier Inc.
Please cite this article as: T.S. Bindiya, E. Elias, Meta-heuristic evolutionary algorithms for the design of optimal multiplier-
less recombination filter banks, Information Sciences (2016), http://dx.doi.org/10.1016/j.ins.2015.12.018
JID: INS
2 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx
13 (NPR) critically sampled NUFBs proposed by Princen [48] is based on modulated filter banks and the adjacent band alias
14 cancellation principle. Of these various methods for designing NUFBs, only a few have linear-phase. The tree-structured
15 method proposed in [61] is the simplest method which can be used to design NPR alias-free linear-phase (LP) non-uniform
16 filter banks. An iterative algorithm is proposed in [33] by Kumar et al. for the design of the NPR tree-structured non-
17 uniform filter bank. However, the unequal sampling factors that can be implemented using this method are limited to
18 powers of two and there is a significant delay because of their cascaded structure. Nagai et al. proposed a new approach
19 to the design of linear-phase PR NUFBs based on frequency domain constraints to eliminate the amplitude and the aliasing
20 distortions [46]. The PR condition and a design procedure for LP finite impulse response (FIR) filter banks using pseudo-
21 QMF are derived in [64]. In this method, however, FIR filters are required to have complex coefficients to reduce the aliasing
22 distortion in the resulting NUFBs. A method for designing LP NUFBs with rational decimators is proposed in [36,66,67] based
23 on the recombination structure given by Cox [13]. Another direct design method for LP NPR NUFBs with integer sampling
24 factors is proposed by Xie et al. [68]. This method is extended to the rational case in [12]. The recombination non-uniform
25 filter bank (RNUFB) is a good choice for realizing NUFBs with rational sampling factors, in which some channels of an M-
26 channel uniform filter bank (UFB) are merged by the synthesis filters of uniform transmultiplexers (TMUXs) [30] with a
27 smaller number of channels in order to generate sub-bands with different sampling factors [36,66–68]. The advantage of
28 the recombination NUFB is that the PR property is structurally imposed if the original UFB and the TMUX have PR property.
29 This paper is based on the recombination filter banks proposed in [36], in which the constituent PR uniform filter banks are
30 designed using the direct method proposed in [52].
31 FIR filters and filter banks require considerably more arithmetic operations and hardware components than their infinite
32 impulse response (IIR) counterparts. This makes their implementation very expensive. Multipliers, being relatively complex
33 units, are the deciding factors in the overall speed, area and power consumption of digital signal processing systems. With
34 the current advancements in VLSI technologies, fast multipliers, operating at speeds exceeding 100 MHz, are available. They
35 employ parallel processing, which requires a large chip area and high power. If the filter banks are employed in high-speed
36 applications such as real-time image compression systems, a separate fast multiplier is required for each filter coefficient,
37 which will result in high hardware complexity [23]. Hence, the need to eliminate the multipliers using equivalent blocks is
38 essential for hardware-efficient signal processing systems. One such popular approach is to replace the multipliers by means
39 of simple shifters and adders/subtractors, which can result in reduced area and power consumption [37]. The implementa-
40 tion approach in which the digital filters are designed with no multipliers is referred to as a multiplier-less or multiplier-free
41 approach. This is achieved by representing the filter coefficients in the signed power-of-two (SPT) space [38]. Canonic signed
42 digit representation [21] is a special case of the SPT space, which uses a minimum number of non-zero bits (SPT terms) to
43 represent a decimal number. The number of shifters and adders required for the implementation of the multiplier-less filter
44 depends on the total number of SPT terms in the filter coefficient representation. Hence, the CSD-represented filter offers
45 minimum complexity, area and power consumption compared to other representations.
46 In order to realize multiplier-less filter banks, the continuous coefficient filter banks are designed first by designing
47 the continuous coefficient analysis and synthesis filters. The next phase is to convert them into the CSD space to obtain
48 multiplier-less filters and multiplier-less filter banks. To bring the complexity down further, the number of adders in the
49 CSD-represented multiplier-less implementation is reduced. To achieve this, the coefficients of the constituent filters in the
50 filter bank are rounded to the discrete values in the CSD space with a restricted number of SPT terms. This rounding of the
51 coefficients may cause a deterioration of the frequency responses of the filters and the filter bank and calls for the use of
52 suitable optimization techniques to improve the response. The higher the number of non-zero bits used to represent the
53 filter coefficients, the closer the rounded filter coefficients are to the unrounded coefficients and the closer the rounded per-
54 formance is to the unrounded performance. However, a higher number of non-zero bits means greater hardware complexity
55 and a higher cost. Thus, there should be a trade-off between the hardware cost and performance characteristics. Because
56 the coefficient space is discrete, the optimization methods developed for the continuous variables are not applicable here.
57 Meta-heuristic algorithms are good alternatives for these types of optimization problems because it is reported that these
58 can finally reach a global solution if the parameters are properly selected with respect to a particular design problem [70].
59 They are especially useful in finding near-optimal solutions in multi-modal, multi-objective, and multidimensional prob-
60 lems. Various nature-inspired meta-heuristic algorithms are employed in various optimization applications in the literature
61 [1,5–9,11,16,24,25,40–43,53,58,62,63,65].
62 The design of multiplier-less reconfigurable channel filters using a ternary-coded genetic algorithm (GA) is reported in
63 [8]. Integer-coded GA and integer-coded artificial bee colony (ABC) algorithms, in which the integer indices of look-up-
64 table entries are used to arrive at the solution, are used for the optimization of the multiplier-less transmultiplexer in
65 [40] and [41], respectively. A design for a multiplier-less frequency response masking (FRM) filter using an integer-coded
66 ABC algorithm and differential evolution (DE) algorithm is presented in [42]. In [4–7], an integer-coded ABC algorithm, DE
67 algorithm, harmony search algorithm (HSA) and gravitational search algorithm (GSA) are proposed for the optimization of
68 reconfigurable channel filters and tree-structured non-uniform filter banks. No design for a multiplier-less recombination
69 filter bank using meta-heuristic algorithms has been reported in the literature so far. In this paper, various meta-heuristic
70 algorithms are proposed for the design of multiplier-less non-uniform RNUFBs.
71 This paper is organized as follows: Section 2 gives an overview of the recombination non-uniform filter banks. In
72 Section 3, the proposed methodology, for designing multiplier-less RNUFBs using meta-heuristic algorithms is discussed.
73 This section also discusses the various meta-heuristic algorithms such as DE, ABC, HSA and GSA. Section 4 gives an example
JID: INS
T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 3
Fig. 1. Structure of recombination NUFB [36].
74 of a design for a two-channel recombination non-uniform filter bank with rational sampling factors {2/3, 1/3} and its pro-
75 posed optimal multiplier-less implementation. Section 5 illustrates the design example for a continuous and multiplier-less
76 four-channel recombination non-uniform filter bank with sampling factors {2/5, 1/5, 1/5, 1/5}. Section 6 concludes the paper.
77 2. Review of recombination filter banks
78 In the recombination filter bank structure, the required channels of an M-channel uniform filter bank are combined
79 by the synthesis filters of a TMUX with a smaller number of channels [36,66,67] as shown in Fig. 1a. Hi (z) and Fi (z) for
80 i = 0, 1, . . . ,M − 1 are respectively the analysis and synthesis filters of the M-channel uniform filter bank. Gl, i (z) and Gl,i (z )
81 for i = 0, 1, . . . ,ml − 1 are respectively the synthesis and analysis filters of the ml -channel uniform TMUX. ↓M and ↑M re-
82 spectively indicate downsampling and upsampling by M. The merging operation is performed by employing the synthesis
83 filters Gl, i (z) of the ml -channel TMUX. A PR uniform TMUX can be derived using a 1-skewed FB [30]. Let {Gl,i (z ), Gl,i_F B (z )}
_F B
84 and {Gl,i (z ), Gl, i (z)} be the analysis/synthesis filter pairs of a uniform FB and a 1-skewed FB respectively. A 1-skewed FB can
85 be derived from the uniform FB using the following equations [30].
Gl,i (z ) = Gl,i_F B (z ) Gl,i (z ) = z−1 Gl,i_F B (z ) (1)
86 If M and ml are co-prime, the equivalent structure of the recombination filter bank can be derived [36,66,67] and is
87 shown in Fig. 1b. The constants ci and ci−1 , i = 0, 1, . . . ,M − 1 are used to control the magnitude responses of the equivalent
88 filters Ĥ (z ) using the expression below [36,67]:
ml −1

Ĥ (z ) = crl +i Hrl +i (zml )Gl,i (zM ) (2)
i=0
89
ml −1

F̂ (z ) = cr−1+i Frl +i (zml )Gl,i (zM ) (3)
l
i=0
90 The advantage of RNUFBs is that if the original and recombination FBs are PR, then the entire RNUFB will be PR, if the
91 different delays, introduced by the transmultiplexers in each branch, are compensated. Here, the M-channel and ml -channel
92 uniform filter banks are derived using the method given in [52]. Let Hi (z) and Fi (z), i = 0, 1, . . ., M − 1 be the analysis and
JID: INS
Fig. 2. Structure of M-channel uniform filter bank [61].
93 synthesis filters respectively of a maximally decimated M-channel uniform filter bank as shown in Fig. 2 [61]. If X(z) is the
94 input signal, the output signal X̂ (z ) of the uniform filter bank can be represented as
1
M−1 M−1
X̂ (z ) = Hk (zW l )Fk (z )X (zW l ) (4)
M
k=0 l=0
95
1
M−1 M−1
= T ( z )X ( z ) + Hk (zW l )Fk (z )X (zW l ) (5)
M
k=0 l=1
96 where W = e− j2π /M and the distortion transfer function T(z) of the uniform filter bank is given by
1
M−1
T (z ) = Hk (z )Fk (z ) (6)
M
k=0
97 The aliasing distortion in the uniform filter bank can be cancelled if the synthesis filters are selected in terms of the analysis
98 filters as given below [52]:
(−1 )k F0 (z )|H−0k |
Fk (z ) = for k = 0, 1, . . ., M − 1 (7)
|H−00 |
Q4
99
where F0 = |H−00 | and H− jl is the alias component (AC) matrix with the ( j + 1 )th row and the (l + 1 )th column deleted.
100 The AC matrix H is given in [52]:

H0 (z ) H1 (z ) . . . HM−1 (z )
H (zW ) H1 (zW ) HM−1 (zW )

0 . . .
. . . .
(8)
. . . .

. . . .
H (zW M−1 ) H1 (zW M−1 ) . . . HM−1 (zW M−1
)
0
101 The transfer function T(z) of the uniform filter bank given in Eq. (6) can be rewritten as
1
M−1
T (z ) = (−1 )k Hk (z )|H−0k | (9)
M
k=0
102 For example, in the case of a three-channel uniform filter bank, to guarantee that it is alias free, the synthesis filters can
103 be designed using Eq. (7) and are given by
F0 (z ) = H1 (zW )H2 (zW 2 ) − H2 (zW )H1 (zW 2 ) (10)
104
F1 (z ) = H2 (zW )H0 (zW 2 ) − H0 (zW )H2 (zW 2 ) (11)
105
F2 (z ) = H0 (zW )H1 (zW 2 ) − H1 (zW )H0 (zW 2 ) (12)
106 With the above choice for the synthesis filters, the Z-transform of the output now becomes
1
Y (z ) = X (z ) H0 (z )F0 (z ) + H1 (z )F1 (z ) + H2 (z )F2 (z )] (13)
3
107 The proposed design for a multiplier-less recombination non-uniform filter bank is illustrated with the help of two ex-
108 amples in Sections 4 and 5.
JID: INS
Table 1
12 bit CSD look-up-table entries.
Index CSD representation Decimal equivalent No. of SPT terms
900 0100-1000010-1 0.2195 4
109 3. Proposed methodology for designing multiplier-less recombination non-uniform filter banks
110 using meta-heuristic algorithms
111 Our aim is to design a totally multiplier-less and less complex recombination filter bank for hardware implementation.
112 Once the continuous coefficient filter bank is designed using the method discussed in Section 2, the filter coefficients are
113 converted to the CSD space. This is done to replace the multipliers in the filter implementation with shifters and adders so
114 that the design becomes multiplier-less. Now, to reduce the number of adders in the implementation, the filter coefficients
115 are rounded with a restricted number of SPT terms. In this paper, CSD conversion is done with the help of a look-up-table
116 (LUT) as explained in Section 3.1. The rounding of the continuous filter coefficients using a finite number of SPT terms may
117 cause degradation in the performance of the RNUFBs and it can be improved by formulating a suitable objective function
118 and using suitable optimization techniques.
119 3.1. Encoding of filter bank coefficients using CSD look-up-table
120 Hardware implementation of the infinite precision FIR filter requires the realization of multipliers, which are the main
121 power-consuming components. If the filter coefficients are represented in the SPT space, multipliers can be replaced by
122 shifters and adders. The number of non-zero bits in the filter coefficient representation decides the number of partial prod-
123 uct additions and hence the adders. CSD representation is a unique representation of the filter coefficients with a minimum
124 number of non-zero bits [21]. A fractional number q is represented in CSD format as [21].

W
q= ci 2R−i (14)
i=1
125 where ci = {−1, 1, 0}, W is the word length of the CSD number and the integer R represents a radix-point in the range
126 0 ≤ R < W. No adjacent bits in the CSD representation can be non-zero, i.e. ci ∗ ci−1 = 0, where ci is the ith bit in the CSD
127 representation.
128 Here, the encoding of the filter coefficients into the CSD space is done using a look-up-table (LUT) approach [74]. A CSD
129 look-up-table is created, consisting of four fields, namely index, CSD representation, decimal equivalent and number of SPT
130 terms. The entries of a typical 12-bit CSD look-up-table in which all 12 bits are used for the fractional part are shown in
131 Table 1. Here, the value of R is 0.
132 One method of converting the filter coefficients into the SPT space is to use ternary coding [8], in which the CSD repre-
133 sentation of the closest counterpart of each infinite precision multiplier coefficient in the LUT is taken and they are concate-
134 nated to form the initial seed. Since, in this method, CSD representation is used to generate the initial seed, the dimension
135 of the optimization problem and hence the time taken for optimization will be high. Also, during the various steps in the
136 optimization, the ternary-coded coefficients may become invalid CSD numbers. Hence, suitable restoration algorithms have
137 to be used to convert them back to the valid CSD format [19]. This increases the computational load of the algorithm.
138 In [44], the coefficients are encoded using the binary indices of a look-up-table and this encoding scheme is found to
139 preserve the canonical nature of the CSD number, but it is necessary for the binary-coded indices of the coefficients to
140 be decoded to access the look-up-table, which increases the computational load. Also, it is proved in [42] that, to attain
141 convergence, this method takes a large amount of running time.
142 Another encoding method is reported in [5,6,40–42], where the CSD filter coefficients are encoded as signed integers
143 by taking the indices corresponding to the magnitude of the CSD filter coefficients from the look-up-table and attaching to
144 them the signs. It is proved in [3] that if ‘D’ is the nearest CSD equivalent of a decimal number ‘d’, then the CSD equivalent
145 of ‘−d’ can be obtained as ‘−D’. Hence, it is appropriate to encode the filter coefficients using the signed indices of the
146 corresponding look-up-table locations. If the decimal filter coefficient is negative, then it is encoded as the negative of the
147 index of the location of its positive counterpart. The encoding and decoding of this method are simple, hence this encoding
148 scheme is used in this paper.
149 Once the continuous filter coefficients have been obtained, they are directly rounded to the nearest discrete value with
150 the specified number of SPT terms, as per the CSD look-up-table. The corresponding index after attaching the sign is taken.
151 The signed indices of all the coefficients are concatenated to get the initial solution for the optimization.
152 Here, the search space contains integers, hence classical gradient-based optimization algorithms cannot be used. Meta-
153 heuristic algorithms are good alternatives for these types of optimization problems because it is reported that these can
154 finally reach a global solution if the parameters are properly selected with respect to a particular design problem [70].
JID: INS
155 3.2. Optimization of the CSD-represented recombination non-uniform filter bank using meta-heuristic algorithms
156 If good optimization techniques are employed and a suitable objective function is formulated, only a limited number of
157 SPT terms are required for the CSD representation of the filter coefficients, in order to meet the given specifications. A filter
158 represented in this way will have a lower number of shifters and adders, thereby reducing the area and power consumption
159 further. In this paper, the aim is to design a multiplier-less RNUFB with a reduced number of adders meeting the given
160 specifications.
161 As discussed in Section 2, the design of the RNUFB includes the design of a set of uniform filter banks. In these uniform
162 filter banks, the synthesis filters are derived from the analysis filters as given in Eq. (7) such that aliasing is cancelled. Hence,
163 to optimize the multiplier-less RNUFB, we can optimize the analysis filters of the uniform filter bank and the synthesis filters
164 of the TMUX first and then these optimized filters can be used to derive the remaining filters and the multiplier-less RNUFB.
165 Here, the joint optimization of the filters is used, i.e. the coefficients of all the analysis filters of the M-channel UFB and
166 synthesis filters of the ml -channel TMUX are used to form the initial seed of the optimization problem. Since all the filters
167 are designed to have linear-phase, only half the number of filter coefficients needs to be considered and this reduces the
168 dimension of the optimization problem. The signed indices of the look-up-table entries of the filter coefficients are concate-
169 nated to form the initial seed. Since we use joint optimization of the filters, all of them will be adjusted simultaneously to
170 get a better filter bank response.
171 These optimized filters are used to design the totally multiplier-less RNUFB. A suitable objective function, known as
172 the cost function or fitness function, which takes care of the performance degradation due to the CSD rounding, is to be
173 formulated for the optimization and is done in the next section.
174 3.2.1. Formulation of the objective function

175 A suitable objective function is required to improve the performances of the CSD rounded filters and CSD rounded RNUFB.
176 Here, the amplitude distortion of the CSD-represented RNUFB and the pass-band ripple and stop-band attenuation of the
177 frequency response at the output of each channel are included as part of the objective function.
178 Let, Fpi and Fsi be the pass-band error and the stop-band error respectively of channel i, where i varies from 1 to n, where
179 n is the number of channels of the filter bank. They are defined as
Fpi = max |(|H (ωi )| − 1 )|, i = 1, 2, . . . , n (15)
ω pi <ωi <ωpi
180
Fsi = max |H (ωi )|, i = 1, 2, . . ., n (16)
ωi <ωsi or ωi >ωsi
181 where ωi is the frequency of the ith channel, ω pi and ωsi respectively are the lower pass-band and stop-band cut-off fre-
182 quencies and ωpi and ωs i respectively are the upper pass-band and stop-band cut-off frequencies of the frequency response
183 at the output of the ith channel. The pass-band error and the stop-band error of each channel are the difference between
184 the zero-phase frequency response, H(ωi ), of the channel to be optimized and the ideal channel response, which is 1 in the
185 pass-band and 0 in the stop-band.
186 Let the amplitude distortion function of the RNUFB in the frequency domain be represented as Tdist (ω). The peak error
187 in the amplitude distortion function of the filter bank, Fdist , can be written as
Fdist = max |(Tdist (ω ) − 1 )| (17)
0<ω<π
188 Now, to reduce the number of adders in the CSD-represented filter bank implementation, a constraint is also added to the
189 objective function using the penalty method [10]. The total number of adders required to implement a filter depends only
190 on the total number of SPT terms in the representation of the filter coefficients and not on how the SPT terms are allocated
191 to the coefficients. Hence, it is desirable to optimize the performance of the filter bank, subject to a fixed value of the total
192 number of SPT terms for the constituent filters in the RNUFB instead of a fixed number of SPT terms for each coefficient
193 [35]. This is done by fixing the average number of SPT terms in the CSD-represented RNUFB. This allocation has more
194 flexibility, because the different filter coefficients can contain a different number of SPT terms such that the performance is
195 maximized. If vH denotes the average number of non-zero SPT terms in the filter coefficients and VH is the required upper
196 bound of vH , the penalty function added to the optimization problem is given as
p(vH , VH ) = max(0, vH − VH ) (18)
197 Thus, the optimization problem for the design of the multiplier-less RNUFB is modeled as minimization of the objective
198 function OF, which is formulated by including Fpi , Fsi , Fdist and the penalty function, p(vH , VH ), as given below:

n
Minimize OF = [α pi Fpi + αsi Fsi ] + α1 Fdist + α2 p(vH , VH ) (19)
i=1
199 3.2.2. Optimization using meta-heuristic algorithms

200 The classical gradient-based optimization uses the function values and their derivatives and is not efficient in non-
201 differentiable or discontinuous problems. In this paper, the performances of the CSD-represented filters are optimized, hence
JID: INS
202 the search space consists of integers only. Hence, here, classical gradient-based optimization algorithms cannot be used [15].
203 The meta-heuristic algorithms are good alternatives for dealing with a discontinuous objective function. They are based on
204 randomization and local search and are found to be suitable for optimization problems where the search space contains
205 integers [70].
206 According to Yang [70], the meta-heuristic algorithms reach the optimal solution through two techniques: intensifica-
207 tion or exploitation and diversification or exploration. The algorithm should generate a diverse range of solutions including
208 potentially optimal solutions, and this is done by randomization. The algorithm should also be able to intensify the search
209 around the neighborhood of potentially good solutions by exploiting past solutions. Any successful meta-heuristic algorithm
210 requires a good balance between these two components. Typically, the solutions start with some randomly generated solu-
211 tions, and gradually reduce their diversification while increasing their intensification at the same time [70].
212 There are two types of meta-heuristic algorithms: trajectory-based and population-based. It is reported in [70] that
213 population-based algorithms might be more efficient for multi-objective, multi-modal optimization problems, since multiple
214 search actions are in parallel [32].
215 Multi-modal functions are functions with two or more local optima [28]. The multi-objective optimization deals with
216 optimizing problems with multiple and conflicting objectives. Unlike single objective optimization, these solutions are ob-
217 tained in such a way that the performance of each objective cannot be improved without sacrificing the performance of at
218 least one other objective [2]. Here, the multi-objective optimization problem is modeled as a single objective optimization
219 problem.
220 Many modern meta-heuristic algorithms are developed which mimic the intelligent characteristics of biological and phys-
221 ical agents. Genetic algorithms have been established as good alternatives for the optimization of multi-modal and multi-
222 dimensional problems. GA is a population-based evolutionary algorithm where, in each iteration, candidate solutions are
223 generated using genetic operations such as reproduction, crossover and mutation. Modern meta-heuristic algorithms such
224 as differential evolution (DE), artificial bee colony (ABC), harmony search (HS) and gravitational search (GS), which have
225 been proved to outperform GA, are used in this paper for solving the optimization problems in the discrete space. The
226 prominent features of the proposed approaches are the following: (i) they require reduced computation time and (ii) they
227 avoid the use of restoration algorithms. The mathematical proof, convergence analysis and discussions about the various
228 meta-heuristic algorithms are available in the literature [20,47,60,69,71].
229 Since the optimization problem posed in this paper is based on the design in a discrete space, the direct application of
230 these algorithms is not possible. The algorithms are tailor made so that, in every exploitation and exploration phase, the
231 candidate solutions turn out to be integers.
232 3.3. Overview of different meta-heuristic algorithms
233 The different steps of the various meta-heuristic algorithms such as DE, HSA, GSA and ABC are detailed below [6].
234 3.3.1. Modified differential evolution (DE) algorithm

235 The DE algorithm is a simple evolutionary algorithm and introduced by Storn and Price [59] and Dasgupta et al. [14]. A
236 modified integer-coded DE algorithm for the discrete space is proposed for the optimization of frequency response masking
237 (FRM) filter and reconfigurable channel filters in [42] and [6], respectively. The various steps of the modified DE algorithm
238 are explained below [6,42].
239 1. Initialization
240 The initial vector is formed by concatenating the signed indices of the CSD rounded coefficients of all the constituent
241 filters of the filter bank. A population of size N is generated by randomly perturbing the initial vector. To have a wider
242 search space, the initial number of vectors in the population is taken as an integer multiple of the chosen value of N. A
243 typical ith vector in the Gth generation is represented as
xi,G = [x1i,G x2i,G x3i,G . . . xDi,G ]
244 where D is the dimension of the solution vector.

245 2. Prioritized enlisting of the solution vectors using the fitness function.
246 The fitness of the candidate solutions is evaluated using the objective function given in Eq. (19) and only N number of
247 the best solutions are passed to the next stage.
248 3. Mutation
249 In this process, a new parameter vector called a mutant vector is generated for each target vector xi, G . This is done by
250 adding the scaled difference between two population vectors to a third vector. Three vectors, xr1 ,G , xr2 ,G and xr3 ,G are
251 randomly chosen and for each target vector, xi, G , the mutant vector, is generated as given below.
vi,G+1 = xr1 ,G + (xr2 ,G − xr3 ,G )F , i ∈ 1, 2, . . ., N (20)
252 where r1 , r2 , r3 are the different random indices lying between {1, 2, . . ., N} and are different from the running index
253 i. F is the scaling factor which controls the amplification of the difference to be added and varies in the range [0, 2].
JID: INS
254 . operation ensures that the values obtained from the mutation operation are integers. In addition, to ensure that the
255 values in the new vector are within the boundaries of the look-up-table, the following step is also included [6,42].

vlb , if v ji,G+1 < vlb
v ji,G+1 = (21)
vub , if v ji,G+1 > vub
256 where vlb and vub respectively are the lower and upper bounds of the look-up-table and j ∈ 1, 2, . . ., D
257 4. Crossover
258 The parameters of the vector obtained from the mutation step are mixed with those of the target vector to obtain the
259 trial vector, ui,G+1 = [u1i,G+1 u2i,G+1 u3i,G+1 ... uDi,G+1 ]. The mixing of the vectors is controlled using a parameter known
260 as a crossover ratio, CR, which is in the range [0, 1] and is done as given below

v ji,G+1 if randi ≤ CR or j = jrandi
u ji,G+1 = (22)
x ji,G otherwise
261 where jrandi is a random integer in the range [1, D] and randi is a random value in the range [0, 1].
262 5. Selection
263 The fitness function of the target vector and the trial vector are evaluated using the objective function given in Eq. (19)
264 and then their values are compared. If the trial vector has a better fitness, the target vector will be replaced by the trial
265 vector in the next generation.
266 6. Termination
267 Steps 3–5 are repeated until the specified number of iterations is reached. Once the algorithm is terminated, the best
268 solution is taken and is decoded to get the optimum filters.
269 3.3.2. Modified artificial bee colony (ABC) algorithm

270 Based on the foraging behavior of the honey bee swarm, Karaboga proposed an ABC algorithm for solving multi-
271 dimensional and multi-modal optimization problems [26–29]. An ABC algorithm for optimizing multi-objective problems
272 is discussed in [2]. A modified integer-coded ABC algorithm is proposed for the optimization of the FRM filter and recon-
273 figurable channel filters in the discrete space in [6] and [42], respectively. The various steps of the modified algorithm are
274 given below [6,42].
275 1. Initialize the food sources

276 The initial food source is formed by concatenating the signed indices of the CSD rounded coefficients of all the con-
277 stituent filters of the filter bank. Other different food sources are generated by randomly perturbing the initial food
278 source. To obtain a wider search space, the initial number of food sources is taken as the integer multiple of the number
279 of employed bees.
280 2. Sort the food sources based on fitness value
281 The amount of nectar in each food source which represents the fitness of the solution vectors is calculated. The solution
282 vectors are sorted based on their fitness values and only N vectors will be passed to the next step.
283 3. Employee bee phase
284 Each employed bee is associated with a food source, i.e. a solution vector. Employed bees search for new food sources
285 in the neighborhood of the current food sources, i.e. they try to find new solution vectors which are adjacent to the
286 existing vectors. A new food source adjacent to the ith food source is found by changing the randomly chosen jth value
287 in the ith food source as in the equation given below:
xt+1 (i, j ) = xt (i, j ) + φ (xt (i, j ) − xt (k, j )) (23)
288 where φ is a random value in the range [−1, 1], j is a random value in the range (1, 2, 3, . . ., D ) and k is a random
289 value in the range (1, 2, 3, . . ., N ). Here ,xt (i, j) is the jth parameter of the ith food source in the tth iteration. D is the
290 dimension of the solution vectors and N is the number of employed bees. To ensure that the newly generated value is
291 an integer, the above equation is modified as follows:
xt+1 (i, j ) = xt (i, j ) + φ (xt (i, j ) − xt (k, j )) (24)
292 Also, to ensure that the values in the new vector are within the boundaries of the look-up-table, the following step is
293 also included [6,42].

vlb , if xt+1 (i, j ) < vlb
xt+1 (i, j ) = (25)
vub , if xt+1 (i, j ) > vub
294 where vlb and vub respectively are the lower and upper bounds of the look-up-table. Now, a greedy selection mechanism
295 is used to select between the old and new solution vectors. The fitness value of the new vector representing the nectar
296 amount of the new food source is calculated and compared with that of the existing food source. If the fitness value of
297 the new vector is equal to or greater than the existing one, it is replaced with the new vector; otherwise, it is retained.
JID: INS
298 4. Onlooker bee phase

299 The employed bees share the information about their food source with the onlooker bees. The onlooker bees calculate
300 the nectar information collected from all the employed bees and, depending on the nectar amount, they select the food
301 source. An onlooker bee selects a food source depending on the probability distribution function associated with that
302 food source. Hence, a solution vector with a high fitness value will get more onlooker bees. Then, the onlooker bees
303 search for better food sources in the neighborhood of the current food source and these will be selected or rejected
304 depending on the fitness value. The search and selection/rejection of the new food source are done by the same method
305 as in the employee bee phase.
306 5. Scout bee phase
307 If the fitness value of a solution vector representing the food source is not improved after a fixed number of iterations,
308 known as limit cycles, that food source will be abandoned and the associated employee bee becomes the scout bee. The
309 scout bee randomly finds a new food source and occupies the position of the abandoned food source.
310 6. Termination
311 Steps 2–5 are repeated until the predetermined maximum number of iterations is reached. Once the algorithm termi-
312 nates, the solution vector with the best fitness is taken and is decoded to obtain the optimum filter in the CSD space.
313 3.3.3. Modified harmony search algorithm (HSA)

314 Inspired by the music improvisation scheme, Geem introduced the harmony search algorithm [34,45,56,75] for the op-
315 timization of mathematical problems. In this paper, the HSA algorithm is modified to be used for the discrete space in [6].
316 The various steps of the modified algorithm are explained below [6].
317 1. Initialization of harmony memory

318 The signed indices of the CSD rounded coefficients of all the constituent filters of the filter bank are concatenated to form
319 the initial harmony vector. Different solutions are generated by perturbing the initial harmony vector. To have a wider
320 search space, the initial number of harmony memory locations is taken as the integer multiple of the chosen harmony
321 memory size (HMS). A typical harmony vector of size D in the kth location of the harmony matrix can be represented
322 as:
xk = [xk,1 xk,2 xk,3 . . . xk,D ]

323 2. Prioritized enlisting of harmony memory locations using the fitness function
324 The fitness function is evaluated for each harmony vector, and the ‘HMS’ number of best solutions will be passed to the
325 next stage.
326 3. Harmony improvisation
327 Harmony improvisation is based on the following steps [34]:
328 • Memory consideration: values of the decision variable xnew i
in the new vector are chosen from any of the values stored
329 in the harmony memory in the range x1,i , x2,i , . . ., xHMS,i with probability equal to the harmony memory considering
330 rate (HMCR).
331 • Pitch adjustment: each new decision variable obtained in the memory consideration step is examined to decide
332 whether it should be pitch adjusted. This is decided by the parameter, pitch adjusting rate (PAR). If pitch adjustment
333 is necessary, it is done as shown below:
xnew
i = xi + rand (1, −1 )FW (i ) (26)
334 where FW(i) is an arbitrary distance bandwidth for the ith design variable. . operation ensures that the new values
335 obtained from the pitch adjustment step are integers.
336 • Random selection: random decision variables are generated for the new vector with a probability equal to (1 −
337 HMCR ).
338 4. Update the harmony memory
339 Now the fitness function is evaluated and if the fitness of the newly generated harmony vector is better than the worst
340 harmony in the harmony memory (HM), the worst harmony is replaced by the new harmony.
341 5. Check the stopping criterion
342 The HSA is terminated when the specified number of iterations is reached. Otherwise, steps 3 and 4 are repeated. After
343 the algorithm terminates, fitness is evaluated for all the harmony vectors which are retained and the best solution is
344 taken and encoded to get the optimized filter coefficients.
345 3.3.4. Modified gravitational search algorithm (GSA)

346 GSA is a population-based algorithm based on the law of gravity and mass interactions [49–51,72]. GSA can be considered
347 as an artificial world of masses. The position of the masses constitutes the solution space and the performance of each
348 solution is measured in terms of its masses using fitness function. The GSA algorithm is modified to be used for optimization
349 in the discrete space in [6]. The various steps of the modified algorithm are explained below [6].
350 1. Initialization of agents
JID: INS
351 Let there be a system of N masses (agents), with the position of the ith agent defined as
xi = (x1i , . . . , xdi , . . ., xni ) for i = 1, 2, 3, . . ., N
352 where xdi represents the position of the ith agent in the dth dimension. Different agents are generated by randomly
353 perturbing the initial solution vector. To obtain a wider search space, the initial number of agents is taken as the integer
354 multiple of N.
355 2. Sort the agents based on fitness value
356 The fitness value of agents is calculated using the objective function and the agents are sorted based on the fitness value.
357 Only N number of agents with the best fitness will be passed to the next step.
358 3. Evaluate fitness and update gravitational constant, G
359 In each iteration, the best fitness, best(t), and worst fitness, worst (t ), of the agents are updated using the equation given
360 below:
worst (t ) = max j∈1,...,N f it j (t ) (27)
361
best (t ) = min j∈1,...,N f it j (t ) (28)
362 where fiti (t) represents the fitness value of the agent i at time t. The gravitational constant depends on the age of the
363 universe and is reduced with time t. The gravitational constant G at each iteration is calculated by means of the following
364 equation:
G(t ) = G0 e−αt/T (29)
365 G0 is set to 100, α is taken as 20 and T is the total number of iterations.
366 4. Calculate the mass of each agent
367 The gravitational and inertial masses of an agent are calculated using the fitness function and are updated by means of
368 the following equations:
Mai = M pi = Mii = Mi , i = 1, 2, 3,. . ., N (30)
369
f iti (t ) − worst (t ) mi (t )
mi (t ) = Mi (t ) = N (31)
best (t ) − worst (t ) j=1 m j (t )
370 where Mai , Mpi , and Mii are the active gravitational mass, passive gravitational mass and inertia mass respectively of the
371 ith agent.
372 5. Calculate acceleration of each agent
373 adi (t ), the acceleration of the agent i at time t, and in dimension d, is given by:
Fid (t )
adi (t ) = (32)
Mii (t )
374 The total force that acts on agent i in a dimension d is equal to the randomly weighted sum of the dth components of
375 the forces applied from other agents, which is given by

N
Fid (t ) = rand j Fidj (t ) (33)
j=1, j
=i
376 where randj is a random number in the interval [0,1]. At a particular time t, the force acting on mass i from mass j can
377 be defined as
M pi (t )Ma j (t ) d
Fidj (t ) = G(t ) (x j (t ) − xdi (t )) (34)
Ri j (t ) + ε
378 where Rij (t) is the distance between two agents i and j and ε is a small constant.
379 6. Update the velocity and position of each agent
380 The next position and velocity of an agent can be calculated as follows:
vdi (t + 1 ) = randi vdi (t ) + adi (t ) (35)
381
xdi (t + 1 ) = xdi (t ) + vdi (t + 1 ) (36)
382 where randi is a uniform random variable in the interval [0, 1]. . denotes rounding to the lower value. This operation
383 ensures that the new candidate solution is an integer.
384 7. Termination
385 Steps 3–6 are repeated until the maximum number of iterations is reached. The agent with the best fitness is taken and
386 is decoded to obtain the optimum filters in the CSD space.
JID: INS
3 2 2 3
C0 C0-1
H0 G0,0 G’0,0 F0
Channel 1, Y1(z)
3 2 2 C1-1 3
C1
X(z) H1 G0,1 G’0,1 F1
Y(z)
3 C2 -1 3
C2
H2 Z-k F2
Channel 2, Y2(z)
(a) Block diagram of recombination NUFB with rational sampling factors {2/3, 1/3}
2 3 3 2
Channel 1, Y1(z)
H0 F0
X(z) Y(z)
3 3
H2 Z-k F2
Channel 2, Y2(z)
(b) Equivalent structure of recombination NUFB with rational sampling factors {2/3, 1/3}
Fig. 3. RNUFB with sampling factors {2/3, 1/3}.
387 4. Design example 1: two-channel recombination non-uniform filter bank with rational sampling
388 factors {2/3, 1/3}
389 All the simulations are done using MATLAB 7.10.0.499 on an Intel®CoreTM i5-2400 processor operating at 3.10 GHz.
390 The general structure of a recombination filter bank is shown in Fig. 1a. A two-channel recombination NUFB with rational
391 sampling factors {2/3, 1/3} is designed using the method described in Section 2. All the filters are chosen to have linear-
392 phase property and are designed using the Parks–McClellan algorithm [55]. The filter bank is made multiplier-less using
393 meta-heuristic algorithms. The performances of the multiplier-less RNUFBs designed using various meta-heuristic algorithms
394 are analyzed. The complexity of implementation of the continuous coefficient filter bank and multiplier-less filter bank using
395 various algorithms is also compared.
396 4.1. Design for a continuous coefficient recombination non-uniform filter bank
397 The block diagram of a two-channel recombination NUFB with rational sampling factors {2/3, 1/3} and its equivalent
398 structure are shown in Fig. 3. Here, H0 (z), H1 (z) and H2 (z) are the analysis filters and F0 (z), F1 (z) and F2 (z) are the synthesis
399 filters of the three-channel UFB. G0, 0 (z) and G0, 1 (z) are the synthesis filters and G0,0 (z ) and G0,1 (z ) are the analysis filters
400 of the two-channel TMUX. The parameters c0 , c1 and c2 as per Fig. 1a are selected as 1, −1, and 1, respectively. z−k indicates
401 the delay which is inserted in channel 2 to compensate for the delay introduced by the TMUX in channel 1.
402 The pass-band and stop-band frequencies of the analysis filters, H0 (z), H1 (z) and H2 (z) of the three-channel UFB and the
403 synthesis filters, G0, 0 (z) and G0, 1 (z) of the two-channel UFB are taken as:
404 H0 (z) : ω p0 = 0.232π , ωs0 = 0.432π
405 H1 (z) : ωs1 = 0.232π , ω p1 = 0.432π , ωp1 = 0.566 π , ωs 1 = 0.766π
406 H2 (z) : ωs2 = 0.566π , ωp2 = 0.766π
407 G0, 0 (z) : ω p0,0 = 0.4π , ωs0,0 = 0.6π
408 G0, 1 (z) : ωs0,1 = 0.4π , ω p0,1 = 0.6π
409 Maximum pass-band ripple: 0.001 dB
410 Minimum stop-band attenuation: 40 dB
411 The magnitude responses of the analysis filters of the three-channel UFB and the synthesis filters of the two-channel UFB
412 are shown in Fig. 4.
413 The lengths of the analysis filters of the three-channel UFB and synthesis filters of the two-channel UFB are chosen as 39
414 and 26, respectively. The other filters are derived from these filters using the method explained in Section 2. The 1-skewed
415 FB and hence the two-channel TMUX are derived from the two-channel UFB using Eq. (1). Combining them as shown in
416 Fig. 3 results in a two-channel RNUFB with sampling factors {2/3, 1/3}. The magnitude responses at the outputs of channel
417 1 and channel 2 are shown in Fig. 5.
418 The total number of coefficients of the analysis filters of the three-channel UFB and synthesis filters of the two-channel
419 UFB is 39 + 39 + 39 + 26 + 26 = 169. When exploiting the symmetry property of the linear-phase FIR filters, only half
420 the number of filter coefficients needs to be considered. The total number of multipliers required to design the continuous
421 coefficient two-channel RNUFB with rational sampling factors {2/3, 1/3} is found to be 86. The initial vector of all the
JID: INS
20 10
G G
H H H 0,0 0,1
1 0
0 2
0
−10
−20 −20
Magnitude response in dB
−30
−40
−40
−60
−50
−80 −60
−70
−100
−80
−120 −90
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency (ω/pi) Normalized frequency (ω/pi)
(a) Magnitude responses of the (b) Magnitude responses of the
analysis filters of three-channel UFB synthesis filters of two-channel UFB
Fig. 4. Magnitude responses of the filters of the continuous coefficient three- and two- channel UFB.
20
Channel 1 Channel 2
0
−20
−40
−60
−80
−100
−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency (ω/pi)
Fig. 5. Magnitude responses at the outputs of channel 1 and channel 2 of the continuous coefficient two-channel RNUFB with sampling factors {2/3, 1/3}.
Table 2
Amplitude distortion and complexity of implementation of the continuous coefficient recombination filter bank with
rational sampling factors {2/3, 1/3}.
Amplitude distortion (dB) Number of multipliers Number of adders
0.05402 86 81
422 optimization techniques is formed by concatenating the signed indices of the CSD rounded analysis filter coefficients of the
423 filter bank. Hence, the number of decision variables is 86. The amplitude distortion of the continuous coefficient RNUFB is
424 given in Table 2. The magnitude responses of the equivalent analysis filters, Ĥ0 (z ) of channel 1 and H2 (z) of channel 2, of
425 the continuous coefficient RNUFB referred to in Fig. 3 b are shown in Fig. 6a. The magnitude responses of the equivalent
426 synthesis filters, F̂0 (z ), of channel 1 and F2 (z) of channel 2 of the continuous coefficient RNUFB referred to in Fig. 3 b are
427 shown in Fig. 6b. The total amplitude distortion, |Y(z)|, of the RNUFB is shown in Fig. 7.
428 4.2. Design of the multiplier-less recombination non-uniform filter bank
429 To reduce the complexity in the hardware implementation, the non-uniform recombination filter bank is made multiplier-
430 less by converting the continuous filter coefficients into the CSD space. As discussed in Section 2, the design of the RNUFB
431 includes the design of a set of uniform filter banks. Also, in these uniform filter banks, the synthesis filters are derived from
432 the analysis filters as given in Eq. (7) such that aliasing is cancelled. Thus, the analysis filters of the uniform filter bank and
JID: INS
20 20
F F
H H 0 2
0 2 0
0
−20
−20
−40
−40
−60
−60
−80
−80
−100
−100 −120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) The frequency responses of the (b) The frequency responses of the
equivalent analysis filters of the equivalent synthesis filters of the
continuous coefficient RNUFB continuous coefficient RNUFB
Fig. 6. RNUFB with sampling factors {2/3, 1/3}.
Amplitude
Distortion
0.04
0.02
−0.02
−0.04
−0.06
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 7. Amplitude distortion of the continuous coefficient RNUFB with sampling factors {2/3, 1/3}.
Table 3
Amplitude distortion and complexity of implementation of the maximum-precision CSD-represented recombination fil-
ter bank with rational sampling factors {2/3, 1/3} when different numbers of bits are used for CSD representation.
Word-length Amplitude distortion Number of adders

used (dB) due to SPT terms
Continuous coefficients 0.05402 0

10 bits 0.108 92
12 bits 0.07126 142
14 bits 0.05838 202
433 the synthesis filters of the TMUX are made multiplier-less first. Then these filters are used to derive the remaining filters to
434 design the multiplier-less RNUFB.
435 4.2.1. Design of the maximum-precision recombination filter bank

436 To decide the word-length to be used, the performances of the maximum-precision RNUFB with different word-lengths
437 are analyzed. Table 3 gives the frequency performance parameters and complexity of the maximum-precision recombination
438 filter bank when different numbers of bits are used for the CSD representation. The amplitude distortion obtained for the
439 maximum-precision RNUFB with a 12-bit and 14-bit CSD is closer to that of the continuous coefficient RNUFB, compared
440 to that of the maximum-precision RNUFB with a 10-bit CSD. However, the number of adders due to the SPT terms for the
441 maximum-precision RNUFB with a 14-bit CSD is much higher compared to that of the maximum-precision RNUFB with a
442 12-bit CSD. Hence, there may be some trade-off between the filter response and complexity.
443 In this example, a 12-bit CSD representation is used instead of a 10-bit or 14-bit CSD in order to reduce the complexity
444 without much degradation of the filter performance.
JID: INS
Table 4
Amplitude distortion and complexity of implementation of the CSD rounded RNUFB with sampling factors {2/3, 1/3}.
Word-length Amplitude Number of Number of

used distortion (dB) structural adders due Total number
adders to SPT terms of adders
Continuous 0.05402 81 0 81
coefficients
Maximum
precision 0.07126 81 142 223
(6 SPTs)
CSD rounded 0.1541 81 118 199
(3 SPTs)
Table 5
Various parameters in the objective function of the two-channel RNUFB with sampling factors {2/3, 1/3}.
α p1 αs1 α p2 αs2 α1 α2 VH
1 5 1 3 5 0.5 2.5
445 4.2.2. Design of the CSD rounded recombination filter bank

446 The number of adders required for the filter implementation depends on the number of SPT terms. The greater the num-
447 ber of SPT terms used to represent the filter coefficients, the closer the rounded filter coefficients are to the unrounded
448 coefficients and the closer the rounded performance is to the unrounded performance. However, a greater number of non-
449 zero bits means greater hardware complexity and a higher cost. Thus, there is a trade-off between hardware cost and per-
450 formance characteristics. To reduce the number of adders, the filter coefficients are rounded to the discrete values in the
451 CSD space with a restricted number of SPT terms. In Table 4, the number of adders required for the implementation of
452 the CSD rounded RNUFB with 3 SPT terms is compared with that of the maximum-precision RNUFB. It can be seen that
453 there is a reduction of 10.76% in the total number of adders required for the implementation of the CSD rounded RNUFB
454 when compared to that of the maximum-precision RNUFB. The amplitude distortion of the filter bank with continuous filter
455 coefficients, with maximum-precision filter coefficients and with the filter coefficients rounded with 3 SPT terms is also
456 compared in Table 4. It can be seen that the complexity of implementation of the CSD rounded filter bank is reduced, but
457 the performance becomes degraded. Hence, suitable meta-heuristic algorithms are deployed to improve the performance.
458 4.3. Proposed design of the optimal totally multiplier-less recombination non-uniform filter bank using meta-heuristic algorithms
459 To optimize the multiplier-less RNUFB, the analysis filters of the uniform filter bank and the synthesis filters of the TMUX
460 are optimized first and then these optimal filters can be used to derive the remaining filters and the optimal multiplier-less
461 RNUFB.
462 Here, the coefficients of all the analysis filters of the M-channel UFB and synthesis filters of the ml -channel TMUX are
463 used to form the initial seed of the optimization problem. Since all the filters are designed to have linear-phase, only half
464 the number of filter coefficients needs to be considered. Also, since the joint optimization of the filters is used, they will
465 all be simultaneously adjusted to get a better filter bank response. These optimized filters are used to design the totally
466 multiplier-less RNUFB.
467 First, the continuous filter coefficients are encoded into the CSD space using the encoding scheme described in
468 Section 3.1. To achieve this, a 12-bit look-up-table (LUT) of size 2048, in which all 12 bits are used for the fractional part,
469 is created. The next step is to formulate a suitable objective function which takes care of the performance degradation of
470 the RNUFB due to the CSD rounding. The objective function given in Eq. (19) is made use of here, where n is 2 for the
471 two-channel filter bank and the required average number of non-zero bits in a filter coefficient, VH , is taken to be 2.5 in this
472 case, since this is found to give the minimum complexity RNUFB without much degradation in the performance. The values
473 assigned to the different parameters of the objective function given in Eq. (19) when it is used for the optimization of the
474 two-channel RNUFB considered here, are given in Table 5.
475 Now, the meta-heuristic algorithms are applied to improve the performance of the RNUFB with the help of the objective
476 function given in Eq. (19).
477 The various parameters used for the integer-coded GA, DE, ABC, HSA and GSA algorithms are given in Tables 6–10, re-
478 spectively. The average amplitude distortions of the optimal multiplier-less RNUFB with sampling factors {2/3, 1/3} for 10
479 simulations are listed and compared in Table 11.
480 The comparison of the average design time taken by the various algorithms is given in Table 12. The evolutionary pro-
481 cesses of various algorithms for the design of the optimal multiplier-less RNUFB with sampling factors {2/3, 1/3} is shown
482 in Fig. 8. The comparison of the complexity of implementation of the recombination filter bank in terms of the number of
483 multipliers and number of adders is given in Table 13. The number of multipliers in RNUFB proposed in [36] is 86, whereas
484 it is zero for the RNUFB proposed in this paper.
JID: INS
Table 6
Parameters of integer-coded GA.
Population Mutation Fraction of Number of elite Number of

size rate population kept chromosomes generations
50 0.005 0.3 10 500
Table 7
Parameters of integer-coded DE algorithm.
Population size, N Scaling factor, F Crossover ratio, CR Number of generations
50 0.7 0.01 500
Table 8
Parameters of integer-coded ABC.
Number of food sources Limit Number of generations
50 50 500
Table 9
Parameters of integer-coded HSA.
Harmony memory size (HMS) Harmony memory considering rate (HMCR) Pitch adjusting rate (PAR) Number of iterations
50 0.95 0.01 500
Table 10
Parameters of integer-coded GSA.
Number of agents Gravitational constant, G0 α Number of iterations
50 100 20 500
Table 11
Amplitude distortion of the RNUFB with sampling factors {2/3, 1/3} designed using various meta-heuristic algorithms (average of 10 simu-
lations).
Algorithm Amplitude distortion (dB) Standard deviation of amplitude distortion
Continuous coefficients 0.05402 –

Maximum-precision (6 SPTs) 0.07126 –
CSD Rounded (3 SPTs) 0.1541 –
Integer-coded GA 0.09127 1.017 × 10−2
Integer-coded DE 0.08035 5.284 × 10−3
Integer-coded ABC 0.06512 2.2548 × 10−3
Integer-coded HSA 0.08398 9.7452 × 10−3
Integer-coded GSA 0.05915 1.9209 × 10−3
Table 12
Design time used by various optimization algorithms for the design of the optimal multiplier-less RNUFB with sampling factors {2/3, 1/3}
(average of 10 simulations).
Algorithm Average of final objective function values Number of iterations Design time (s)
Integer-coded GA 0.1410 500 690.11

Integer-coded DE 0.1277 500 548.21
Integer-coded ABC 0.1098 500 1030.83
Integer-coded HSA 0.1304 500 42.74
Integer-coded GSA 0.1043 500 614.442
485 It can be seen that DE, ABC, HSA, and GSA respectively result in a multiplier-less RNUFB with 3.59%, 5.38%, 5.38% and
486 6.28% savings in the number of adders in the implementation compared to that of the maximum-precision RNUFB. The
487 amplitude distortions of the ABC and GSA optimized RNUFB are less than that of the maximum-precision RNUFB. It can be
488 inferred that HSA is the fastest algorithm, but the amplitude distortion of the HSA optimized RNUFB is more compared to
489 that of the ABC and GSA optimized RNUFB. GSA results in the least complex RNUFB with the least amplitude distortion.
490 However, the design time for the GSA is approximately fifteen times that for the HSA. When HSA is run for 1000 seconds, it
491 is found that the value of the objective function is reduced to 0.1031. Hence, if the design time is of primary concern, HSA
JID: INS
0.2
GA
0.19 DE
ABC
HSA
0.18 GSA
Minimum cost
0.17
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0 50 100 150 200 250 300 350 400 450 500
Number of iterations
Fig. 8. The evolution processes of various algorithms for the design of the optimal multiplier-less RNUFB with sampling factors {2/3, 1/3}.
Table 13
Complexity of implementation of the RNUFB with sampling factors {2/3, 1/3} designed using various meta-heuristic algorithms
(average of 10 simulations).
Optimization techniques Number of Adders due to Number of Total number Number of

SPT terms SPT terms structural adders of adders multipliers
Continuous – – 81 81 86
coefficients [36]
Maximum-
precision 227 142 81 223 0
(6 SPTs)
CSD 203 118 81 199 0
rounded
(3 SPTs)
Integer- 213 128 81 209 0
coded GA
Integer- 213 134 81 215 0
coded DE
Integer- 215 130 81 211 0
coded ABC
Integer- 215 130 81 211 0
coded HSA
Integer- 213 128 81 209 0
coded GSA
492 can be chosen. For applications in which the complexity and amplitude distortion are more important, GSA can be selected
493 for the design of the multiplier-less RNUFB.
494 5. Design example 2: four-channel recombination non-uniform filter bank with rational sampling
495 factors {2/5, 1/5, 1/5, 1/5}
496 A four-channel recombination NUFB with rational sampling factors {2/5, 1/5, 1/5, 1/5} is also designed using the method
497 described in Section 2. The parameters c0 , c1 , c2 , c3 and c4 are selected as 1, −1, 1, 1 and 1, respectively. Here too, all the
498 filters are designed using the Remez exchange algorithm.
499 5.1. Design of the continuous coefficient non-uniform recombination filter bank
500 A four-channel linear-phase RNUFB with rational sampling factors {2/5, 1/5, 1/5, 1/5} is chosen as shown in Fig. 9. z−k1 ,
501 z−k2 ,
and z−k3 indicate the different delays which are inserted in channels 2, 3 and 4 respectively to compensate for the
502 delay introduced by the TMUX in channel 1. The pass-band and stop-band frequencies of the analysis filters of the five-
503 channel UFB and the synthesis filters of the two-channel UFB are taken as:
504 H0 (z) : ω p0 = 0.16π , ωs0 = 0.24π
505 H1 (z) : ωs1 = 0.16π , ω p1 = 0.24π , ωp1 = 0.36 π , ωs 1 = 0.44π
506 H2 (z) : ωs2 = 0.36π , ω p2 = 0.44π , ωp2 = 0.56 π , ωs 2 = 0.64π
507 H3 (z) : ωs3 = 0.56π , ω p3 = 0.64π , ωp3 = 0.76 π , ωs 3 = 0.84π
JID: INS
5 2 2 5
c0 c0-1
H0 G0,0 G’0,0 F0
Channel 1, Y1(z)
5 2 2 5
c1 c1-1
H1 G0,1 G’0,1 F1
5 c2 -1 5
c2
Channel 2, Y2(z)
X(z) H2 Z-k1 F2 Y(z)
5 c3 -1 5
c3
Channel 3, Y3(z)
H3 Z-k2 F3
5 c4-1 5
c4
Channel 4, Y4(z)
H4 Z-k3 F4
(a) Block diagram of four-channel recombination NUFB

2 5 5 2
Channel 1, Y1(z)
H0 F0
5 5
Channel 2, Y2(z)
H2 Z-k1 F2
X(z) 5 5 Y(z)
Channel 3, Y3(z)
H3 Z-k2 F3
5 5
Channel 4, Y4(z)
H4 Z-k3 F4
(b) Equivalent structure of four-channel recombination NUFB

Fig. 9. Recombination NUFB with rational sampling factors {2/5, 1/5, 1/5, 1/5}.
20 10
G0,0 G0,1
H0 H1 H2 H3 H4 0
0
−10
−20
−20
−30
−40
−40
−60
−50
−80 −60
−70
−100
−80
−120 −90
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) Magnitude responses of the (b) Magnitude responses of the synthesis

analysis filters of the five-channel UFB filters of the two-channel UFB
Fig. 10. Magnitude responses of the filters of the continuous coefficient five- and two- channel UFBs.
508 H4 (z) : ωs4 = 0.76π , ω p4 = 0.84π

509 G0, 0 (z) : ω p0,0 = 0.42π , ωs0,0 = 0.58π
510 G0, 1 (z) : ωs0,1 = 0.42π , ω p0,1 = 0.58π
511 Maximum pass-band ripple: 0.001dB
512 Minimum stop-band attenuation: 40 dB
513 The magnitude responses of the analysis filters of the three-channel UFB and the synthesis filters of the two-channel UFB
514 are shown in Fig. 10. The lengths of the analysis filters of the five-channel UFB and the synthesis filters of the two-channel
515 UFBs are chosen as 85 and 34, respectively. Combining the five-channel UFB and two-channel TMUX as shown in Fig. 9
516 results in the four-channel RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5}. The total number of multipliers required to
JID: INS
20
Channel 1 Channel 2 Channel 3 Channel 4

0
−20
−40
−60
−80
−100
−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 11. Magnitude responses at the outputs of channel 1 and channel 2 of the continuous coefficient two-channel RNUFB with sampling factors {2/5, 1/5,
1/5, 1/5}.
20
H0 H2 H3 H4
0
−20
−40
−60
−80
Channel 1
−100 Channel 2
Channel 3
Channel 4
−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 12. The frequency responses of the equivalent analysis filters of the continuous coefficient RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5}.
0.05
Amplitude Distortion
0.04
0.03
0.02
0.01
−0.01
−0.02
−0.03
−0.04
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 13. Amplitude distortion of the continuous coefficient RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5}.
517 implement this continuous coefficient four-channel NUFB is found to be 249. The magnitude responses at the outputs of
518 channel 1, channel 2, channel 3 and channel 4 are shown in Fig. 11. The magnitude responses of the equivalent analysis
519 filters Ĥ0 (z ), H2 (z), H3 (z) and H4 (z) of the continuous coefficient RNUFB referred to in Fig. 9 are shown in Fig. 12. The total
520 amplitude distortion, |Y(z)|, of the continuous coefficient RNUFB is shown in Fig. 13.
JID: INS
Table 14
Amplitude distortion and complexity of implementation of the maximum-precision CSD-represented RNUFB with ratio-
nal sampling factors {2/5, 1/5, 1/5, 1/5} when different numbers of bits are used for CSD representation.
Word-length used Amplitude distortion (dB) Number of adders due to SPT terms
Continuous coefficients 0.04169 0

10 bits 0.1446 203
12 bits 0.0633 331
14 bits 0.04191 481
Table 15
Amplitude distortion and complexity of implementation of the RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5}.
Word-length Amplitude Number Number of adders Total number

used distortion (dB) of adders due to SPT terms of adders
Continuous 0.04169 242 0 242

coefficients
Maximum-
precision 0.0633 242 331 573
(6 SPTs)
CSD 0.1826 242 279 521
rounded
(3 SPTs)
Table 16
Various parameters in the objective function of the four-channel RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5}.
α p1 αs1 α p2 αs2 α p3 αs3 α p4 αs4 α1 α2 VH
1 1 1 1 1 1 1 1 10 1 2
521 5.2. Design of the totally multiplier-less recombination non-uniform filter bank using meta-heuristic algorithms
522 Table 14 gives the amplitude distortion and the complexity of implementation of the maximum-precision recombination
523 filter bank with rational sampling factors {2/5, 1/5, 1/5, 1/5} when different word-lengths are used for CSD representation
524 and is compared with that of the continuous coefficient RNUFB. Here, the amplitude distortion obtained for the maximum-
525 precision RNUFB with 12-bit and 14-bit CSD representation are closer to those of the continuous coefficient RNUFB, com-
526 pared to those of the maximum-precision RNUFB with 10-bit CSD representation. However, the number of adders due to
527 the SPT terms for the maximum-precision RNUFB with 14-bit CSD representation is much higher compared to that of the
528 maximum-precision RNUFB with 12-bit CSD representation.
529 Hence, in this example too, 12-bit CSD representation is used instead of 10 or 14 bits, in order to reduce the complexity
530 without much degradation of the filter performances.
531 The amplitude distortion of the filter bank with continuous coefficients and with the coefficients rounded with 3 non-
532 zero bits in the CSD space are given in Table 15. Meta-heuristic algorithms such as modified integer-coded GA, ABC, DE, HSA
533 and GSA are deployed to optimize the performances of the CSD rounded RNUFB.
534 The objective function used for the optimization of the four-channel RNUFB is the same as that given in Eq. (19), where
535 n is 4 for the four-channel filter bank. The values assigned to the different parameters for the optimization of the RNUFB
536 are given in Table 16.
537 Here, the required average number of non-zero bits in a filter coefficient is taken to be 2, since this gives the minimum
538 complexity without affecting the performance too much.
539 The various parameters used for the meta-heuristic algorithms DE, ABC, HSA and GSA for the optimization of the four-
540 channel RNUFB are given in Tables 7–10, respectively. The amplitude distortions of the RNUFB with sampling factors {2/5,
541 1/5, 1/5, 1/5} when it is made multiplier-less using various meta-heuristic algorithms are listed in Table 17.
542 The evolutionary processes of various algorithms when they are used for the design of the optimal multiplier-less RNUFB
543 with sampling factors {2/5, 1/5, 1/5, 1/5} are shown in Fig. 14. The comparison of the design time of the different algorithms
544 is given in Table 18. The complexity of implementation of the recombination filter bank with sampling factors {2/5, 1/5, 1/5,
545 1/5} designed using various meta-heuristic algorithms is given in Table 19.
546 As in the case of the two-channel RNUFB discussed in Section 4, here too HSA is found to be the fastest algorithm.
547 However, the amplitude distortion of the HSA optimized RNUFB is more than that of the GSA optimized RNUFB and results
548 in a more complex RNUFB. GSA results in the RNUFB with the least amplitude distortion. However, the least complex RNUFB
549 is designed with the ABC algorithm, but this algorithm is observed to be very slow. Hence, when the design specifications
550 require the least complex filter bank and if there is no speed constraint, the ABC algorithm can be chosen. If amplitude
551 distortion is of primary importance, GSA can be chosen, which also results in an RNUFB with a slightly higher complexity
JID: INS
Table 17
Amplitude distortion of the optimized RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5} designed
using various meta-heuristic algorithms (average of 10 simulations).
Algorithm Amplitude distortion (dB)
Continuous coefficients 0.04169

Maximum-precision (6 SPTs) 0.0633
CSD rounded (3 SPTs) 0.1826
Integer-coded DE 0.09990
Integer-coded ABC 0.07359
Integer-coded HSA 0.06862
Integer-coded GSA 0.04441
0.4
0.35
Minimum cost
0.3
0.25
DE
0.2
ABC
HSA
GSA
0 50 100 150 200 250 300 350 400 450 500
Number of iterations
Fig. 14. The evolution processes of various algorithms for the design of the optimal multiplier-less RNUFB with sampling factors {2/5, 1/5, 1/5, 1/5}.
Table 18
Design time used by various optimization algorithms for the design of the optimal multiplier-less RNUFB with sampling
factors {2/5, 1/5, 1/5, 1/5} (average of 10 simulations).
Algorithm Average of final objective function values Number of iterations Design time (s)
Integer-coded DE 0.2273 500 918.05

Integer-coded ABC 0.2428 500 1865.74
Integer-coded HSA 0.3385 500 58.35
Integer-coded GSA 0.1960 500 1219.34
Table 19
Comparison of complexity of implementation of RNUFBs with sampling factors {2/5, 1/5, 1/5, 1/5} designed using various meta-heuristic algo-
rithms (average of 10 simulations).
Optimization Number of SPT Adders due to SPT Number of adders Total of number of Number of
techniques terms terms adders multipliers
Continuous – – 242 242 249

coefficients [36]
Maximum- 564 331 242 573 0
precision
(6 SPTs)
CSD 512 279 242 521 0
rounded
(3 SPTs)
Integer- 499 279 242 521 0
coded DE
Integer- 517 284 242 526 0
coded ABC
Integer- 546 313 242 555 0
coded HSA
Integer- 520 287 242 529 0
coded GSA
JID: INS
552 than that using the ABC algorithm. When speed is of primary importance and if the complexity is not of concern, HSA can
553 be selected.
554 6. Conclusion
555 The recombination non-uniform filter bank approach is a method for implementing near perfect reconstruction NUFBs
556 with rational sampling factors. By placing the TMUX in the appropriate positions, the RNUFB which extracts the required
557 non-uniform frequency bands from the input can be designed. In this paper, totally multiplier-less optimal recombination
558 non-uniform filter banks are proposed. This is achieved by converting the constituent filter coefficients into the CSD space.
559 This approach reduces the complexity by replacing the multipliers with adders and shifters. Approximation of the continu-
560 ous filter coefficients into the discrete space degrades the performances of the filter banks, which are improved by various
561 meta-heuristic algorithms. The performances of the DE, ABC, HSA and GSA optimized filter banks are compared in terms
562 of amplitude distortion of the filter bank, design time and complexity of implementation. A near perfect reconstruction to-
563 tally multiplier-less two-channel RNUFB with sampling factors {2/3, 1/3} and another four-channel RNUFB with sampling
564 factors {2/5, 1/5, 1/5, 1/5} are designed and the complexity is compared with that of the continuous coefficients. It is found
565 that GSA gives good frequency response characteristics for RNUFBs for both examples used. However, in each case, HSA is
566 found to be the fastest algorithm. Thus, the method proposed in this paper results in non-uniform filter banks with rational
567 sampling factors which are multiplier-less, have linear-phase and near perfect reconstruction.
568 References
569 [1] S. Agrawal, O. Sahu, Artificial bee colony algorithm to design two-channel quadrature mirror filter banks, Swarm Evol. Comput. 21 (2015) 24–31.
570 [2] R. Akbari, R. Hedayatzadeh, K. Ziarati, B. Hassanizadeh, A multi-objective artificial bee colony algorithm, Swarm Evol. Comput. 2 (2012) 39–52.
571 [3] F. Ashrafzadeh, B. Nowrouzian, A.T. Fuller, A novel modified branch-and-bound technique for discrete optimization over canonical signed-digit number
572 space, in: Proceedings of the 1998 IEEE International Symposium on Circuits and Systems, ISCAS’98., vol. 5, IEEE, 1998, pp. 391–394.
Q5 573 [4] T. Bindiya, E. Elias, Design of multiplier-less reconfigurable non-uniform channel filters using meta-heuristic algorithms, Int. J. Comput. Appl. 59 (11)
574 (2012) 1–11.Foundation of Computer Science, New York, USA
575 [5] T. Bindiya, E. Elias, Metaheuristic algorithms for the design of multiplier-less non-uniform filter banks based on frequency response masking, Soft
576 Comput. 18 (8) (2014a) 1529–1547.
577 [6] T. Bindiya, E. Elias, Modified metaheuristic algorithms for the optimal design of multiplier-less non-uniform channel filters, Circuits Syst. Signal Process.
578 33 (3) (2014b) 815–837.
579 [7] T. Bindiya, E. Elias, Design of totally multiplier-less sharp transition width tree structured filter banks for non-uniform discrete multitone system, AEU
580 Int. J. Electron. Commun. 69 (3) (2015) 655–665.
581 [8] T. Bindiya, V.S. Kumar, E. Elias, Design of low power and low complexity multiplier-less reconfigurable non-uniform channel filter using genetic
582 algorithm, Glob. J. Res. Eng. 12 (6) (2012) 7–19.
583 [9] D. Bose, S. Biswas, A.V. Vasilakos, S. Laha, Optimal filter design using an improved artificial bee colony algorithm, Inf. Sci. 281 (2014) 443–461.
584 [10] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
585 [11] O. Castillo, G. Huesca, F. Valdez, Evolutionary computing for topology optimization of type-2 fuzzy controllers, in: Hybrid Intelligent Systems, Springer,
586 2007, pp. 163–178.
587 [12] X. Chen, X. Xie, G. Shi, Direct design of near perfect reconstruction linear phase nonuniform filter banks with rational sampling factors, in: Proceedings
588 of the 2006 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2006, vol. 3, IEEE, 2006, p. III.
589 [13] R.V. Cox, The design of uniformly and nonuniformly spaced pseudoquadrature mirror filters, IEEE Trans. Acoust. Speech Signal Process. 34 (5) (1986)
590 1090–1096.
591 [14] S. Dasgupta, S. Das, A. Biswas, A. Abraham, On stability and convergence of the population-dynamics in differential evolution, Ai Commun. 22 (1)
592 (2009).
593 [15] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms, vol. 16, John Wiley & Sons, 2001.
594 [16] M.B. Dowlatshahi, H. Nezamabadi-Pour, M. Mashinchi, A discrete gravitational search algorithm for solving combinatorial optimization problems, Inf.
595 Sci. 258 (2014) 94–107.
596 [17] A. Eghbali, H. Johansson, On efficient design of high-order filters with applications to filter banks and transmultiplexers with large number of channels,
597 IEEE Trans. Signal Process. 62 (5) (2014) 1198–1209.
598 [18] A. Fernandez-Vazquez, G.J. Dolecek, Generalized Chebyshev filters for the design of IIR filters and filter banks, Circuits Syst. Signal Process. 33 (7)
599 (2014) 2237–2250.
600 [19] A.T. Fuller, B. Nowrouzian, F. Ashrafzadeh, Optimization of FIR digital filters over the canonical signed-digit coefficient space using genetic algorithms,
601 in: Proceedings of the 1998 Midwest Symposium on Circuits and Systems, 1998, IEEE, 1998, pp. 456–459.
602 [20] W.J. Gutjahr, Convergence analysis of metaheuristics, in: Matheuristics, Springer, 2010, pp. 159–187.
603 [21] R. Hartley, et al., Subexpression sharing in filters using canonic signed digit multipliers, IEEE Trans. Circuits Syst. II Analog Digital Signal Process. 43
604 (10) (1996) 677–688.
605 [22] P.-Q. Hoang, P. Vaidyanathan, Non-uniform multirate filter banks: Theory and design, in: Proceedings of the IEEE International Symposium on Circuits
606 and Systems, 1989, IEEE, 1989, pp. 371–374.
607 [23] B.-R. Horng, H. Samueli, et al., The design of low-complexity in linear-phase FIR filter banks using powers-of-two coefficients with an application to
608 subband image coding, IEEE Trans. Circuits Syst. Video Technol. 1 (4) (1991) 318–324.
609 [24] A. Jain, A. Goel, A multiobjective optimization method for designing M-channel NPR cosine modulated filter bank for image compression, Engineering
610 7 (02) (2015) 93.
Q6 611 [25] S. Kalathil, E. Elias, Efficient design of non-uniform cosine modulated filter banks for digital hearing aids, AEU Int. J. Electron. Commun. (2015),
612 doi:10.1016/j.aeue.2015.05.015.
613 [26] D. Karaboga, An Idea Based on Honey Bee Swarm for Numerical Optimization, Technical Report -TR06, Erciyes University, Engineering Faculty, Com-
614 puter Engineering Department, 2005.
615 [27] D. Karaboga, B. Basturk, Artificial bee colony (ABC) optimization algorithm for solving constrained optimization problems, in: Foundations of Fuzzy
616 Logic and Soft Computing, Springer, 2007, pp. 789–798.
617 [28] D. Karaboga, B. Basturk, A powerful and efficient algorithm for numerical function optimization: artificial bee colony (ABC) algorithm, J. Glob. Optim.
618 39 (3) (2007) 459–471.
619 [29] D. Karaboga, B. Basturk, On the performance of artificial bee colony (ABC) algorithm, Appl. Soft Comput. 8 (1) (2008) 687–697.
620 [30] R.D. Koilpillai, T.Q. Nguyen, P. Vaidyanathan, Some results in the theory of crosstalk-free transmultiplexers, IEEE Trans. Signal Process. 39 (10) (1991)
621 2174–2183.
JID: INS
622 [31] J. Kovačević, M. Vetterli, Perfect reconstruction filter banks with rational sampling factors, IEEE Trans. Signal Process. 41 (6) (1993) 2047–2066.
623 [32] S. Koziel, X.-S. Yang, Computational optimization, methods and algorithms, vol. 356, Springer Science & Business Media, 2011.
624 [33] A. Kumar, G.K. Singh, S. Anurag, Design of nearly perfect reconstructed non-uniform filter bank by constrained equiripple FIR technique, Appl. Soft
625 Comput. 13 (1) (2013) 353–360.
626 [34] K.S. Lee, Z.W. Geem, A new meta-heuristic algorithm for continuous engineering optimization: harmony search theory and practice, Comput. Methods
627 Appl. Mech. Eng. 194 (36) (2005) 3902–3933.
628 [35] D. Li, J. Song, Y. Lim, A polynomial-time algorithm for design digital filters with power-of-two coefficients, in: Proceedings of the 1993 IEEE Interna-
629 tional Symposium on Circuits and Systems, ISCAS ’93, 1993, pp. 84–87.
630 [36] L. Li, X. Xie, S. Chan, Design of perfect-reconstruction nonuniform filter banks with linear-phase property, in: Proceedings of the Seventeenth European
631 Signal Processing Conference, 2009, IEEE, 2009, pp. 383–387.
632 [37] Y. Lian, Y.J. Yu, Guest editorial: Low-power digital filter design techniques and their applications, Circuits Syst. Signal Process. 29 (1) (2010) 1–5.
633 [38] Y.C. Lim, R. Yang, D. Li, J. Song, Signed power-of-two term allocation scheme for the design of digital filters, IEEE Trans. Circuits Syst. II Analog Digital
634 Signal Process. 46 (5) (1999) 577–584.
635 [39] B.W.-K. Ling, C.Y.-F. Ho, K.-L. Teo, W.-C. Siu, J. Cao, Q. Dai, Optimal design of cosine modulated nonuniform linear phase FIR filter bank via both
636 stretching and shifting frequency response of single prototype filter, IEEE Trans. Signal Process. 62 (10) (2014) 2517–2530.
637 [40] V. Manoj, E. Elias, Design of multiplier-less nonuniform filter bank transmultiplexer using genetic algorithm, Signal Process. 89 (11) (2009) 2274–2285.
638 [41] V. Manoj, E. Elias, Artificial bee colony algorithm for the design of multiplier-less nonuniform filter bank transmultiplexer, Inf. Sci. 192 (2012) 193–203.
639 [42] M. Manuel, E. Elias, Design of frequency response masking FIR filter in the canonic signed digit space using modified artificial bee colony algorithm,
640 Eng. Appl. Artif. Intell. 26 (1) (2013) 660–668.
641 [43] P. Melin, F. Olivas, O. Castillo, F. Valdez, J. Soria, M. Valdez, Optimal design of fuzzy classification systems using PSO with dynamic parameter adaptation
642 through fuzzy logic, Expert Syst. Appl. 40 (8) (2013) 3196–3206.
643 [44] P. Mercier, S.M. Kilambi, B. Nowrouzian, Optimization of FRM FIR digital filters over CSD and CDBNS multiplier coefficient spaces employing a novel
644 genetic algorithm, J. Comput. 2 (7) (2007) 20–31.
645 [45] A. Milad, Harmony search algorithm: strengths and weaknesses, J. Comput. Eng. Inf. Technol. (2013).
646 [46] T. Nagai, T. Futie, M. Ikeh, Direct design of nonuniform filter banks, in: Proceedings of the 1997 IEEE International Conference on Acoustics, Speech,
647 and Signal Processing, ICASSP-97, vol. 3, IEEE, 1997, pp. 2429–2432.
648 [47] S. Olafsson, Handbooks in operations research and management science, Metaheuristics 13 (2006) 633–654.
649 [48] J. Princen, The design of nonuniform modulated filterbanks, IEEE Trans. Signal Process. 43 (11) (1995) 2550–2560.
650 [49] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, GSA: a gravitational search algorithm, Inf. Sci. 179 (13) (2009) 2232–2248.
651 [50] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, BGSA: binary gravitational search algorithm, Natural Comput. 9 (3) (2010) 727–745.
652 [51] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, Filter modeling using gravitational search algorithm, Eng. Appl. Artif. Intell. 24 (1) (2011) 117–122.
653 [52] P. Saghizadeh, A.N. Willson Jr, A new approach to the design of critically sampled M-channel uniform-band perfect-reconstruction linear-phase FIR
654 filter banks, IEEE Trans. Signal Process. 46 (6) (1998) 1544–1557.
655 [53] S. Saha, R. Kar, D. Mandal, S. Ghoshal, Optimal IIR filter design using gravitational search algorithm with wavelet mutation, J. King Saud Univ. Comput.
656 Inf. Sci. 27 (1) (2015) 25–39.
Q7 657 [54] V. Sakthivel, R. Chandru, E. Elias, Design of multiplier-less MDFT filter banks with perfect reconstruction using ABC algorithm, Int. J. Comput. Appl. 96
658 (1) (2014).
659 [55] T. Saramäki, Y.C. Lim, Use of the Remez algorithm for designing FRM based FIR filters, Circuits Syst. Signal Process. 22 (2) (2003) 77–97.
660 [56] H. Sarvari, K. Zamanifar, Improvement of harmony search algorithm by using statistical analysis, Artif. Intell. Rev. 37 (3) (2012) 181–215.
661 [57] G. Shi, X. Xie, X. Chen, W. Zhong, Recent advances and new design method in nonuniform filter banks, in: Proceedings of the 2006 International
662 Conference on Communications, Circuits and Systems Proceedings, vol. 1, IEEE, 2006, pp. 211–215.
663 [58] A. Sombra, F. Valdez, P. Melin, O. Castillo, A new gravitational search algorithm using fuzzy logic to parameter adaptation, in: Proceedings of the 2013
664 IEEE Congress on Evolutionary Computation (CEC), IEEE, 2013, pp. 1068–1074.
665 [59] R. Storn, K. Price, Differential evolution – a simple and efficient heuristic for global optimization over continuous spaces, J. Glob. Optim. 11 (4) (1997)
666 341–359.
667 [60] N.S. Thomaidis, V. Vassiliadis, Stochastic convergence analysis of metaheuristic optimisation techniques, in: Towards Advanced Data Analysis by Com-
668 bining Soft Computing and Statistics, Springer, 2013, pp. 343–357.
669 [61] P.P. Vaidyanathan, Multirate Systems and Filter Banks, Pearson Education India, 1993.
670 [62] F. Valdez, P. Melin, O. Castillo, Evolutionary method combining particle swarm optimization and genetic algorithms using fuzzy logic for decision
671 making, in: Proceedings of the IEEE International Conference on Fuzzy Systems, FUZZ-IEEE 2009, IEEE, 2009, pp. 2114–2119.
672 [63] F. Valdez, P. Melin, O. Castillo, A survey on nature-inspired optimization algorithms with fuzzy logic for dynamic parameter adaptation, Expert Syst.
673 Appl. 41 (14) (2014) 6459–6466.
674 [64] S. Wada, Design of nonuniform division multirate FIR filter banks, IEEE Trans. Circuits Syst. II Analog Digital Signal Process. 42 (2) (1995) 115–121.
675 [65] Y. Wei, Y. Wang, Design of low complexity adjustable filter bank for personalized hearing aid solutions, IEEE/ACM Trans. Audio Speech Lang. Process.
676 23 (5) (2015) 923–931.
677 [66] X. Xie, S. Chan, T. Yuk, Design of perfect-reconstruction nonuniform recombination filter banks with flexible rational sampling factors, IEEE Trans.
678 Circuits Syst. I Regul. Pap. 52 (9) (2005) 1965–1981.
679 [67] X. Xie, S. Chan, T. Yuk, Design of linear-phase recombination nonuniform filter banks, IEEE Trans. Signal Process. 54 (7) (2006) 2809–2814.
680 [68] X. Xie, X. Chen, G. Shi, A simple design method of linear-phase nonuniform filter banks with integer decimation factors, in: Proceedings of the Forty-
681 Eighth Midwest Symposium on Circuits and Systems, 2005, IEEE, 2005, pp. 724–727.
682 [69] Y. Xin-She, Nature-inspired mateheuristic algorithms: Success and new challenges, J. Comput. Eng. Inf. Technol. (2012).
683 [70] X.-S. Yang, Harmony search as a metaheuristic algorithm, in: Music-Inspired harmony Search Algorithm, Springer, 2009, pp. 1–14.
684 [71] X.-S. Yang, Nature-Inspired Metaheuristic Algorithms, Luniver Press, 2010.
685 [72] S. Yazdani, H. Nezamabadi-pour, S. Kamyab, A gravitational search algorithm for multimodal optimization, Swarm Evol. Comput. 14 (2014) 1–14,
686 doi:10.1016/j.swevo.2013.08.001.
687 [73] S. Yin, Y. Zhou, S. Chan, An efficient method for designing of modulated filter banks with causal-stable IIR filters, J. Signal Process. Syst. 78 (2) (2015)
688 187–197.
689 [74] Y.J. Yu, Y.C. Lim, A novel genetic algorithm for the design of a signed power-of-two coefficient quadrature mirror filter lattice filter bank, Circuits Syst.
690 Signal Process. 21 (3) (2002) 263–276.
691 [75] W.G. Zong, Music-inspired Harmony Search Algorithm: Theory and Applications, Springer, 2009.

Bindiya 2016

Uploaded by

Copyright:

Available Formats

You might also like

Bindiya 2016

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bindiya 2016

Uploaded by

Copyright:

Available Formats

JID: INS

ARTICLE IN PRESS [m3Gsc;January 5, 2016;17:54]

Information Sciences xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Meta-heuristic evolutionary algorithms for the design of

2 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 3

Fig. 1. Structure of recombination NUFB [36].

77 2. Review of recombination filter banks

Gl,i (z ) = Gl,i_F B (z ) Gl,i (z ) = z−1 Gl,i_F B (z ) (1)

4 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

Fig. 2. Structure of M-channel uniform ﬁlter bank [61].

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 5

Index CSD representation Decimal equivalent No. of SPT terms

900 0100-1000010-1 0.2195 4

119 3.1. Encoding of ﬁlter bank coeﬃcients using CSD look-up-table

6 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

174 3.2.1. Formulation of the objective function

199 3.2.2. Optimization using meta-heuristic algorithms

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 7

232 3.3. Overview of different meta-heuristic algorithms

234 3.3.1. Modiﬁed differential evolution (DE) algorithm

xi,G = [x1i,G x2i,G x3i,G . . . xDi,G ]

244 where D is the dimension of the solution vector.

vi,G+1 = xr1 ,G + (xr2 ,G − xr3 ,G )F , i ∈ 1, 2, . . ., N (20)

8 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

269 3.3.2. Modiﬁed artiﬁcial bee colony (ABC) algorithm

275 1. Initialize the food sources

xt+1 (i, j ) = xt (i, j ) + φ (xt (i, j ) − xt (k, j )) (23)

xt+1 (i, j ) = xt (i, j ) + φ (xt (i, j ) − xt (k, j )) (24)

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 9

298 4. Onlooker bee phase

313 3.3.3. Modiﬁed harmony search algorithm (HSA)

317 1. Initialization of harmony memory

xk = [xk,1 xk,2 xk,3 . . . xk,D ]

345 3.3.4. Modiﬁed gravitational search algorithm (GSA)

350 1. Initialization of agents

10 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 11

Fig. 3. RNUFB with sampling factors {2/3, 1/3}.

12 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

Amplitude distortion (dB) Number of multipliers Number of adders

428 4.2. Design of the multiplier-less recombination non-uniform ﬁlter bank

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 13

Fig. 6. RNUFB with sampling factors {2/3, 1/3}.

Word-length Amplitude distortion Number of adders

Continuous coeﬃcients 0.05402 0

435 4.2.1. Design of the maximum-precision recombination ﬁlter bank

14 T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx

Word-length Amplitude Number of Number of

445 4.2.2. Design of the CSD rounded recombination ﬁlter bank

T.S. Bindiya, E. Elias / Information Sciences xxx (2016) xxx–xxx 15

Population Mutation Fraction of Number of elite Number of

50 0.005 0.3 10 500

Population size, N Scaling factor, F Crossover ratio, CR Number of generations

50 0.7 0.01 500

Number of food sources Limit Number of generations

50 0.95 0.01 500

vi,G+1 = xr1 ,G + (xr2 ,G − xr3 ,G )F , i ∈ 1, 2, . . ., N (20)

xt+1 (i, j ) = xt (i, j ) + φ (xt (i, j ) − xt (k, j )) (24)