Documente Academic
Documente Profesional
Documente Cultură
a r t i c l e in f o a b s t r a c t
Available online 18 March 2010 Satellite communications technology has a tremendous impact in refining our world. The frequency
Keywords: assignment problem is of a fundamental importance when it comes to providing high-quality
Combinatorial optimization transmissions in satellite communication systems. The NP-complete frequency assignment problem in
Differential evolution satellite communications involves the rearrangement of frequencies of one set of carriers while keeping
Evolutionary computations the other set fixed in order to minimize the largest and total interference among carriers. In this paper,
Frequency assignment Problem we present a number of algorithms, based on differential evolution, to solve the frequency assignment
NP-complete problem. We investigate several schemes ranging from adaptive differential evolution to hybrid
Satellite communication algorithms in which heuristic is embedded within differential evolution. The effectiveness and
robustness of our proposed algorithms is demonstrated through solving a set of benchmark problems
and comparing the results with a number of previously proposed techniques that solve the same
problem. Experimental results show that our proposed algorithms, in general, and hybrid ones in
particular, outperform the existing algorithms both in terms of the quality of the solutions and
computational time.
Crown Copyright & 2010 Published by Elsevier Ltd. All rights reserved.
0305-0548/$ - see front matter Crown Copyright & 2010 Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.cor.2010.03.004
ARTICLE IN PRESS
A.A. Salman et al. / Computers & Operations Research 37 (2010) 2152–2163 2153
in [9] finds better solutions by taking care of constraint satisfaction algorithms are indeed very effective and efficient to solve for the
and the goal function optimization simultaneously. frequency assignment problem.
Sanz et al. [10,11] presented hybrid methods that combine The rest of the paper is organized as follows. A brief
the Hopfield neural network with simulated annealing [10] introduction to FAP in satellite communications is presented in
and genetic algorithm [11] to solve the frequency assignment Section 2. Section 3 provides an overview for Differential
problem. In their technique, fast digital Hopfield neural network Evolution algorithm. Our proposed methods are presented in
manages the problem’s constraints and simulated annealing Section 4. Experimental results are reported and discussed in
or genetic algorithm searches for better quality solutions. Section 5. Finally, in Section 6 we draw some conclusions.
Wang et al. [12,13] presented a noisy chaotic neural network
with variable threshold to solve the frequency assignment
problem. In their approach, the optimization term was separated 2. Problem definition
from the constraint term in the cost function by assigning different
neurons with variable thresholds. This technique obtains better Satellite communications technology has become of paramount
solutions compared with both techniques proposed in [9,10]. importance in global economy and culture [1,2]. In satellite
In this work, we propose new, fast, and scalable algorithms communication systems, two or more satellites in the same orbit
based on differential evolution (DE) to solve the frequency experience cochannel interference when using the same frequency
assignment problem. Differential evolution is a relatively new which severely degrades the quality of transmission in the system.
evolutionary algorithm, proposed by Storn and Price [14], which Therefore, to provide high-quality transmissions, cochannel inter-
has been successfully applied to solve a wide range of optimiza- ference minimization becomes an important design issue in
tion problems [15–17]. DE is similar to others evolution satellite communications [3]. Frequency assignment problem in
algorithms where a population of individuals are used to search satellite communication systems is described as a typical example
for an optimal solution [18]. The key difference is that mutation in of constrained combinatorial optimization problem [9–13].
DE algorithm is an arithmetic combination of individuals [18] This paper follows the FAP formulation given in [7] which has
whereas in traditional evolutionary algorithms, it is the result of been applied in numerous algorithms such as [9–13]. Fig. 1
small perturbations to the genes of an individual. Moreover, in DE, illustrates intersystem cochannel interference between two
the trial solutions are generated by adding weighted difference adjacent satellites when operating on the same frequency band.
vectors to the target vector followed by a recombination In such a system, each carrier can be described as a collection of
(or crossover) step to produce an offspring which is only accepted consecutive unit segments in a frequency band [6], hence, the
if it improves the fitness of the parent individual. DE is easy to interference between two M-segment systems can be described
implement, requires little parameter tuning [19] and exhibits fast by an M M matrix (denoted as E) where each element, eij, stands
convergence [20]. Nevertheless, due to its continuous nature, for the cochannel interference cost when segment ‘‘i’’ in system 2
applying DE algorithm to solve combinatorial problems is still uses a common frequency with segment ‘‘j’’ in system 1
limited. In this work, we investigate the potential use of DE to (interference cost between segment ‘‘i’’ in system 2 and
provide effective and efficient solution for FAP in satellite segment ‘‘j’’ in system 1). Therefore, the FAP can be described as
communications. The major obstacle in tailoring DE towards our the rearrangement of the frequencies assigned to system 2
needs is that DE is introduced as global optimizer over continuous carriers to minimize the largest interference and at the same
search spaces while FAP in satellite communications is of time minimizing the sum of total interference (hence multi-
combinatorial nature with discrete decision variables. Therefore, objectives FAP [9]). For example, Fig. 3 shows the interference
to apply DE on FAP we have either to find an appropriate mapping matrix for the segmented frequency banks systems in Fig. 2 and a
algorithm to convert floating-point vectors to actual frequency possible optimal reassignment of the frequencies as shown in
assignment solution, or to adapt basic DE operators to present the Fig. 4. Therefore, to solve FAP, the following three constraints
FAP solution. We propose several schemes to solve FAP problem (C1, C2, C3) must be fulfilled [9–13]:
using DE such as adaptive differential evolution, standard DE with
appropriate mapping, and hybrid algorithms (where a greedy C1 Every segment in system 2 must be assigned to a segment in
heuristic is embedded within adaptive or standard differential system 1.
evolution) as it will be discussed later. A thorough analysis has C2 Every segment in system 1 can be assigned to at most one
been performed to provide sufficient evidences that the proposed segment in system 2.
Sat 1
Sat 2
Interference
System 1 System 2
ES 1 ES 2
Fig. 1. Outline of cochannel interference. Fig. 2. Cochannel interference model of the system in Fig. 1.
ARTICLE IN PRESS
2154 A.A. Salman et al. / Computers & Operations Research 37 (2010) 2152–2163
C3 All the segments of each carrier in system 2 must be assigned mutation operator favors exploration, however, as evolution
to consecutive segments in system 1 in the same order. progresses, the mutation operator favors exploitation [21]. Hence,
DE automatically adapts the mutation increments (i.e. search step)
Moreover, the objectives of FAP are to minimize the largest to the best value based on the stage of the evolutionary process and
interference and to minimize the sum of total interference, hence, not based on a predefined probability density function. DE has been
in order to minimize the peak of interference between the successfully applied to solve a wide range of optimization problems
systems (largest interference), the first objective function can be such as clustering [19], unsupervised image classification [22],
described as follows [10,11]: digital filter design [23], and other non-linear function/process
optimization [24–27]. In short, DE is generally considered as a
F1 ¼ maxðei,j fi,j Þ 8i,j where i,j ¼ 1, . . . ,M ð1Þ
reliable, accurate, robust, and fast optimization technique used in
where M and ei,j represent the number of segments and cochannel many practical optimization problems.
interference cost between ‘‘i’’ and ‘‘j’’, respectively. Moreover, fi,j is Unlike other evolutionary algorithms, differential evolution (DE)
a boolean function that returns 1 if segment ‘‘i’’ has been assigned does not use a probability distribution function to introduce
to segment ‘‘j’’, and 0 otherwise. The second objective function of variations into the population. Instead, DE uses the difference
FAP requires that the total interference of the systems to be between randomly selected vectors (individuals) as the source of
minimum resulting in a second optimization function as follows: variation for a third vector referred to as the target vector. Trial
solutions are generated by adding a weighted difference vectors to
M X
X M
F2 ¼ eij fij ð2Þ the target vector. This process is referred to as the mutation operator
i¼1j¼1 where the target vector is mutated. A recombination, or crossover
step is then applied to produce an offspring which is only accepted if
it improves the fitness of the parent individual. The basic DE
3. Differential evolution algorithm is described in more detail below with reference to the
three evolution operators: mutation, crossover, and selection:
Differential evolution is an evolution algorithm proposed by Mutation: For each parent, xi(t) of generation t, a trial vector,
Storn and Price [14] which is similar to other evolutionary vi,j(t), is created by mutating a target vector. Randomly select the
algorithms where a population of individuals is used to search for target vector xi3 ðtÞ, with ia i3 . Then, two individuals xi1 ðtÞ, and
an optimal solution [18]. The main difference between traditional xi2 ðtÞ are randomly selected with i1 a i2 a i3 ai and the difference
evolutionary algorithms and DE is that in traditional evolutionary vector xi1 xi2 is calculated. The trial vector is then calculated as
algorithms, mutation results in small perturbations to the genes of vi,j ðtÞ ¼ xi3 ðtÞ þFðxi1 ðtÞxi2 ðtÞÞ ð3Þ
an individual while in DE the mutation is an arithmetic combination
of individuals [18]. At the beginning of the evolution process, the DE where the term Fðxi1 ðtÞxi2 ðtÞÞ A ð0,1Þ represents the mutation
step size and F is a scaling factor used to control the amplification
of the differential variation.
C11 C12 C13
Crossover: DE follows a discrete recombination approach
S11 S12 S13 S14 S15 S16 where elements from the parent vector xi(t) are combined with
elements from the trial vector, vi(t), to produce the offspring mi ðtÞ.
C21 S21 20 20 40 0 25 25
Assuming a binomial crossover, the offspring can be described
S22 50 10 30 0 55 * using the following equation:
C22 (
S23 * 50 30 0 15 55 vi,j ðtÞ if randð0,1Þ oCR or j ¼ r
mi,j ¼ x ðtÞ otherwise ð4Þ
i,j
C23 S24 30 20 45 0 15 35
C11
System #1 C12 C13
S11 S12 S13 S14 S15 S16 Segment
directly from the parent. Even when CR ¼0, at least one of the directing the search toward optimal solution, then selecting a
parameters of the offspring will be different from the parent proper objective function is of paramount importance [29]. Our
(forced by the condition j¼r). objective function (fitness function) is a weighted sum of both, the
Selection: DE implements selection procedure by forcing FAP’s constraints violation and the values of the optimization
the generated offspring, mi ðtÞ to replace the parent, xi(t), only if functions (i.e. largest and total interference). The weight of
the fitness of the offspring is better than that of the parent. violating any constraint is higher than that of the optimization
Algorithm 1 shows the pseudo-code for the DE: function to ensure producing feasible solutions. Therefore, we
propose the following fitness function:
Algorithm 1. Simple DE algorithm.
FðTÞ ¼ WH ðC1þ C2 þ C3Þ þ W1 F1 þW2 F2 ð5Þ
Initialize a population P;
Evaluate P; where WH is the weighting coefficient of constraints violation and
repeat Wi represents the weighting coefficients for objective ‘‘i’’. More-
over, Ci denotes constraint ‘‘i’’ of FAP whereas Fi is the objective
for each individual in the population do
function (largest and total interference, respectively) as was
Let i ,i , and i be three random integers in ½1,s;
1 2 3
described in Section 2.
Let r ¼ randð1,Nd Þ;
for j ¼ 1 to Nd do
4.1. DE with permutation linkage (DEPL)
if randð0,1Þ o CR or j ¼ r then
v ðtÞ ¼ x þ Fðx x Þ;
i,j i3 ,j i1 ,j i2 ,j
In this algorithm we adapt DE by employing the idea of linkage
else crossover [30,31] and discrete permutation [32]. In DEPL, each
vi,j ðtÞ ¼ xi,j ; individual is represented by discrete numbers where a target
individual is modified by permuting its values and then the
end
linkage crossover is utilized in an attempt to create a valid
end
solution. A detailed description of the various steps used in DEPL
if f ðvi Þ of ðxi Þ then algorithm are given below:
xi ¼ vi ; Representation: Each individual represented by an array of size
end M where M is number of segments and is shown in Fig. 5. In this
end figure, the value of kth gene of i th individual specifies the
until the terminating condition is achieved; segment from system ‘‘1’’ which has been assigned to the kth
segment in system ‘‘2’’. This representation has the advantage
that it can always satisfy the first and the second constraints,
hence C 1 and C 2 in Eq. (5) can always be set to zero while C 3
Price and Storn [28] proposed ten different strategies for DE
must be checked for each solution. The solution will be considered
based on the individual being perturbed, the number of indivi-
feasible when C3 evaluates to zero.
duals used in the mutation process, and the type of crossover used.
Initialization: Each gene is initialized with a random segment
The strategy described above is known as ‘‘DE/rand/1’’ using the
from system ‘‘1’’ such that there are no two genes assigned to the
notation ‘‘DE/x/y’’ where ‘‘x’’ represents the individual being
same segment, hence, the initial population starts with indivi-
perturbed, and ‘‘y’’ gives the number of difference vectors used
duals which do not necessary satisfy constraint C3 but always
to perturb ‘‘x’’ [27]. Therefore, ‘‘DE/rand/1’’ strategy means that
fulfill constraints C1 and C2.
the target vector is randomly selected and only one difference
Mutation: In this approach, the mutation operator is applied to
vector is used. This strategy is considered the most successful and
rearrange the starting positions (i.e. the first assigned segment) of
widely used one [27]. Other DE strategies include ‘‘DE/best/1’’,
carriers to investigate new solution as described in the Pseudo-
‘‘DE/best/2’’, ‘‘DE/rand/2’’ and ‘‘DE/randto-best/1’’. A detailed
code given in Algorithm 2.
discussion of these and other strategies can be found in [29].
Algorithm 2. A pseudo-code for mutation operator in DEPL
method.
4. Proposed approaches
for each carrier do
One of the major challenges in tailoring DE to be applied to FAP if randð0,1Þ oCR or j ¼ r then
in satellite communication problem is that DE is introduced as a Let s ¼ current starting segment for the carrier;
i
global optimizer over continuous search space while FAP is of a Let sx a si is a randomly selected segment and
sx can be used as newly
combinatorial nature with discrete decision variables. Therefore,
starting segment for the carrier;
to apply DE on FAP we have either to find an appropriate mapping
Swap the places of s and s ;
algorithm to convert floating-point vectors to the actual fre- i x
quency assignment solutions, or adapt basic operators of DE to end
present the FAP solutions. In this paper we consider both of these end
approaches in addition to approaches in which heuristics are
embedded within DE. The remainder of this section describes the Crossover: In crossover operator, we employ the idea of linkage
proposed methods and how DE has been tailored to solve FAP by: crossover [30,31] with minor modifications. In our approach, the
(1) identifying the representation of population individuals, strength of linkages between components of individual structure
(2) describing the mechanism of generating initial population,
(3) describing any modification to DE operators (e.g. mutation, C21 C22 C23 C24
(i.e. segments) are determined dynamically for each individual Initialization: Each gene is initialized with a random value
and are controlled by carriers’ lengths. For example; if carrier i between [0,1] representing the carrier’s priority.
with length ci ¼3 is assigned to segment 2 from system ‘‘1’’, then Mutation and Crossover: PDE approach uses standard DE
according to this rule, there is a strong linkage between the operators as it was described in Section 3 and are used to
consecutive segments 2, 3 and 4 in system ‘‘1’’. generate new priority values for each carrier.
Mapping to FAP solutions: By ranking the carriers according to
Selection: All proposed algorithms use the standard selection
their priorities, the segments from system ‘‘1’’ are assigned in a
operator, thus in the selection phase, the generated offspring mi ,
consecutive manner. For example, given the priority assignment
replaces parent, xi only if its fitness is better than its parent.
to each carrier as shown in Fig. 6 and given the same example of
The starting segment of each carrier in the offspring mi is either
FAP problem used earlier in Section 2, the carriers will be
inherited from the mutated individual (vi) or inherited from the
arranged as C24, C23, C21, C22. Since both C24 and C22 can carry
parent individual (xi). All remaining segments are re-arranged to
two segments each, while C23 and C21 can carry one segment,
be consecutive with the current starting segment. Algorithm 3
then segments 1, 2 from system ‘‘1’’ will be assigned to C24,
shows the pseudo-code to produce the offspring mi in DEPL.
segment 3 from system ‘‘1’’ will be assigned to C23, segments 4
from system ‘‘1’’ will be assigned to C21, and segment 5, 6 from
Algorithm 3. A pseudo-code for crossover operator in DEPL system ‘‘1’’ will be assigned to C22.
method.
C21 20 20 40 0 25 25
2 1 3 4
C23 30 20 45 0 15 35
In this method, DE priority version as introduced in Section 4.2 Fig. 8. Total interference cost of each carrier using interference matrix in Fig. 3.
is used to rearrange the carriers, however, the segments from
system ‘‘1’’ are assigned to each carrier using one of the following
heuristics:
S11 S12 S13 S14 S15 S15
1. First heuristic (DEH1): From the available consecutive segments
C21 20 20 40 0 25 25
in system ‘‘1’’ (i.e. not been assigned yet), select the ones which
produce the lowest largest interference with current carrier in C22 50 30 30 15 55 *
system ‘‘2’’.
2. Second heuristic (DEH2): From the available consecutive C23 30 20 45 0 15 35
segments in system ‘‘1’’, select the ones which produce the
lowest total interference with current carrier in system ‘‘2’’. C24 45 25 25 10 50 *
3. Third heuristic (DEH3): From the available consecutive seg-
ments in system ‘‘1’’, select the ones which produce both the Fig. 9. The largest cost interference of each carrier using the matrix interference in
lowest largest interference and the lowest total interference Fig. 3.
Table 1
Specifications of the FAP instances.
Table 2
Comparison of the results obtained by proposed methods (best/average/standard deviation).
coefficient for problem’s constraints was set to 1 106 while the for benchmarks BM4 and BM5 using 50 runs. In the figure, the
weighting coefficients W 1 and W 2 were set to 50.0 and 0.01, vertical axis represents the function value and the horizontal axis
respectively, when the objective was to obtain solution with low represents the number of generations needed. It can be seen from
largest interference. In the case when the objective was to obtain the figure that the fastest time to converge is achieved by DEH1
low total interference solution, W1 and W2 were set to 0.5 and and DEH3 whereas DEPL requires larger time to converge when
5.0, respectively. compared to other algorithms.
they are all competitive in terms of the averaged recorded best largest DEH3 algorithm for BM6 and BM8 are shown in Figs. 11 and 12,
interference, computational time, and the best and averaged total respectively, where a black square stands for assigning the carrier to
interference. In BM1–BM5 we can see that generally DEH1 and DEH3 the corresponding segment, while the white square stands for no
outperformed all other algorithms with averages very close to optimal assignment. For BM7; DEH1 and DEH3 achieved the same best largest
solutions and with the lowest computational time among all. interference as reported by NCNN-VT. Even thought NCNN-VT
However, none of our proposed algorithms reached the best total obtained same instances which has better total interference, on
interference achieved by NCNN-VT in BM5, except for DEH2 which average, DEH1 and DEH3 outperform it with much less
achieved better total interference but with higher largest interference. computational time. It is clear that all our proposed algorithms in
BM6 and BM8 demonstrate clearly the effectiveness and robustness this paper reached better or at least the same performance obtained
of our proposed algorithms, by achieving better solutions in terms of by HopSA and GNN.
best largest interference and total interference with very low
computational time (except for DEPL). New best solutions found by
5.2.2. Pair-wise and Global comparison
Fig. 13 shows a pair-wise and a global comparison among all
algorithms summarized from Tables 2 and 3. The comparison has
been performed by observing the number of times the average of
best largest interference for each algorithm was better, worse or
the same compared to the average of every other algorithm for all 8
benchmarks. Each box in Fig. 13 compares the algorithm in the left
side to the algorithm on the top, where ‘‘4 ’’ , ‘‘o ’’ and ‘‘¼’’ signs
indicate the number of times the average of the algorithm on the
left was better, worse or the same, respectively, compared to the
algorithms listed on the top. For example, the average of DEH1
algorithm was better than the average of DEH2 in five benchmarks,
same in three benchmarks and worse in zero benchmarks. For the
global comparison, an extra box labeled (‘‘All’’) is used. The
numbers in this box indicate the performance of each algorithm
in left side comparing to all other algorithms combined. So based
on results shown in box ‘‘All’’, we can rank the algorithms in an
increasing order of best solutions obtained as: DEH3, DEH1, PDE,
DEH2 and PPDE, NCNN-VT, HopSA, DEPL, and GNN.
Table 3
Performance of GNN [9], HopSA [10] and NCNN-VT [13] methods on eight instances (from [13]).
Largest Best/Ave Total Best/ave Largest Total Largest Best/Ave/SD Total Best/ave/SD
Table 4
Computational time (measured in seconds) comparison between all methods.
DEH1 0.00 (0.00) 0.00 (0.00) 0.16 (0.15) 0.09 (0.05) 0.26 (0.26) 3.71 (1.56) 58.40 (25.63) 1.26 (0.84)
DEH2 0.00 (0.00) 0.00 (0.00) 0.03 (0.02) 0.11 (0.08) 0.06 (0.15) 4.22 (1.73) 43.18 (13.91) 1.69 (1.17)
DEH3 0.00 (0.00) 0.00 (0.00) 0.28 (0.25) 0.17 (0.10) 0.64 (0.56) 12.03 (4.89) 69.64 (44.41) 4.62 (2.08)
DEPL 2 (1.73) 2.38 (2.05) 222.8 (118.94) 224.98 (132.93) 237.3 (113.99) 1076.76 (204.42) 809.17 (264.84) 759.24 (231.15)
PDE 0.00 (0.00) 0.00 (0.00) 0.18 (0.12) 0.62 (0.39) 0.71 (0.49) 5.75 (3.39) 18.42 (12.61) 3.40 (2.04)
PPDE 0.00 (0.00) 0.00 (0.00) 1.36 (0.83) 1.48 (0.69) 1.84 (1.24) 15.55 (8.08) 25.58 (16.75) 7.68 (2.10)
Segments
1 60
Carrier
18
Fig. 11. Best solution obtained by DEH3 for instance BM 6, with largest interference ¼ 32, total interference ¼ 981.
Segments
Carrier 1 50
15
Fig. 12. Best solution obtained by DEH3 for instance BM 8, with largest interference ¼620, total interference ¼ 12 653.
tested, excellent rate of reaching the optimum solutions, as well population was generated for each algorithm as has been
as, they are very fast (excluding DEPL) compared to other. described in Section 4. Each proposed algorithm was run 30
times for each problem instance with maximum number of
generations set to 1500. Since almost all literature investigation
5.3. Experimental set 2 concentrates on achieving minimum largest interference not
minimum total interference, hence no literature results does exit
The second set of experiments has been conducted with the for comparison.
aim to produce solutions with minimum total interference. Tables 6 and 7 list the results of the conducted experiments.
Therefore, W 1 was set to 0.5 and W 2 was set to 5.0. The initial Table 6 lists a representative set of results (best/average/standard
ARTICLE IN PRESS
A.A. Salman et al. / Computers & Operations Research 37 (2010) 2152–2163 2161
>2 >16
HopSA =3 =22
<3 <26
>20
NCNN-VT =20
<24
Fig. 13. A Global comparison of FAP algorithms averaging over 50 runs: ‘‘ 4’’ , ‘‘¼’’, ‘‘o ’’ refers to better, same, worse than in n times.
Table 5
Comparison between NCNN-VT [13] and DEH1, DEH3.
deviation for minimum largest and total interference) when the algorithm obtained the best total interference in all benchmarks
proposed algorithms were applied to benchmarks BM1–BM8 and (except in BM3) with the shortest time and with an excellent rate
Table 7 compares the proposed algorithms in terms of their of achieving the optimum solution. Further, DEH3 was able to
average computational time. achieve comparable results in this experiment as in the previous
rom Tables 6 and 7, it can be seen that all proposed algorithms experiment (when minimum largest interference was the main
reached the same lowest total interference for BM1 and BM2. objective) demonstrating its effectiveness in either case. DEPL,
However, for BM4–BM8, DEH1 algorithm which was able to find PDE and PPDE algorithms found the optimal total interference in
the best solutions in first set of experiments, failed to find the best BM1–BM5, however, their performance differs for BM6–BM8. The
solution in this experiment . This due to the fact that the heuristic longest computational time recorded for all benchmarks was
embedded within DEH1 always looks for the minimum largest when running DEPL algorithm which we believe is caused by the
interference regardless of the total interference. Moreover, DEH2 representation used in the algorithm (i.e. Nd ¼M).
ARTICLE IN PRESS
2162 A.A. Salman et al. / Computers & Operations Research 37 (2010) 2152–2163
Table 6
Comparison between proposed methods for experiment set 2 (best/average/standard deviation).
Table 7
Computational time comparison.
6. Conclusions [2] Whalen DJ. Communications satellites: making the global village possible,
2008 /http://www.hq.nasa.gov/office/pao/History/satcomhistory.htmlS.
[3] Jeruchim M. A survey of interference problems and applications to
In this paper, we have presented a number of novel approaches geostationary satellite networks. In: Proceedings of the IEEE, vol. 65, 1977.
to solve the frequency assignment problem in satellite commu- p. 317–31.
nication systems. All proposed algorithms in this work are [4] Murphy RA, Pardalos PM, Resende MGC. Frequency assignment problems.
Kluwer, MA: Norwell; 1999.
inspired by Differential Evolution algorithm. Several schemes [5] Aardal KI, van Hoesel SPM, Koster AMCA, Mannino C, Sassano A. Models and
such as: adaptive differential evolution which represents FAP solution techniques for frequency assignment problems. Annals of Operations
solutions directly, standard DE with appropriate mapping algo- Research 2007;153(1):79–129.
[6] Mizuike T, Ito Y. Optimum frequency assignment for reduction of cochannel
rithm to present FAP solutions, and hybrid algorithms in which
interference. Transactions on IEICE 1986;J69-B(9):921–32.
heuristics are embedded within adaptive or standard differential [7] Mizuike T, Ito Y. Optimization of frequency assignment. IEEE Transactions on
evolution have been investigated. The effectiveness and robust- Communications 1989;37(10):1031–41.
ness of our proposed algorithms are demonstrated through [8] Kurokawa T, Kozuka S. A proposal of neural network for the optimum frequency
assignment problem. Transactions on IEICE Journal 1993;76-B-II(10):811–9.
solving a set of benchmark problems and comparing results [9] Funabiki N, Nishikawa S. A gradual neural-network approach for frequency
obtained with other existing methods. Experimental results have assignment in satellite communication systems. IEEE Transactions on Neural
shown that our proposed algorithms in general and hybrid ones in Networks 1997;8(6):1359–70.
[10] Sanz SS, Mozos RS, Calzon CB. A hybrid Hopfield network-simulated
particular are robust and were able to find comparable and even
annealing approach for frequency assignment problem in satellite commu-
better solutions (i.e. for some benchmarks) with low computa- nications systems. IEEE Transactions on Systems, Man, and Cybernetics—Part
tional time. Therefore, algorithms based on differential evolution B: Cybernetics 2004;34(2):1108–16.
paradigm provide a viable approach for solving frequency [11] Sanz SS, Calzon CB. A hybrid neural-genetic algorithm for the frequency
assignment problem in satellite communications. Applied Intelligence
assignment problem in satellite communication systems. 2005;3(22):207–17.
[12] Liu W, Shi H, Wang L. Minimizing interference in satellite communications
using chaotic neural networks. In: IEEE third international conference on
natural computation (ICNC 2007), 2007. p. 441–4.
Acknowledgments [13] Wang L, Liu W, Shi H. Noisy chaotic neural networks with variable thresholds
for the frequency assignment problem in satellite communications. IEEE
Transactions on Systems, Man, and Cybernetics—Part C: Applications and
The authors would like to thank Prof. L. Wang and his team, at
Reviews 2008;38(2):209–17.
the School of Electrical Engineering, Nanyang Technological [14] Storn R, Price K. Differential evolution—a simple and efficient adaptive
University, Singapore, for providing us the benchmarks. scheme for global optimization over continuous spaces. Technical report TR
95-012, International Computer Science Institute, Berkeley, CA; 1995.
[15] Storn RM, Price KV. Differential evolution—a simple and efficient heuristic for
global optimization over continuous spaces. Journal of Global Optimization
References
1997;11(4):341–59.
[16] Rae A, Parameswaran S. Synthesizing application-specific heterogeneous
[1] Pelton JN, Oslund RJ, Marshall P. Communications satellites: global change multiprocessor using differential evolution. IEICE Transactions on Funda-
agents. Mahwah, NJ: Lawrence Erlbaum; 2004. mentals E 2001;84-A(12):3125–31.
ARTICLE IN PRESS
A.A. Salman et al. / Computers & Operations Research 37 (2010) 2152–2163 2163
[17] Rzadca K, Seredynshi F. Heterogeneous multiprocessor scheduling with [25] Angira R, Babu B. Evolutionary computation for global optimization of non-
differential evolution. In: Proceedings of congress on evolutionary computa- linear chemical engineering processes. In: Proceedings of International
tion, 2005. p. 2840–7. Symposium on Process Systems Engineering and Control, Mumbai, 2003.
[18] Feoktistov V, Janaqi S. Generalization of the strategies in differential p. 87–91.
evolution. In: Proceedings of the 18th international parallel and distributed [26] Abbass H (editor). A memetic pareto evolutionary approach to artificial
processing symposium, 2004. p. 165–70. neural networks. Lecture notes in artificial intelligence, vol. 2256. Berlin,
[19] Paterlini S, Krink T. High performance clustering with differential evolution. Germany: Springer-Verlag; 2002.
In: Proceedings of the IEEE congress on evolutionary computation, vol. 2, [27] Babu B, Jehan M. Differential evolution for multi-objective optimization. In:
2004. p. 2004–11. Proceedings of the IEEE congress on evolutionary computation, vol. 4, 2003.
[20] Karaboga D, Okdem S. A simple and global optimization algorithm for
p. 2696–703.
engineering problems: differential evolution algorithm. Turkish Journal of
[28] Price K, Storn R. De web site, 2005. URL /http://www.ICSI.Berkeley.edu/
Electrical Engineering 2004;12(1):53–60.
storn/code.htmlS.
[21] Xue F, Sanderson A, Graves R. Pareto-based multi-objective differential
[29] Storn R. On the usage of differential evolution for function optimization. In:
evolution. In: Proceedings of the IEEE congress on evolutionary computation,
The North American fuzzy information processing society conference,
vol. 2, 2003. p. 862–9.
[22] Omran M, Engelbrecht A, Salman A. Differential evolution methods for Berkeley, 1996. p. 519–23.
unsupervised image classification. In: Proceedings of the IEEE congress on [30] Salman AA, Mehrotra K, Mohan CK. Adaptive linkage crossover. In: Proceedings
evolutionary computation, vol. 2, 2005. p. 966–73. of ACM symposium on applied computing, 1998. p. 338–42.
[23] Storn R. Differential evolution design for an iir-filter with requirements for [31] Salman AA, Mohan CK. Linkage crossover operator for genetic algorithms.
magnitude and group delay. Technical report TR 95-026, International PhD thesis, Syracuse University, Syracuse, NY; 1999.
Computer Science Institute, Berkeley, CA; 1995. [32] Tasgetiren MF, Pan Q-K, Suganthan PN, Liang Y-C. A discrete differential
[24] Babu B, Angira R. Optimization of non-linear functions using evolutionary evolution algorithm for the no-wait flowshop scheduling problem with total
computation. In: Proceedings of the 12th ISME international conference on flowtime criterion. In: Proceedings of the 2007 IEEE symposium on
mechanical engineering, India, 2001. p. 153–7. computational intelligence in scheduling, 2007. p. 338–42.