Professional Documents
Culture Documents
Research Article
ISSN 1751-8725
Sidelobe-level suppression for linear and Received on 28th January 2016
Revised on 31st May 2016
circular antenna arrays via the cuckoo search– Accepted on 19th July 2016
doi: 10.1049/iet-map.2016.0083
chicken swarm optimisation algorithm www.ietdl.org
Abstract: An antenna array application with high directivity and low sidelobe level (SLL) is an important technology that
can enhance the reliability and validity of a communication system. In recent years, swarm intelligence optimisation
algorithms have been widely used in the design of antenna arrays. In this study, a hybrid heuristic swarm intelligence
optimisation algorithm called cuckoo search–chicken swarm optimisation (CSCSO) is proposed to optimise the
excitation amplitude of a linear antenna array (LAA) and the excitation amplitude and spacing between the array
elements of a circular antenna array (CAA). The maximum SLL will be obtained while the mainlobe width is fixed.
CSCSO combines cuckoo search (CS) and chicken swarm optimisation (CSO) and thus, it will have the excellent global
search capability of CS as well as introduce the hierarchy mechanism of CSO to improve algorithm precision. Chaos
theory is utilised to determine the initial solution of the algorithm, while Levy flight weight coefficient is used to
improve the convergence rate in CSCSO. Simulation results show that CSCSO has a better performance in terms of
solution accuracy and convergence rate in the radiation pattern optimisation of LAA and CAA compared with the
standard CSO, CS, and particle swarm optimisation algorithms.
2.1.1 Linear antenna array: Fig. 1a shows the geometry of an 2.2 Fitness function
LAA with 2N isotropic antenna elements placed on the x-axis. The
array factor can be described as follows 2.2.1 LAA fitness function: A low sidelobe generally denotes a
good beam pattern. Therefore, we design the fitness function by
N comparing the maximum sidelobe to optimise the beam pattern.
S f = In exp j knd cos f + wn , (1) The fitness function can be defined as follows
n=−N
n=0
S f p
where k represents the wave number (k = 2π/l). In and jn represent fitness = max 20log10 , f [ − , fn , (7)
the excitation amplitude and the phase of the nth element, S f0 2
respectively. Assuming that In = I−n, jn = j−n, and 2N array
elements are placed symmetrically along the x-axis, the array where [−π/2, fn] represents the sidelobe region and f0 represents the
factor can be simplified as follows mainlobe direction (f0 = 0°). The purpose of the linear array
optimisation problem is to determine a set of array element
N excitation amplitudes that can minimise the fitness function value.
S f =2 In cos knd cos f . (2)
n=1
2.2.2 CAA fitness function: The fitness function of CAA is
designed to minimise the SLL while the first null bandwidth is
In this study, jn is set to be 90°. The excitation amplitude In ranges fixed. Thus, we use the following fitness functions
from 0 to 1. The distance among adjacent elements is equal, and the
value is d = l/2. 2
Fitness = F1 + F2 /Smax , (8)
2.1.2 Circular antenna array: Fig. 1b shows the CAA 2 2
composed of N elements. Supposing that N elements are in the F1 = S wnu1 +S wnu2 , (9)
same radiation direction, the normalised array factor can be written
2 2
as follows F2 = max S wms1 , S wms2 , (10)
N where jnu is the angle at a null. In CAA, jnu1 and jnu2 are the first
S u, w = In exp j ka sin(u) cos w − wn + fn , (3) deep nulls on both sides of the main lobe, and the mainlobe width is
n=1 defined as [jnu1, jnu2]. jms1 and jms2 are the angles where the
maximum SLL is attained during the optimisation process from
where the lower band [0, jnu1] to the upper band [jnu2, 360]. Smax is the
maximum value of the array factor. Thus, the fitness function is
2p N
designed to obtain the lowest maximum SLL. Similarly, the
ka = a= di , (4) purpose of the CAA optimisation problem is to identify the
l i=1
excitation amplitudes and positions of array elements that can
N minimise the fitness value.
2p di
wn = i=1
, (5)
ka
3 Algorithms
where In and fn are the excitation amplitude and excitation phase of
the nth element, respectively, and jn is the phase of the nth element
3.1 Cuckoo search
in the x–y plane. To cause the mainlobe to locate at (θ0, j0), the
excitation phase should be set as In nature, CS for a suitable nest to lay their eggs in a random way. To
simulate the above process, three ideal states need to be set as
fn = −ka sin (u0 ) cos (w0 − wn ). (6) follows
where u and v present the random numbers of standard normal + S2 ∗ Rand ∗ xtr2,j − xti,j , (19)
distribution, β = 1.5. f is defined as follows
f −f
1/b S1 = exp i ri , (20)
G(1 + b) × sin (p × (b/2)) abs fi + 1
f= . (15)
G (1 + b)/2 × l × 2(b−1)/2
S2 = exp fr2 − fr1 , (21)
The cuckoo location update formula is as follows
where rand is a random number between [0, 1]. r1 represents the
index of the rooster (r1 ∈ [1, …, N ]] and r2 represents the index
f×u t (r2 ∈ [1, …, N]) of the hens in the same group, r1 ≠ r2.
xt+1 = xti + a0 (xi − xbest ). (16)
|v|1/b
i
Chicks can only forage around their mother hens, and this
situation can be formulated as follows
The locations and directions of the update are random. Hence, CS
can move the solution from one region to another, thereby
i,j = xi,j + FL ∗ xm,j − xi,j ,
xt+1 t t t
(22)
avoiding the fall into local optimum situation and enhancing the
early global search capability.
where xtm, j represents the ith hen. The adjusted parameter FL used in
3.2 Chicken swarm optimisation formula (22) represents the individual differences of the chicks.
The information of the solutions flows in a single direction, and
A new bio-inspired algorithm, called CSO, was proposed in 2014 for the entire optimisation process follows with the optimal solution.
optimisation problems by imitating the hierarchy and crowd In most cases, this algorithm can converge rapidly to the optimal
behaviour of chickens. This algorithm can process the optimisation solution.
problem effectively by extracting the chicken swarm intelligence.
The CSO algorithm is based on the following rules: 3.3 Cuckoo search–chicken swarm optimisation
(i) A number of groups exist in the chicken swarm, and each group 3.3.1 Population initialisation optimisation: In CS, the
contains a rooster, several hens, and chicks. Competitions occur population initialisation is random so that it cannot guarantee the
among different individuals. quality of each solution. The random initialisation of solutions
(ii) The best individual in a randomly selected group will be may cause the solutions be located far from the optimal one. If the
retained to the next generation. quality of the initial solution can improve, then the optimisation
(iii) The percentages of various levels of chicken swarms are fixed, efficiency must be raised. Thus, we can use the ergodicity of
and the number of the hens is the greatest. chaos theory to solve the problem of population initialisation.
(iv) The number of the available chickens n is fixed. The hierarchal Chaos is an irregular movement formed by the development of a
order, dominance relationship, and mother–child relationship in a deterministic system, which is similar to a random phenomenon.
group will remain unchanged. These statuses only can be updated Chaos contains a universal constant in a seemingly chaotic
every some (G) time step. motion. A certain initial value can be used to infer the new system
after the movement and at the same time traced back to the form of Using chaos theory to optimise the initial solution of the swarm
the past system. intelligent optimisation algorithm can cause the solutions to distribute
The logistic mapping model [18] is a widely used chaotic model in the swarm space evenly. Formula (24) can map the variable scope
that possesses a simple regression equation, better ergodicity than of chaos theory to the variable range of optimisation problems
other models, and more homogeneous solutions within the scope
of its distribution. Thus, we use the logistic mapping model as a
method and it can be defined as follows xi = Lb + (Ub − Lb)zi , (24)
The weight coefficient is beneficial to expanding the search space. 3.3.4 Iteration steps of CSCSO: The pseudo codes of CSCSO
When w is large, the global search capability is enhanced; on the are shown in Fig. 2.
contrary, the local search capability is enhanced. During the
optimisation process, the algorithm generates an optimal value in
each generation. Thus, we can establish a contact of the weight 4 Simulations and analysis
coefficient and the number of the same optimal values in several
generations. If the generations in which the same optimal value In this section, LAA and CAA optimisations using the proposed
appears are small, the global optimal solution may be located in CSCSO and other algorithms are simulated based on the Matlab
the near range and thus, the value of w is designed as a small platform. The CPU frequency is 2.40 GHz and the memory of the
value, thereby allowing the algorithm to search locally. By computer is 4GB. First, the main parameters of the CSCSO
contrast, if the generations in which the same optimal value algorithm are tuned by simulations. Second, the influences of
appears increase gradually, the number of w should increase improved components are presented. Finally, we have shown the
correspondingly. The search step length is then increased, and the performances of the beam pattern obtained by CSCSO and other
algorithm can jump out of the local optimum. To achieve the algorithms. Moreover, the convergence rate of these algorithms is
above conclusion, the weight coefficient can be defined as follows presented, and we have also shown the comparison results. In
addition, the performance of CSCSO for designing the real-world
w = a + (b − a) exp (m/G)0.3 . (26) antenna is verified by EM.
m records the times of generations in which the current optimal value 4.1 Parameter selection
appears; when the optimal solution appears for the first time, m is
defined to be 1. As for G, it should be set to an appropriate value, 4.1.1 Weighted levy flight: The following restriction is
which is problem-based. On the one hand, if the value of G is observed for the variable parameters: a ∈ [0, 2] and b ∈ [0, 2].
large, it will reduce the speed of global searching. On the other Each parameter is tested in a specific range with 50 repetitions to
hand, if the value of G is small, the algorithm may fall into the obtain the parameter value that provides the optimal solutions. For
local optimum. Meng et al. [15] presented that G∈[2, 20] may CSCSO, the results of different combinations of a and b are
achieve good results for most problems, so the value of G is set to shown in Fig. 3, and hence the parameter values are chosen as
be 5 in this paper. If the number of iterations increases gradually, a = 0.4, b = 0.9.
4.1.2 Population size: The population sizes are tested in a chaos are lower than that the values without chaos except the ninth
special range with 200 repetitions for LAA and 400 repetitions for and tenth tests for four elements LAA case, while it is the fourth,
CAA. The processes are repeated 20 runs and the average seventh and ninth tests in four elements CAA case. Similar
maximum SLL values are listed in Table 1. As can be seen from situations can be reflected in Fig. 5 for 16-elememt LAA and CAA.
the table that CSCSO cat obtains significantly lower maximum The maximum SLL obtained by random method in some tests can
SLL with the population size of 20 than that of 10 and 15. achieve better results because the random initial solutions may have
Meanwhile, the values of the average maximum SLL obtained by a certain probability to get the better location. However, the initial
the population size of 20 are similar with that of 25 and 30. More solution optimisation-based chaos theory has absolute advantages
population size means higher algorithm complexity. Thus, for most cases, which leads the proposed CSCSO algorithm can
considered the CPU time and other factors, the population size is obtain its optimal result faster compared with the other algorithms.
chosen as 20 for CSCSO in the simulation. Moreover, in order to
ensure the fairness of the comparisons, the population sizes of
other benchmark algorithms are also set to be 20. 4.2.2 Weighted Levy flight mechanism: Fig. 6 helps to show
the convergence rate of CSCSO algorithm for optimising the
16-elements LAA and CAA with and without a weighted Levy
4.2 Influences of the improved components flight mechanism, respectively. It can be seen that the weighted
Levy flight mechanism can improve the convergence speed of
4.2.1 Initial solution optimisation based on chaos: To CSCSO algorithm in both LAA and CAA cases because the steps
verify the influences of the chaotic based solution initialisation of the Levy flight can be adjusted dynamically according to the
mechanism, the values of initial solutions obtained with and situation in the current iteration. Moreover, 16 array elements
without chaos theory are simulated. The tests are conducted optimisation indicates that the number of individuals in each
independently for ten times for each case and the results are population is 16 and it can be seen as a higher dimensional
graphically illustrated in Figs. 4 and 5. It can be seen from Fig. 4 solution, which can reflect the performance of the weighted levy
that the initial solution values of the maximum SLL obtained by flight mechanism effectively.
4.3 Beam pattern optimisation paper. As shown in Fig. 7 and Table 3, all algorithms could obtain
the optimal results by about 50 iterations in this sample and the
In this section, LAA and CAA optimisations based on the CSO, CS, beam pattern in Fig. 7a are overlapped, this is because there are
PSO and proposed CSCSO are simulated and the comparison results only two excitation amplitudes that need to be optimised, and
are shown. For fair comparisons, all the common parameters of thus, the complexity of the optimisation problem is much lower.
above algorithms, such as population size, dimensions and However, the convergence rate for the proposed CSCSO algorithm
maximum number of iterations, are set to be the same. The other is much better than that of CSO, CS and PSO, which proves the
related parameters of these algorithms are listed in Table 2. effectiveness of CSCSO algorithm during the process of
4-elements LAA beam pattern optimisation.
4.3.1 LAA optimisation: We utilise the proposed CSCSO to C.16-element LAA: Similar to the previous examples, Table 4
optimise the excitation amplitude of each array element, and the shows the CSCSO results for a 16-element LAA, and Fig. 7c
maximum SLL performance of CSCSO is compared with that of shows a comparison among the radiation patterns using the
CSO, CS, PSO, and uniform excitation amplitude linear array different optimisation algorithms. The maximum SLLs optimised
(UECLA). The spacing between the array elements is normalised by CSCSO, CSO, CS, PSO, and UECLA are −32.85, −23.31,
to 0.5l, and the iterations are defined as 200. The excitation −28.33, −27.19, and −13.15 dB, respectively. Fig. 7d shows the
amplitude distribution in LAA is symmetrical; thus, only N/2 array convergence rate of the optimisation process obtained by the four
elements should be considered in the simulation. algorithms. The lowest initial value −16.6978 dB is produced by
A.4-element LAA: The simulation codes were run for 20 times CSCSO, whereas those obtained by other algorithms are
independently, and the optimal simulation results are shown in this −15.6846, −14.6970, and −15.6846 dB, respectively. It can be
Uniform [1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, −6.76
1.0000] [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000,
0.5000, 0.5000] = ∑ = 8.0
CSO [0.6283, 0.5153, 0.5885, 0.6135, 0.3944, 0.5312, 0.3976, 0.5942, 0.4054, 0.7303, 0.1892, 0.4667, 0.6719, 0.3248, 0.6847, −11.05
0.6695] [0.3792, 0.5135, 0.4455, 0.6047, 0.3966, 0.4091, 0.4055, 0.2905, 0.3147, 0.3449, 0.6142, 0.8092, 0.4595, 0.4168,
0.6446, 0.4187] = ∑ = 7.4672
CS [0.5963, 0.6605, 0.5163, 0.4530, 0.4038, 0.3714, 0.8692, 0.5959, 0.6087, 0.7265, 0.9777, 0.6850, 0.6729, 0.3736, 0.4460, −12.48
0.9879] [0.3931, 0.6398, 0.6200, 0.8352, 0.7933, 0.4801, 0.0621, 0.3328, 0.8855, 0.6517, 0.3317, 0.3871, 0.5363, 0.0023,
0.2148, 0.2197] = ∑ = 7.3855
PSO [0.5362, 0.2352, 0.2006, 0.6577, 0.4248, 0.2902, 0.8953, 0.8550, 0.7485, 0.5199, 0.4071, 0.8612, 0.0370, 0.5204, 0.3720, −11.91
0.6138] [0.9189, 0.2688, 1.0000, 0.8799, 0.7194, 0.2939, 0.2701, 0.5398, 0.9921, 0.4005, 0.0000, 0.9218, 0.3046, 0.1337,
0.4256, 0.0149] = ∑ = 8.084
CSCSO [0.6120, 0.6244, 0.9208, 0.3895, 0.7350, 0.4645, 0.7365, 0.6685, 0.9976, 0.8612, 0.5508, 0.2393, 0.7363, 0.5500, 0.8908, −15.55
0.8621] [0.3848, 0.3986, 0.8054, 0.7385, 0.7026, 0.5564, 0.3781, 0.4399, 0.4045, 0.4621, 0.8968, 0.7638, 0.7115, 0.8049,
0.4665, 0.2587] = ∑ = 9.1731
(UECCA). The iteration is set to 400 because more parameters need and CS have constant convergence curves, the accuracy of them is
to be optimised with CAA than with LAA, while the number of much worse than CSCSO. These phenomena indicate that CSCSO
elements is equal. has better accuracy for CAA optimising because the worst solution
A.4-element CAA: Fig. 8a shows the beam pattern of 4-element can become suboptimal one by using different position update
CAA obtained by the four algorithms and the results are compared formulas in hierarchical mechanism of CSCSO, and thereby
with UECCA. The excitation amplitude and spacing between the improving the precision of results effectively. Moreover, by using
array elements are recorded in Table 5. The table also shows the the non-linear variation weight coefficient for normal Levy flight,
maximum SLL obtained by these algorithms. The convergence the convergence rate of the proposed CSCSO algorithm can be
curves of the CSO, CS, PSO, and CSCSO are shown in Fig. 8b enhanced. The excitation amplitude and spacing between the array
with the fixed mainlobe width, i.e. 80°. Similar to 4-element LAA elements are recorded in Table 6. Fig. 8c shows that the beam
case, the performance of the proposed CSCSO algorithm for pattern obtained by CSCSO has the lowest maximum SLL when
optimising 4-element CAA is not so better than that of CSO, CS, the mainlobe width (40°) is fixed.
and PSO because the number of array elements is less. However, It can be seen from the preceding examples that the maximum
different from the LAA case, CAA needs to optimise the spacing SLL of the beam patterns obtained by CSCSO are lowest
between the elements so that enhancing the complexity of the significantly compared with other algorithms in each case.
problem, and this means the increasing of the dimensions of the Moreover, CSCSO is more prominent when the number of array
solution. Thus, the accuracy and the convergence rate of CSCSO elements is big. In addition, CSCSO has better values of initial
are better than other algorithms and this can be reflected from the solutions, faster convergence rate, and higher accuracy of the beam
figure. pattern optimisations. Hence, CSCSO has more advantages on the
C.16-element CAA: As mentioned above, there is an antenna array optimisation problem than other mentioned
inconspicuous advantage of CSCSO compared with benchmark algorithms.
algorithms under 4-element CAA case. However, once the number
of array elements increased to 16, the accuracy gaps and the
convergence rates among these algorithms become apparently. 4.4 EM simulation
Fig. 8d shows that the CSO obtained the reduced maximum SLL
of −11.05 dB in the 115th iteration, and then it got into local To verify the performance of the real-world antenna array, we
optimum. However, the proposed CSCSO algorithm continued designed a 4-element LAA for EM simulations based on HFSS
exploring for better solution and finally settled at a maximum SLL 13.0. Fig. 9a shows the physical structure and parameters of the
of −15.55 dB by the 400th iteration, which is much lower than antenna and Fig. 9b shows the beam pattern (also in polar
that of −12.48 dB of CS and −11.91 dB of PSO. Although PSO coordinate form) of the antennas with different excitation