Journal of Electromagnetic Analysis and Applications, 2011, 3, 213-219 doi:10.4236/jemaa.2011.

36035
Published Online June 2011 (http://www.SciRP.org/journal/jemaa) Copyright 2011 SciRes. JEMAA 213

On

Chebyshev Array Design Using Particle Swarm Optimization

Mohammed H. Bataineh
Department of Electrical Engineering, Faculty of Engineering, Qassim University, Qassim, Saudi Arabia. Email: [email protected] Received March 10th, 2011; revised March 20th, 2011; accepted April 12th, 2011.

ABSTRACT
In this paper, the evolutionary algorithm of particle swarm optimization (PSO) is applied to synthesis an optimal linear array in the Chebychev sense. Equiripple radiation patterns may be obtained by synthesizing the excitation currents feeding the array or by carefully choosing the interelement spacing. The desired equal side lobes level is achieved si-multaneously with the narrowest possible beamwidth (high directivity). Though the optimization problem may become nonlinear, convex and/or sometimes nonconvex, it can be handled using an efficient, a robust and a nongradient based particle swarm optimizer algorithm. In order to effectively utilize this algorithm it is important to define an appropriate objective, or cost, function that return a single number to enable the PSO algorithm minimizing it. In this paper, the objective function is formulated taking into consideration the level of the side lobes as well as the main beam width. In addition to satisfy the objective function, the obtained results using the proposed technique are in agreement with those available in the literature. Keywords: Optimization, PSO, Linear Antenna Arrays, Chebyshev Radiation Pattern

1. Introduction
When antennas elements are arrayed in a certain geomet-rical configuration the signal induced on them are com-bined to form the array output. A plot of the array re-sponse as a function of angle is normally referred to as the array pattern or beam pattern. Synthesizing the array pattern of antenna arrays has been a subject to several studies and investigations. The trade-off between the side-lobe levels (SLL) and the half-power beamwidth (HPBW) stimulate the question answered primarily by Dolph [1] of obtaining the narrowest possible beamwidth for a given side-lobe level or the smallest side-lobe level for a given beamwidth. This was possible by using the orthogonal functions of Chebyshev [2] in order to design an optimum radiation pattern. However, for large number of elements this procedure becomes quite cumbersome since it requires matching the array factor expression with an appropriate Chebyshev function [3]. To over-come this deficiency, Safaai-Jazi [4] proposed a new formulation for the design of Chebyshev arrays based on solving a system of linear equations. Iterative procedure was used to produce the desired pattern [5]. Formulating the synthesis problem as a convex optimization, which may be solved by interior-point methods, has been presented in [6]. Shpak et al. [7,8] discussed an improved method for the design of linear arrays with prescribed nulls. The forgoing mentioned investigations either re-quire analytical formulae or evaluating the gradient of some cost function, which sometimes become formidable to evaluate. As alternatives, neural network and evolu-tionary algorithms techniques were used in order to re-duce the side lobes of linear arrays [9,10]. Genetic algo-rithms (GA) and particle swarm optimization are well-known evolutionary algorithm techniques. In GA, a sample of possible solutions is assumed then mutation, crossover, and selection are employed based on the con-cept of survival of the fittest [10]. On the other hand, PSO is a much easier algorithm in which each possible solution is represented as a particle in the swarm with a certain position and velocity vector [11,12]. The position and velocity of each particle are updated according to some fitness function [11]. Some studies have been de-voted to compare between the GA and PSO [20,21] and a general conclusion has been reached the PSO shows bet-ter performance due to its greater implementation sim-plicity and minor computational time. Since it has been introduced by Kennedy and Eberhard [13], the PSO is being applied to many fields of endeavor. On Chebyshev Array Design Using Particle Swarm Optimization 214

Surprisingly, it has been applied to the design of low dispersion fiber Bragg gratings [14], and to the design of corrugated horn antenna [15]. Other applications can be found in [16]. This investigation is devoted to the design of Chebyshev linear antenna arrays by considering vari-ous affecting parameters using the PSO. The paper is organized as follows. In Section II, the particle swarm optimization algorithm is overviewed. A background about Chebyshev polynomials is addressed in Section III. Formulation of linear array pattern synthesis is presented in Section IV. Simulation examples and results are given in Section V. Conclusions drawn and hints to further investigation are pointed out in Section VI.

2. Particle Swarm Optimization (PSO) Algorithm

PSO is a stochastic optimization technique that has been effectively used to solve multidimensional discontinuous optimization problems in a variety of fields [17]. The stochastic behavior of this technique can be controlled by one single factor which can be chosen to end up with a deterministic strategy that does not need gradient infor-mation. Compared with other evolutionary 9 techniques, the PSO is much simpler and easier to implement with guaranteed convergence [18]. The concept of the PSO is derived from an analogy of the social behavior of a swarm of bees in the field, for example. Without any prior knowledge, each bee re-ferred to as agent or particle in the PSO jargon, in the swarm starts with random position and velocity with an aim to find the location of highest density of flowers in the field. Then, during their search, each bee updates its velocity and position based on two pieces of information. The first is its ability to remember the location of most flowers it personally found (particle best will be denoted later as pbest). The second represents the location of most flowers found by all the bees of the swarm (global best will be denoted later as gbest) at the present instant of time. This process of updating velocities and positions continues and will result in one of the bees would find a location with a highest density of flowers in the field. Eventually, all the bees (solutions) will be drawn to this location since they will not be able to find any other bet-ter location. This represent a convergence of the algo-rithm and the optimum solution is obtained [11,13,17, 20-24]. The main steps of the PSO algorithm are given below and will be elaborated in the remaining portion of this section. 1) Definition of the solution space: In this step, the minimum and maximum value for each dimension in the N-dimensional optimization problem are specified to define the solution space of the problem. 2) Definition of the fitness function: The fitness func-tion is a problem dependent measure of the goodness of a position (N-dimensional vector) that represents a valid solution of the problem. It should be carefully selected to represent the goodness of the solution and return a single number. 3) Random initialization of the swarm positions and velocities: The positions and velocities of the particles are randomly initialized. To help the discussion let us refer to these matrices as X and V, respectively. Both matrices have a dimension of M N. Here, M represents the number of particles or swarm size and N represents the dimension of the optimization problem. However, it is preferred to randomly initialize the X matrix within the solution space for faster convergence. To complete the initialization step, each row in the matrix X is labeled as the individual best position for each particle (pbest). In addition, the positions in X are plugged into the fitness function and the returning numbers are compared, the position with the best returning number is labeled as the global best position (gbest). In this context, the word best could mean highest or lowest depending on the optimiza-tion problem. The problem considered here is a minimi-zation of the fitness function. 4) Update of velocity and position: The following sub-steps are carried out for each particle individually: a) Velocity update: The velocity of the mth particle in the nth dimension is updated according to the following equation
1111122tttttmnmnnmnmntttnmnmnvwvcpbestxcgbestx (1) The superscripts t and t1 denote the time index of the current and the previous iterations, n1 and n2 are two different, uniformly distributed random numbers in the interval [0,1]. The relative weights of the personal best position versus the global best position are specified by the parameters c1 and c2, respectively. Both c1 and c2 are typically set to a value of 2.0 [19]. The parameter w is called the inertial weight, and it is a number in the range [0,1] that specifies the weight by which the parti-cles current velocity depends on its previous velocity, and the distance between the particles position and its personal best and global best positions. Empirical studies [11] have shown that the PSO algorithm converges faster if w is linearly damped with iterations, for example starting at 0.9 at the first iteration and finishing at 0.4 in the last iteration. b) Position update: In this step, the position of the mth particle in the nth dimension is updated according to the following equation

1tttmnmnmn

xxtv (2)

Copyright 2011 SciRes. JEMAA On Chebyshev Array Design Using Particle Swarm Optimization 215

where represents a given time step (usually chosen to be one). t c) Fitness evaluation: The updated N-dimension po-sition in the previous step is plugged in the fitness function and the returning number is compared with that cor-responding to the pbest, if the returning number is better, this updated N-dimension position is labeled as the new pbest. In addition, if the returning number is better than that corresponding to gbest, the updated N-dimension position is also labeled as the new gbest. 5) Checking the termination criterion: In this step, the algorithm may be terminated if the number of iteration equals a pre-specified maximum number of iteration or the returning number corresponding to gbest is close enough to a desired number. If none of the above condi-tions is satisfied, the process is repeated starting at step 4.

3. Chebyshev Polynomials
One of the most eminent methods used to equate the sidelobes arising in the radiation pattern of antenna ar-rays is to utilize a set of polynomials referred to as Che-byshev polynomials, after the Russian mathematician Pafnuti Chebyshev (1821-1894) [2]. These polynomials are originated as possible solutions of a second order ordinary differential equation with variable coefficients. This eigenvalue equation may take the form
210xyxyy (3) with an additional requirements that y(1), (1), y(1), and (1) are to be bounded. A change of variables according to yy cosx (4) transforms Equation (3) to 22d0dTT (5) where cosTy. The transformation used in Equation (4) results in a second order linear ordinary differential equation with constant coefficients as given by Equation (5). Solutions of Equation (5) depend on the eigenvalue and could be written as cossin00ABTCD (6) where A, B, C, and D are arbitrary constants. The solu-tion would be bounded if B = D = 0 and n with n being an integer. The eigenfunctions, corresponding to nontrivial solutions are given by cos,0,1,2,nnn (7) In terms of the original independent variable x, the ei-genvalues and eigenfunctions are given by

coscos1coshcosh 1nnnnTxnxxTxnxx (8) Here, the dependent variable T is in honor of Cheby-shev (often spelled Tchebysceff). Using the Euler iden-tity and substituting for x from Equation (4), the argu-ment cosn could be expanded in terms of its funda-mental argument, cosn. A few examples are given below, for different orders, for the purpose of illustration 012233244()1()()12()34()188TxTxxTxxTxxxTxxx (9) A general recursion formula may be deduced from Equations (9) and is given by 112()2()()() nnnTxTxTxTx (10) Chebyshev polynomials constitute a set of orthogonal functions with respect to a weighting function, . The weighting function could be found be recasting Equation (7) in standard Sturm-Liouville form. Carrying out this step to find [2] ()wx 21()1wxx (11) With respect to this weighting function, Chebyshev polynomials are orthogonal, i.e., they satisfy
211 11()()()d0, mnwxTxTxxmn (12) These polynomials oscillate with unit amplitude in the interval 1x and become monotonically increasing or decreasing, depending on their order, outside this range. This property of Chebyshev polynomials enabled Dolph to use them to design an equiripple radiation patterns.

4. Linear Array Pattern Synthesis

The focus here will be on an N-element array distributed on a line with equal interelement spacing, d. The array factor may be written as
10Nnnn

AFI (13)

Here, expzj and coskd. The an-gle measured from the line of the array, the is
Copyright 2011 SciRes. JEMAA On Chebyshev Array Design Using Particle Swarm Optimization 216

wavenumber k = 2/; being the wavelength, and is the difference in phase excitation between the elements. In is the magnitude of the current of the nth element. For symmetrical current distribution Equation (13) may be written as [3] 2cos12NnnAFIn (14) for even number of elements. In the case of odd number of elements, the array factor could be written as [3]
/21

12012cosNnnAFIIn (15) In Equation (15), I0 denotes the current of the center element. The case of unequal interelement spacing will be considered also. This requires a straightforward modi-fication in Equations (14) and (15), however. The corre-spondence between an N-element array factor (Equation (14) and Equation (15)) and a Chebyshev polynomial of order (N 1) is carried out to match the coefficients of similar terms thus giving the required current excitations nI. It is clear that this procedure requires elaborate computation especially for large number of elements. The problem is circumvented here as an optimization one. Starting from an arbitrary, or random, current excitation, the PSO algorithm is used in order to return current ex-citation that leads to an equiripple pattern with the nar-rowest possible beamwidth. The procedure of using the PSO to obtain the required current excitations nI may be summarized in the following steps 1) Specify N, d, and .
2) Start with an arbitrary current excitation 0nI. 3) Obtain the array factor AF () using either Equa-tion (14) or (15). 4) The side-lobe levels maxiand their locations AFmaxi 1,iare determined to be used in the next step; max2,, . Here, max denotes the number of peaks appear in side-lobe regions. 5) The side-lobe levels maxi are forced not to exceed a certain level, say , and this is achieved by minimizing the following cost, or fitness, or error, or objective function AF
maxmax21iiAF (16) 6) Return to step 2 while the number of iterations or minimum error criteria is not attained, otherwise stop. Indeed the process described above gave equiripple ra-diation pattern but not necessarily the narrowest possible beamwidth. Thus the cost function should be amended to incorporate a condition on the beamwidth also. Equation (16) is therefore modified to read

maxmax221iiiAFBWcB (17) The positive weighting factors i and are added to give each factor a certain influence on the obtained re-sults. The numerically calculated beamwidth (BWc) is compared with the desired beamwidth (BWd), which may be determined by an analytic formula (e.g. [4]). Equation (17) satisfies the requirement of the PSO algo-rithm that the fitness function should return a single number representing the target of the minimization proc-ess. The optimization here is carried out using a fixed spacing, which is between /2, the elements. However, this spacing is not an optimum spacing. Looking for such an optimum spacing is also undertaken. Using an opti-mum spacing, which is usually greater than /2, extends the visible space. Hence, more sidelobes and narrower patterns are expected. To further demonstrate the capa-bilities of the PSO algorithm, the locations of the array elements are also considered as the optimized parameters. Though the problem becomes nonlinear, the PSO re-turned the required spacings that lead to equiriple pattern. These points will be elaborated by considering specific examples in the following Section.
5. Illustrative Examples
In this section, the capabilities of the PSO algorithm in the synthesis of antennas array are demonstrated by three examples. First, the PSO algorithm was used to find the current excitations that result in equiripple array factor. In this example, the interelement spacing was set to /2 and all excitation currents are assumed to be equiphase. Figure 1 shows the radiation pattern of a 10-element array using the excitation currents returned by the PSO with fitness function as defined in Equation (16) (dotted line) and in Equation (17) (solid line). In the former case, the goal was to obtain an equiripple array with side lobes level of 20 dB. Although the side lobes are at the de-sired level, the obtained array is not optimum in a Che-byshev sense. To overcome this problem, the fitness function was modified as given in Equation (17) to achieve the narrowest possible main lobe besides ob-taining an equiripple pattern. With this modification of the fitness function, the PSO returned excitation currents that resulted in the desired narrow main lobe and equal side lobe level. The returned current excitations of this case along with various other examples using

different array sizes are given in Table 1. These results are in ex-cellent agreement with those presented elsewhere (e.g. [4]). In the second example, the PSO was used to find more directive patterns than that obtained by limiting the spacing between the elements to /2. Such optimum spacing undoubtedly extends the visible space and results
Copyright 2011 SciRes. JEMAA On Chebyshev Array Design Using Particle Swarm Optimization 217

Figure 1. Radiation pattern of 10-element antennas array synthesized using PSO using two different fitness functions. Table 1. Magnitude of Current Excitation of /2-spacing for antenna array synthesis using PSO.
Number of Elements Side Lobes Level (SLL) Magnitude of Current Excitations (normalized with respect to the edge element) 1.7357, 1.0

20 dB 40 dB 20 dB 40 dB
2.6687, 1.0

10

20 dB 40 dB

1.8388,1.6827,1.27 63,1.0 6.1224,5.1492,2.96 32,1.0 1.5568,1.4383,1.21 26,0.9295, 1.0 8.0551,6.7374,4.66 08, 2.5283, 1.0