chapter 4 linear antenna array synthesis … · models its processes after the principles of...
TRANSCRIPT
76
CHAPTER 4
LINEAR ANTENNA ARRAY SYNTHESIS USING
GENETIC ALGORITHM
4.1 INTRODUCTION
Improving antenna characteristics such as directivity, input
impedance, bandwidth, sidelobe level, null depth level and size has always
been the goal of antenna researchers. It is important for the antenna
researchers to have the latest tools to effectively design antennas that meet the
given specifications. Optimization techniques are used to either synthesize an
antenna from given radiation characteristics or simply improve existing
antenna designs. Search routines that utilize numerical methods to provide
radiation properties of antenna generally consume a considerable amount of
time. Therefore, a great deal of work has been devoted to achieving
optimization routines that rapidly and accurately search out an optimum
solution.
Recently, a unique optimization scheme based on genetic
algorithms (GA) (Teruel and Rajo-Iglesias 2006, Stephen 2008, Goldberg and
Holland 1998) has been used to solve a number of electromagnetic problems.
Genetic algorithms, the adaptive heuristic search algorithms, are
programming techniques that mimic biological evolution as a problem-
solving strategy, based on the evolutionary ideas of natural selection and
genetic. Genetic algorithms tend to evolve in an environment in which there is
77
a very large set of candidate solutions and in which the search space is uneven
and has many hills and valleys. Genetic algorithms are useful and efficient
when
The search space is large , complex and poorly understood
Domain knowledge is scarce or expert knowledge is difficult to
encode to narrow the search space
No mathematical analysis is available
Traditional search methods fail
The genetic algorithm is a robust, stochastic search method that
models its processes after the principles of natural selection and evolution.
These genetic algorithms are very useful for finding optimum antenna designs
that maximize or minimize certain radiation properties. The GA provides
optimal solutions by successively creating populations that improve over
many generations. Selecting, mating and mutating the previous population
each creates a new generation. This process continues until the population
converges to a single optimal solution. The selection process is based on the
rating of each member relative to the population. This rating is done by fitness
testing of each individual. The fitness value can be the gain, axial ratio, input
impedance, size, sidelobe level or any combination thereof. This fitness value
is received from some type of numerical code that can provide the radiation
characteristics of the antenna. A MATLAB code for SLL reduction and null
depth level is used in this thesis as a fitness function from which the fitness
value is obtained.
This chapter discusses the deployment of Genetic algorithm
optimization method for synthesis of antenna array radiation pattern in
adaptive beamforming. The synthesis problem discussed is to find the weights
78
of the Uniform Linear Antenna array elements that are optimum to provide
the radiation pattern with maximum reduction in the sidelobe level (Basak
et al 2008).
4.2 GENETIC ALGORITHM AS AN OPTIMIZATION TOOL
Holland performed much of the foundational work in Genetic
Algorithm during 1960-1970. His goal of understanding the processes of
natural adaptation and designing biologically-inspired artificial systems led to
the formulation of the simple genetic algorithm.
Since its conception, genetic algorithms have enjoyed global use by
many researchers and scientists in many different areas. Although computer
scientists can take much of the credit for the development of GA, areas such
as business, science, and engineering have put the GA to good use.
Researchers, who historically have been obsessed with better and cheaper,
find particular interest in the GA. In Electromagnetics, most all problems are
of a non-differentiable type. Some type of numerical solution is required to
calculate the important characteristics of the problem. When the optimization
is a goal, these problems lend themselves very well to the use of genetic
algorithms. Antenna design is an area of Electromagnetics that has recently
benefited from the use of Gas. Haupt (1994,1995, 2004, 2008), recently
compiled a great deal of work on antenna design using genetic algorithms.
Array thinning and array synthesis are two areas where genetic algorithms
have proven their usefulness. In both of these cases, the GA has provided
results that exceed previous attempts to improve the array designs.
Genetic algorithms are typically implemented using computer
experiments in which an optimization problem is specified. For this problem,
members of a space of candidate solutions, called individuals, are represented
79
using abstract representations called chromosomes. GA consists of an
iterative process that evolves a working set of individuals called a population
toward an objective function, or fitness function. Traditionally, solutions are
represented using fixed length strings, especially binary strings, but
alternative encodings have been developed.
The key terms related to genetic algorithm are
Individual – any possible solution
Population – group of all individuals
Search space – all possible solutions to the problem
Chromosome – blueprint for an individual
Trait – possible aspect of an individual
Allele – possible settings of a trait
Locus – the position of a gene on a chromosome
Genome – collection of all chromosomes for an individual
The evolutionary process of a GA is a highly simplified and
stylized simulation of biological version. It starts from a population of
individuals randomly generated according to some probability distribution,
usually uniform and updates this population in steps called generations. In
each generation, multiple individuals are randomly selected from the current
population based upon some application of fitness, bred using crossover, and
modified through mutation to form a new population. The genetic algorithm
operators are defined below
Crossover – exchange of genetic material (substrings) denoting
rules, structural components, features of a machine learning,
search, or optimization problem
80
Selection – the application of the fitness criterion to choose
which individuals from a population will go on to reproduce
Reproduction – the propagation of individuals from one
generation to the next
Mutation – the modification of chromosomes for single
individuals
Current GA theory consists of two main approaches – Markov
chain analysis and schema theory. Markov chain analysis is primarily
concerned with characterizing the stochastic dynamics of a GA system, i.e.,
the behavior of the random sampling mechanism of a GA over time. The most
severe limitation of this approach is that while crossover is easy to implement,
its dynamics are difficult to describe mathematically. Markov chain analysis
of simple GAs has therefore been more successful at capturing the behavior of
evolutionary algorithms with selection and mutation only. These include
evolutionary algorithms (EAs) and evolutionary strategies. A schema is a
generalized description or a conceptual system for understanding knowledge-
how knowledge is represented and how it is used. According to this theory,
schemata represent knowledge about concepts: objects and the relationships
they have with other objects, situations, events, sequences of events, actions,
and sequences of actions.
4.2.1 Flow Chart for Simple Genetic Algorithm
The flow of simple Genetic algorithm can be explained through the
flowchart representation as shown in Figure 4.1.
81
Initial population (firstgeneration) / best fit and
new population ofremaining generations
Fitness on costfunction evaluationsfor each generation
Selection of best fitindividuals/ rejectionof least fit population
Breed using geneticoperators crossover
and mutation
Generate newoffspring
Figure 4.1 GA reproduction cycle
The cycle is repeated until a termination condition has been reached
such as
1. A solution that satisfies the minimum criteria.
2. Reaching the specified number of generations.
3. Reaching the specified Computation time.
4. Arriving at fitness value and
5. Manual inspection.
82
Traditionally, solutions are represented using fixed length strings,
especially binary strings, but alternative encodings have also been developed.
The genetic algorithms provide a directed random search in
complex landscapes. In nature, all living organisms have certain the physical
characteristics or traits, known as the phenome. The phenomes are encoded
into a set of genetic data structure, termed as genome. A particular set of
genetic information is a genotype, and likewise a particular set of physical
characteristics, or traits, is a phenotype. There may, or may not, be a direct
one-to-one mapping of genotypes to phenotypes. These physical
characteristics determine how well suited to its environment a particular
organism is. The suitability of a given organism to its environment is usually
measured as its fitness. Computationally, it is usual to evaluate the fitness of
an organism directly, without considering any kind of phenome.
4.2.2 Encoding
In GAs encoding the solutions of the problem into chromosomes is
a key issue. John Holland, used a single binary bit string. The problem
associated with encoding is that some individuals may correspond to
infeasible or illegal solutions. This may become very severe for constrained
optimization problems and combinatorial optimization problems. Infeasible
solution is one, that chromosome lies outside the feasible region of given
problem. Penalty methods can be used handle infeasible chromosomes. Illegal
solutions have chromosomes that do not represent a solution. Repair
techniques are usually adopted to convert an illegal chromosome to legal one.
Many different methods for encoding the genetic information are in common
use today; Tree encoding, real-valued arrays, permutations, Gray encoding
and so on.
83
In general encoding methods can be classified as follows:
i) Binary encoding
ii) Real-number encoding
iii) Integer or literal permutation encoding
4.2.2.1 Binary encoding
Binary encoding (i.e., the bit strings) are the most common
encoding used. Much of existing GAs theories is based on the assumption of
using binary encoding. The binary code doses not preserve the locality of
points in the phenotype space.
4.2.2.2 Real number encoding
Real number encoding performs better than binary encoding for
function optimization and constrained optimizations problems. In real number
encoding, the structure of genotype space is identical to that of the phenotype.
Therefore, it is easy to form effective genetic operators by borrowing useful
techniques from conventional methods.
4.2.2.3 Permutation encoding
Permutation encoding is best used for combinational optimization
problems because it is useful in searching for the best permutation or
combination of items subject to constrain. The encoding schemes can be
better understood while dealing with genetic operators such as cross over and
mutation.
4.3 GENETIC ALGORITHM OPERATORS
There are two basic genetic algorithms operators which are
crossover and mutation. These operators perform a blind search by working
84
together to explore – i.e. to investigate new and unknown areas in search
space and exploit- to make use of knowledge of solutions previously found in
search space to help in find better solutions, by creating new variants in the
chromosomes. It is confirmed that mutation operator play the same important
role as that of the crossover
4.3.1 Crossover or Recombination
Crossover plays important role in the design and implementation of
robust evolutionary systems. In most GAs, individuals are represented by
fixed-length strings and crossover operates on pairs of individuals (parents) to
produce new strings (offspring) by exchanging segments from the parents’
strings. Traditionally, the number of crossover points (which determines how
many segments are exchanged) has been fixed at a very low constant value of
1 or 2.
4.3.1.1 Single point crossover
A commonly used method for crossover is called single point
crossover which is shown in Figure 4.2. In this method, a single point
crossover position (called cutpoint) is chosen at random and the parts of two
parents after the crossover position are exchanged to form two offspring
Figure 4.2 Single point crossover
Parent2
Parent1 Offspring1
Offspring2
85
4.3.1.2 Multi point crossover
Multi-point crossover is a generalization of single point crossover,
introducing a higher number of cut-points. In this case multi positions are
chosen at random and the segments between them are exchanged as shown in
Figure 4.3.
Figure 4.3 Multi point crossover
4.3.1.3 Uniform crossover
Uniform crossover does not use cut-points, but simply uses a global
parameter to indicate the likelihood that each variable should be exchanged
between two parents as shown in Figure 4.4.
[
Figure 4.4 Uniform crossover
Parent1 Offspring1
Parent2 Offspring2
Parent 1 Offspring 1
Parent 2 Offspring 2
86
4.3.2 Mutation
Mutation is a common operator used to help preserve diversity in
the population by finding new points in the search pace to evaluate. When a
chromosome is chosen for mutation, a random change is made to the values of
some locations in the chromosome. A commonly used method for mutation is
called single point mutation. Though, a special mutation types used for varies
problem kinds and encoding methods.
4.3.2.1 Single point mutation
Single gene (chromosome or even individual) is randomly selected
to be mutated and its value is changed depending on the encoding type used,
as shown in Figure 4.5.
[[Figure 4.5 Single point mutation
4.3.2.2 Multi point mutation
Multi genes (chromosomes or even individuals) are randomly
selected to be mutated and there values are changed depending on the
encoding type used which is shown in Figure 4.6.
Figure 4.6 Multi point mutation
OffspringParent
Parent Offspring
87
4.3.3 Selection
Selection is the process of determining the number of times a
particular individual is chosen for reproduction and, thus, the number of
offspring that an individual will produce. Selection provides the driving force
in genetic algorithms. With too much force, genetic search will terminate
prematurely. While with too little force, evolutionary progress will be slower
than necessary. Typically, a lower selection pressure is indicated at the start of
genetic search in favor of a wide exploration of the search space, while a
higher selection pressure is recommended at the end to narrow the search
space. In this way, the selection directs the genetic search toward promising
regions in the search space and that will improve the performance of genetic
algorithms. Many selection methods have been proposed, examined and
compared. The most common types are
1) Roulette wheel selection
2) Rank selection
3) Tournament selection
4) Steady state selection
5) Elitism
4.3.3.1 Roulette wheel selection
Roulette wheel selection is most common selection method used in
genetic algorithms for selecting potentially useful individuals (solutions) for
crossover and mutation. In roulette wheel selection which is shown in
Figure 4.7, as in all selection methods, possible solutions are assigned fitness
by the fitness function. This fitness level is used to associate a probability of
selection with each individual. While candidate solutions with a higher fitness
will be less likely to be eliminated, there is still a chance that they may be.
With roulette wheel selection there is a chance some weaker solutions may
88
survive the selection process; this is an advantage, as though a solution may
be weak, it may include some component which could prove useful following
the recombination process. The analogy to a roulette wheel can be envisaged
by imagining a roulette wheel in which each candidate solution represents a
pocket on the wheel; the size of the pockets is proportionate to the probability
of selection of the solution. Selecting N individual from the population is
equivalent to playing N games on the roulette wheel, as each candidate is
drawn independently.
Figure 4.7 Roulette wheel selection
4.3.3.2 Rank selection
In ranking selection, as shown in Figure 4.8, the individuals in the
population are sorted from best to worst according to their fitness values.
Each individual in the population is assigned a numerical rank based on
fitness, and selection is based on this ranking rather than differences in
fitness. The advantage of this method is that it can prevent very fit individuals
from gaining dominance early at the expense of less fit ones, which would
reduce the population's genetic diversity and might hinder attempts to find an
acceptable solution. The disadvantage of this method is that it required sorting
the entire population by rank which is a potentially time consuming
procedure.
89
(a)
(b)
Figure 4.8 Rank selection effect (a) before ranking (b) after ranking
4.3.3.3 Tournament selection
This method randomly chooses a set of individuals and picks out
the best individual for reproduction. The number of individual in the set is
called the tournament size. A common tournament size is 2, this is called
binary tournament. By adjusting tournament size, the selection pressure can
be made arbitrarily large or small. For example, using large Tournament size
has the effect of increasing the selection pressure, since below average
individuals are less likely to win a tournament while above average
individuals are more likely to win it.
4.3.3.4 Steady state selection
The steady state selection will eliminate the worst of individuals in
each generation. It works as follows; the offspring of the individuals selected
90
from each generation go back into the pre-existing population, replacing some
of the less fit members of the previous generation.
4.3.3.5 Elitism
Elitism is an addition to many selection methods that force genetic
algorithms to retain some number of the best individual at each generation. It
improves the selection process and save the best individuals. With elitist
selection, the quality of the best solution in each generation monotonically
increases over time. Without elitist selection, it is possible to lose the best
individuals due to stochastic errors
4.4 GENETIC ALGORITHMS PARAMETERS
One of the more challenging aspects of using genetic algorithms is
to choose the configuration parameter settings. Discussion of GA theory
provides little guidance for proper selection of the settings. The population
size, the mutation rate, and the type of recombination have the largest effect
on search performance. They are used to control the run of a GA. They can
influence the Population and the Reproduction part of the GAs. In traditional
GAs the parameters has fixed values. Some guidelines are used in selecting
these parameter settings are given in the following subsections.
4.4.1 Population Size
The population size is one of the most important parameters that
play a significant role in the performance of the genetic algorithms. The
population size dictates the number of individuals in the population. Larger
population sizes increase the amount of variation present in the initial
population at the expense of requiring more fitness evaluations. It is found
that the best population size is both applications dependent and related to the
91
individual size (number of chromosomes within). A good population of
individuals contains a diverse selection of potential building blocks resulting
in better exploration.
If the population loses diversity the population is said to have
“premature convergence” and little exploration is being done. For larger
individuals and challenging optimization problems, larger population sizes are
needed to maintain diversity (higher diversity can also be achieved through
higher mutation rates and uniform crossover) and hence better exploration.
Many researchers suggest population sizes between 25 16 and 100 individual,
while others suggest that it must be very much larger.
4.4.2 Crossover Rate
Crossover rate determines the probability that crossover will occur.
The crossover will generate new individuals in the population by combining
parts of existing individuals. The crossover rate is usually high and
‘application dependent’. Many researchers suggest crossover rate to be
between 0.6 and 0.95.
4.4.3 Mutation Rate
Mutation rate determines the probability that a mutation will occur.
Mutation is employed to give new information to the population (uncover
new chromosomes) and also prevents the population from becoming saturated
with similar chromosomes, simply said to avoid premature convergence.
Large mutation rates increase the probability that good schemata will be
destroyed, but increase population diversity. The best mutation rate is
‘application dependent’. For most applications, mutation rate is between
0.001 and 0.1 while for automated circuit design problems, it is usually
between 0.3 and 0.8.
92
Genetic algorithms work on two types of spaces alternatively:
coding space and solution space, or in other words, genotype space and
phenotype space. Genetic operators (crossover and mutation) work on
genotype space, while evolution and selection work on phenotype space. The
selection is the link between chromosomes and the performance of decoded
solutions. The mapping from genotype space to phenotype space has a
considerable influence on the performance of genetic algorithms.
4.5 GENETIC ALGORITHM – A TOOL FOR SYNTHESIS OF
UNIFORM LINEAR ANTENNA ARRAY
In this thesis, two different approaches have been adopted to get a
deeper insight into the application of GA for the synthesis of linear antenna
array,
(1) Using Simple GA
(2) Using GA optimization toolbox in MATLAB
In general, larger population sizes tend to converge to an optimum
solution with fewer generations. This is because the solution space is
sufficiently covered with a larger population size. This luxury may not be
affordable if the fitness function is too time consuming. Also, it may be
redundant to use such a thorough initial search space. A tradeoff occurs
between population size and the time needed to converge to a solution. A
small population converges quickly, but may not find a global optimum. A
large population converges slowly, but more confidently finds the global
optimum. Judicious choices of gene and population size are important, and
dependent on the particular problem. Gene length (number of alleles or bits)
is set relative to how accurate an answer is needed. Larger gene length allows
more quantization levels. The gene length affects the accuracy of the solution.
93
It is generally accepted to choose a population size that is 2 to
3 times larger than the total chromosome length (total number of bits). For
instance a chromosome with 4 genes (parameters) that each contains 10
alleles (bits) should have a population of 80 – 120. Larger populations require
more simulations per generation, but in general require fewer generations to
find a solution. This is merely because the initial search space is so large. This
may not be necessary for simple problems. However, more complex
problems, the majority of the initial population may be unacceptable and the
next generation is filled with the very few members that survived the selection
process. In fact, sometimes the initial population is filled with no acceptable
solutions at all. This is remedied by allowing a larger initial population and
thus a more thorough initial search.. Elitist strategy is important in order to
maintain a monotoniously increasing species fitness. The probabilities of
crossover and mutation should be set to 0.6 < pcross < 0.9 and 0.001 < pmut <
0.01. The genetic algorithm itself does not converge down to the exact
solution. It is very good at hunting down the approximate solution but does
not achieve the perfect solution each time. This fact becomes obvious when
multiple runs of the GA turns up many different answers which all lie very
close to one another. Therefore the use of another search routine after the
initial GA search may be necessary. To take advantage of the GA that is
already being used, simply create a new initial population that varies by ~5%
from the previous GA result and rerun the GA. This will force the GA to
significantly reduce its search space. There is no current theory that proves
that the genetic algorithm is capable of achieving the correct solution 100% of
the time. This is because the GA is dependent on the correct setup of the
problem. For instance, if the GA converges quickly to a wrong answer the
population size is too small. These user issues are difficult to quantify into an
exact formula for the setup of the GA. Considering all these criteria, we have
formulated and selected the parameters.
94
4.5.1 Problem Formulation
Consider an array of antenna consisting of N number of elements. It
is assumed that the antenna elements are symmetric about the center of the
linear array. The far field array factor of this array with an even number of
isotropic elements (2N) can be expressed as
1( ) 2 cos(2 sin )N
n nnAF a d (4.1)
where an is the amplitude of the nth element is the angle from broadside and
dn is the distance between position of the nth element and the array center. The
objective is to find an appropriate set of required element amplitude an that
achieves interference suppression with maximum sidelobe level reduction.
To find a set of values which produces the desired array pattern, the
algorithm is used to minimize the following cost function given in
Equation (4.2)
90
90( )[ ( ) ( )]o dcf W F F (4.2)
where Fo( ) is the pattern obtained using our algorithm and Fd( ) is the
pattern desired. Here it is taken to be the Chebyshev pattern with SLL of
-13dB and W( ) is the weight vector to control the sidelobe level in the cost
function. The value of weight vector is to be selected based on experience and
knowledge. The two different approaches used in our investigation are
detailed as follows.
4.5.2 ULA Synthesis using Simple GA
A continuous GA with a population size 10 and a mutation rate of
0.35 is run for a total of 500 generations using MATLAB and the best result is
95
found for every iterations. The cost function is the minimum sidelobe level
for the antenna pattern. Figure 4.9 shows that the antenna array with N = 8
elements has been normalized for a gain of 0dB along the angle 0° and the
maximum relative sidelobe level of -15dB.
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-30
-25
-20
-15
-10
-5
0
Figure 4.9 Optimized Radiation pattern with reduced Sidelobe level of
-15dB for N=8 elements
The convergence of the algorithm for maximum reduction in the
relative sidelobe level with N = 8 elements is depicted in Figure 4.10.
0 10 20 30 40 50 60 70 80 90 100-22
-21
-20
-19
-18
-17
-16
-15
-14
-13
-12
generation
Figure 4.10 Convergence of sidelobe level with respect to evolving
generations for N=8 elements
96
It starts from -13dB which is the optimized value of chebychev
pattern for the RSLL and after 8 iterations it reaches -18.8dB and after 43
generations it converges to a maximum reduction of -21dB. Figure 4.11
shows the optimized radiation pattern with relative sidelobe level of -15dB
with N=16 and Figure 4.12 shows its convergence curve. The convergence
curve shows that it converges to -19.3dB after 54 generations. Changing the
number of elements causes the contiguous GA to get different optimum
weights. Among N=8, 16, 20, and 24, N=20 performed well and thus selected
as optimized element number. The corresponding array pattern for N=8, 16,
20, and 24 are shown in Figure 4.13. In this, the radiation pattern for N=20
has the best directivity with minimum relative sidelobe level of -14.67dB
below the main beam.
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-30
-25
-20
-15
-10
-5
0
Figure 4.11 Optimized Radiation pattern with reduced sidelobe level of
-15 dB for N = 16 elements
97
0 10 20 30 40 50 60 70 80 90 100-20
-19
-18
-17
-16
-15
-14
-13
generation
Figure 4.12 Convergence of sidelobe level with respect to evolving
generations for N=16 elements
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-30
-25
-20
-15
-10
-5
0N=24N=20N=16N=8
Figure 4.13 The optimized radiation pattern with reduced sidelobe level
for N=8, 16, 20 and 24
Figure 4.14 shows the convergence of sidelobe level for N=20.
Figure 4.15 shows the optimized radiation pattern with RSLL of -18.7dB with
N=20 elements.
98
0 10 20 30 40 50 60 70 80 90 100-20
-19
-18
-17
-16
-15
-14
-13
generation
Figure 4.14 Convergence of sidelobe level with respect to evolvinggenerations for N=20 elements
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-30
-25
-20
-15
-10
-5
0
Figure 4.15 The optimized radiation pattern with reduced sidelobe level
for number of elements N = 20
Figure 4.16 shows the convergence curve for N=24 elements.
Figure 4.17 shows the optimized radiation pattern with RSLL of -14.97dB
with N=24 elements.
99
0 10 20 30 40 50 60 70 80 90 100-19
-18
-17
-16
-15
-14
-13
generation
Figure 4.16 Convergence of sidelobe level with respect to evolving
generations for N=24 elements
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-30
-25
-20
-15
-10
-5
0
Figure 4.17 The optimized radiation pattern with reduced sidelobe level
for number of elements N= 24
The obtained costs are ranked from best to worst. The most among
suitability criteria is to discard the bottom half and to keep the top half of the
list. But in our program the selection criteria is to discard any chromosome
that has relative sidelobe level less than -15dB. Table 4.1 shows the cost
100
function relative to the population that has a SLL less than -15 dB. Among 10
populations only 5 are selected. This limitation speeds up the convergence of
the algorithm. After this natural selection the chromosomes mate to produce
off springs. Mating takes place by pairing the surviving chromosome. Once
paired, the offspring consists of genetic material from both parents.
Table 4.1 Population and Respective Cost Function Values
Index Chromosome (weight vector)Relativesidelobe
level (dB)
1 0.8933 0.9059 0.7956 0.7167 0.5880 0.3725 0.2648 0.2232 -26.9682
2 0.8933 0.7659 0.5982 0.5391 0.9296 0.7216 0.7412 0.2059 -19.0861
4 0.4635 0.7659 0.6338 0.5391 0.9296 0.2594 0.7412 0.2059 -18.2047
7 0.8933 0.7659 0.8712 0.4302 0.9296 0.2594 0.7412 0.2059 -17.5515
5 0.8374 0.4270 0.7824 0.5322 0.5425 0.8238 0.9641 0.3051 -17.5173
Figure 4.18 shows the amplitude excitation for optimized antenna
array as given in Table 4.2.
0 5 10 15 20 250
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
elements
weight vector for N = 20 elements
Figure 4.18 Amplitude distribution for optimized antenna array with
N=20 elements
101
Table 4.2 Amplitude excitation values For N=20 elements
corresponding to Figure 4.16
Wn Amplitude excitationW1
W2
W3
W4
W5
W6
W7
W8
W9
W10W11W12W13W14W15W16W17W18W19W20
0.90280.96450.72590.69100.70260.94910.77890.34780.50970.63190.53580.56250.46960.48280.24110.54640.00600.10600.40430.4334
4.5.2.1 Roulette wheel selection
In this thesis, the following parameters are defined; maxgen = 500,
maxfun = 1000 and mincost = -50dB. Population is generated randomly. Then
it is sorted based on its cost - minimum sidelobe level. For choosing mates for
reproduction Roulette wheel selection is used. Each weight vector is assigned
a probability of selection on the basics of either its rank in the sorted
102
population or its cost. Rank order selection is the easiest implementation of
roulette wheel selection Figure 4.19 Shows the Roulette wheel selection
probabilities for five parents in the mating pool. The chromosome with low
sidelobe level has higher percent chance of being selected than the
chromosomes with higher sidelobe level. In this case first or the best weight
vector has a 43% chance of being selected.
9% 3%17%
28%
43%
Figure 4.19 Roulette wheel probabilities for five parents in the mating pool
As more generations are added, the percent chance of weight vector
being selected changes. Figure 4.20 shows the Roulette wheel selection for
seven parents in the mating pool.
25%
33%
18%
1%4%
7%
12%
Figure 4.20 Roulette wheel probabilities for seven parents in the mating
pool
103
The best weight vector has 33% chance of being selected. The
roulette wheel selection needs to be computed only once, because the number
of parents in the mating pool remains constant from generation to generation.
The Genetic Algorithm has converged well for a variant of options
mentioned above with some trade-offs to have main impact on convergence
speed.
4.5.3 ULA Synthesis using GA Solver Optimization Tool Box
Optimization toolbox with GA-Genetic Algorithm solver in
MATLAB has been used in this thesis to find the amplitude excitations to
achieve minimum sidelobe level of -50 dB. Half the number of elements is
used as the number of variables with the Lower Bound (LB) = 0 and Upper
Bound(UB) = 1. The details of the other parameters set are as follows.
Population size = 20; Selection function = Roulette; Reproduction
(Elite count) = 1; Mutation function = Adaptive feasible; Crossover function
= Single point
A. Case 1
Number of variables = 8; Number of array elements =16;
The experiment has been conducted for 25 times and the best
results are presented here.
Figure 4.21 shows four different plots viz 1) Best fitness 2) Best
individual 3) Score Diversity and 4) Array pattern.
104
0 50 100-60
-40
-20
0
Generation
Best: -48.9263 Mean: -48.8641
1 2 3 4 5 6 7 80
0.5
1
Number of variables (8)
Current Best Individual
-49 -48.8 -48.6 -48.4 -48.20
5
10
15Score Histogram
Score (range)-50 0 50
-60
-40
-20
0
Best f itnessMean fitness
Figure 4.21 Showing (a) best fitness (b) best individual (c) score
histogram and (d) radiation pattern for N = 16 elements
Best result of – 48.9263dB sidelobe level is obtained with a mean
value of -48.8641dB. The number of variables is selected as 8, as the antenna
array consists of even number of elements which is symmetric about the
center. The Score Histogram shows, among 20 of the population, 12
individuals give the best score <-48 dB. It converges to -48dB only after 75
generations.
Figure 4.22 shows that the sidelobe level is reduced to – 36.7213dB
with a mean value of -38.6051dB. The Score Histogram shows 13 individuals
get the score < -36.6 dB. The amplitude excitations of best individuals
are obtained as w1 = 0.9853; w2 = 0.9242; w3 = 0.8215; w4 = 0.6698;
w5 = 0.5218; w6 = 0.3527; w7 = 0.2316; w8 = 0.1406;
105
0 50 100-40
-30
-20
-10
0
Generation
Best: -36.7213 Mean: -36.6051
1 2 3 4 5 6 7 80
0.5
1
Number of variables (8)
Current Best Individual
-36.8 -36.6 -36.4 -36.2 -360
2
4
6
8Score Histogram
Score (range)-50 0 50
-60
-40
-20
0
Best fitnessMean fitness
Figure 4.22 Showing (a) best fitness (b) best individual (c) score
histogram and (d) radiation pattern for N = 16 elements
The same is tabulated in Table 4.3 for 16 elements. The sidelobe
levels are almost constant for 6 sidelobes and the last one is wider and less
than the remaining. The convergence takes place in 80 generations.
Table 4.3 Amplitude Excitations of a 16 Element Array
W1 W2 W3 W4 W5 W6 W7 W8 W9 W10 W11 W12 W13 W14 W15 W16
0.14 0.23 0.35 0.52 0.67 0.92 0.82 0.98 0.98 0.92 0.82 0.67 0.52 0.35 0.25 0.14
B. Case 2:
Number of variables = 10; Number of array elements = 20;
The experiment is repeated for 10 variables. Figure 4.23 shows that
the sidelobe level is reduced to -31.147dB whereas the mean is -30dB. All the
individuals lie within the range of -30.5dB to -31.5dB. The main beam width
is narrower but the sidelobes are wider.
106
0 50 100-40
-30
-20
-10
0
Generation
Best: -31.1473 Mean: -30.8601
1 2 3 4 5 6 7 8 9 100
0.5
1
Number of variables (10)
Current Best Individual
-31.5 -31 -30.5 -300
2
4
6Score Histogram
Score (range)-50 0 50
-60
-40
-20
0
Best f itnessMean f itness
Figure 4.23 Showing (a) best fitness (b) best individual (c) score
histogram and (d) radiation pattern for N = 20 elements
A similar case is run and the result is shown in Figure 4.24.
0 50 100-40
-30
-20
-10
Generation
Best: -32.1697 Mean: -31.9476
1 2 3 4 5 6 7 8 9 100
0.5
1
1.5
Number of variables (10)
Current Best Individual
-32.5 -32 -31.5 -310
2
4
6Score Histogram
Score (range)-50 0 50
-60
-40
-20
0
Best fitnessMean f itness
Figure 4.24 Showing (a) best fitness (b) best individual (c) score
histogram and (d) radiation pattern for N = 20 elements
107
C. Case 3
The experiments are conducted with 22, 42, and 62 elements for 25
runs and their performance are tabulated in Table 4.4 and compared with that
of a table given in Haupt(2004).
Table 4.4 Performance Comparison of Antenna Array with Different
Number of Elements
N 8 16 20 24RSLL (dB) -16.1 -15 -18.7 -14.97Final Convergence (dB) -21.3 -19.3 -19.4 -18.7No. of generations 43 54 31 42
Table 4.5 shows the performance characteristics of five algorithms
for an average of 25 runs with random seed values of the amplitude weights.
Table 4.5 Comparisons of Optimized Sidelobes for Three Different Array
Sizes using Other Algorithms (Haupt 2004) and Genetic
Algorithm
Algorithm
22 Elements 42 Elements 62 ElementsMedianSidelobe
Level(dB)
MedianFunction
Calls
MedianSidelobe
Level (dB)
MedianFunction
Calls
MedianSidelobe
Level (dB)
MedianFunction
Calls
BFGS -30.3 1007 -25.3 2008 -26.6 3016DFP -27.9 1006 -25.2 2011 -26.6 3015Nelder Mead -18.7 956 -17.3 2575 -17.2 3551Steepest descent -24.6 1005 -21.6 2009 -21.8 3013Our proposed work(Genetic Algorithm )
-22.3 830 -20.3 940 -20.9 860
108
Genetic algorithm performs well when compared to Nelder Mead
but poorer when compared to the remaining algorithms. But the function calls
are minimum than all other algorithm. Hence it is cost effective in terms of
computational time. Genetic algorithm shows the best results of median
sidelobe level of -32.04dB with median function calls of 700 when the array
size is 16 elements. Among the three cases the number of elements of the
antenna array with N = 16 performed very well with narrow main beam width
and reduced sidelobe level and minimum number of function calls which cost
less computation time and less complexity.
From our study it is realized that Genetic algorithm has many
variables to control and trade-offs to consider as detailed below
Number of Chromosomes and initial random Population: more
number of chromosomes provide better sampling number,
solution space but at the cost of slow convergence.
Generating the random list, the type of probability distribution
and weighting of the parameters has a significant impact on the
convergence time.
Roulette selection method is employed to decide which
chromosome to discard.
Crossover the chromosome for mating, the chromosome may be
paired from top to bottom randomly best to worst.
Mutation rate is selected to mutate a particular chromosome.
Mutate does not permit the algorithm to get stuck at local
minimum.
109
Stopping Criteria, set in this program are maxgen = 500,
maxfun = 1000 and mincost = -50dB.
In our investigation the Genetic Algorithm has converged well for a
variant of options mentioned above with some trade offs to have main impact
on convergence speed.
4.6 SUMMARY
In this investigation two different approaches of namely (a) Simple
GA and (b) Genetic algorithm Solver in Optimization toolbox of MATLAB
are used to obtain maximum reduction in sidelobe level relative to the main
beam on both sides of 0°. The specialty of the Genetic algorithm is that it can
optimize the large number of discrete parameters. Genetic algorithm is an
intellectual algorithm searches for the optimum element weight of the array
antenna. This investigation demonstrated the different ways to apply Genetic
algorithm such as varying number of elements, to optimize the array pattern.
Adaptive feasible mutation with single point crossover and Roulette selection
showed the performance improvement by reducing the sidelobe level below
-30dB in most of the cases with number of variables as 8 and minimum
function calls when compared to the other methods shown in Table 4.4. The
best result of -48.9dB is obtained for 16 elements proving that this method is
efficient with much of the computation time and complexity are reduced.
Unlike Simple GA (SGA), the Genetic algorithm solver from the optimization
toolbox of MATLAB is used with adaptive feasible mutation, which enables
search in broader space along randomly generated directions to produce new
generations. This improves the performance greatly to achieve the maximum
reduction in sidelobe level with minimum function calls. Experiments proved
the effectiveness of this method.