IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 45, NO.

3, MARCH 1997 343

Genetic Algorithm Optimization

Applied to Electromagnetics:
A Review
Daniel S. Weile and Eric Michielssen, Member, IEEE

Invited Review Paper

Abstract—Genetic algorithms are on the rise in electromagnet- applied to problems in fields ranging from engineering to
ics as design tools and problem solvers because of their versatility economics and artificial intelligence and their use is becoming
and ability to optimize in complex multimodal search spaces. widespread in electromagnetics [11]–[13]. GA’s can easily
This paper describes the basic genetic algorithm and recounts its
history in the electromagnetics literature. Also, the application be applied to problems involving nondifferentiable functions
of advanced genetic operators to the field of electromagnetics and discrete search spaces. Also, they have an inherent par-
is described, and design results are presented for a number of allel nature which renders them attractive for electromagnetic
different applications. optimization. Motivated by Darwin’s theory of descent with
Index Terms— Genetic algorithms. modification by natural selection, GA’s evolve a population of
candidate designs toward a goal. The stochastic nature of GA’s
should not be confused with directionlessness as evolution is
I. INTRODUCTION emphatically not a random walk; the principle of survival of

D URING the latter half of the nineteenth century, the

biological sciences underwent a revolution when Charles
Darwin discovered the processes by which nature selects and
the fittest leaves little room for interpretation.
This study considers the history and current directions
of GA’s in electromagnetics. In addition, it explains the
optimizes organizms fit for life [1]. About the same time, GA itself and provides the reader with guidelines for the
Gregor Mendel learned the basic laws of genetic inheritance implementation of a successful GA. Section II introduces
which elucidate by what means evolution takes place [2]. The GA’s and their operators. Section III recounts the applica-
advent of computers and powerful computational techniques tion of GA’s in the electromagnetics literature. Advanced
now enables us to apply Nature’s optimization processes in genetic techniques which demonstrate the wide applicability
the form of genetic algorithms (GA’s) to devices built using of GA’s to electromagnetic device optimization are discussed
Maxwell’s equations. in Section IV. Conclusions and avenues for future research
The objective functions that arise in electromagnetic op- on the application of GA’s to electromagnetic problems are
timization problems are often highly nonlinear, stiff, multi- presented in Section V.
extremal, and nondifferentiable. In addition, they are almost
always computationally expensive to evaluate. Historically, II. GENETIC ALGORITHMS: AN INTRODUCTION
the vast majority of research efforts related to the design
In their most basic form, GA’s are function optimizers, i.e.,
of electromagnetic systems using optimization methods has
methods for seeking extrema of a given objective function
relied on deterministic optimization methods (DOM’s) [3].
where . In the electromag-
DOM’s are known to have important drawbacks when ap-
netic community, function optimizers of all types are used
plied to multiextremal and stiff optimization problems [4]
primarily in design and in solving inverse problems. In design
and often lead to highly interactive and expensive design
problems, the are parameters describing salient features of
procedures. GA’s (along with Monte Carlo techniques and
a design, and is a measure of system performance to be
simulated annealing) belong to a small but growing class of
either maximized or minimized. The can be continuous
so-called global optimizers which are stochastic in nature and,
or discrete in nature, and complex, real or integer valued.
therefore, less prone to converge to a weak local optimum
Whatever the nature of the search space, though, GA’s differ
than DOM’s [5]–[10]. In various forms, GA’s have been
from most traditional optimization methods in two ways.
First, they do not necessarily operate directly on the design
Manuscript received August 7, 1996; revised September 30, 1996. This parameters , and second, they simultaneously optimize entire
work was supported in part by the National Science Foundation under Grant
ECS-9502138, The AFOSR under Grant F49620–96–1–0025, and an NSF populations of designs at once, not a single design at a time.
Graduate Fellowship. We discuss both of these characteristics in turn.
The authors are with the Department of Electrical and Computer Engi- In general, GA’s do not operate directly on the parameter
neering, Center for Computational Electromagnetics, University of Illinois at
Urbana-Champaign, Urbana, IL 61801 USA. vector , but on a symbolic representation of , known as
Publisher Item Identifier S 0018-926X(97)02284-9. a chromosome. A chromosome is a concatenation of genes
0018–926X/97$10.00  1997 IEEE
344 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 45, NO. 3, MARCH 1997

which decode to the , and is symbolically denoted as population with a probability proportional to its fitness value
(1)
S s (5)
where is the genetic length and there is a correspondence
between the and the given by
where s with probability

s (6)
(2)
where . In allelic GA’s, the genes are selected
from a finite alphabet (most often the binary alphabet ), Ranking selection comes in many different forms, but a
popular version places copies of the best designs
whose members are known as alleles, which together decode
into . The parameter controls the greediness of the
to a discrete or (discretized) continuous parameter. For exam-
ranking algorithm; a large will force a fast convergence
ple, in binary-coded GA’s, real-valued parameters can be
at the likely expense of a thorough global search. Faster
decoded through a simple linear transformation
than both ranking and roulette-wheel selection is stochastic
binary tournament selection which iteratively chooses pairs of
(3)
chromosomes from and places the better one in until
it is satiated. Invoking Darwin’s ideas, the selection operator
where takes values ranging from to , and is responsible for the algorithm’s convergence, and is the only
. This type of allelic decoding can be avoided entirely, operator which involves the objective function values .
however, through the use of real-coded GA’s which use Selection is followed by crossover, which serves to hy-
real genes (instead of selecting them from an alphabet) bridize design traits by creating a new population
for representing real-valued parameters through the simple C , again of size . Crossover can be denoted as the
genotype to phenotype mapping so that union over a composition of operators:
[14].
The second difference between GA’s and more standard C c ch ch (7)
optimizers is that GA’s do not work on a single chromosome
at a time, but on a whole population of chromosomes where the operator “ch” chooses a random chromosome
(4) from and the operator “c” maps a pair of chromosomes
and
The GA itself is composed of operators which produce a
to another pair according to the rule
succession of populations whose members will have gener-
ally improving objective function values. Given a population with probability
c (8)
, a single GA iteration starts with probability
by evaluating the vector of where the hybrids and are given by
objective function values associated with chromosomes
. The GA then applies the genetic operators of selection, (9)
crossover and mutation to to produce .
To commence this iteration, the GA generates a random The crossover locus in (9) is chosen randomly from in-
tegers between and . This specific single locus
initial population , although a priori knowledge sometimes
implementation is known as one-point crossover. Motivated
is invoked to seed with good chromosomes. With the
by the fact that one-point crossover disallows the possibility
creation of , the population enters the main GA loop which
of alleles residing at opposite ends of the chromosome from
is iterated on each successive population . Genetic manipu-
ever staying together during a cross, alternative mechanisms
lation of begins with the selection of good chromosomes on
have been proposed. For example, in two-point crossover the
the basis of their objective function values . The selection chromosomes to be mated are split at two loci and exchange
operator produces a new population S , also of the central set of genes, allowing the alleles on the ends
designs which, on average, will be better than those to stay together. In uniform crossover, each of the parents’
in . This can be achieved by many different schemes, genes is exchanged with a given probability. The crossover
but the most common methods are roulette-wheel, ranking, operator for real-coded GA’s has not been as well codified or
and stochastic binary tournament selection. For simplicity, we studied as those for binary-coded GA’s, but one technique for
discuss selection in the context of maximization problems, accomplishing it can be found in [15]. Crossover is the main
but all of the selection schemes described can be applied to search tool of the GA since it combines chromosomes which
minimization problems either directly or with a change of contain genetic information which is known to be useful.
objective function through a linear transformation. Roulette Finally, the mutation operator creates a new population
wheel selection randomly places each design in the next M of size by randomly perturbing genes. If
WEILE AND MICHIELSSEN: GENETIC ALGORITHM OPTIMIZATION APPLIED TO ELECTROMAGNETICS 345

we denote the members of the population after crossover as These theorems imply (indirectly) the following implemen-
, mutation for allelic algorithms can be defined as tation guidelines.
• GA’s work best if the optimal solution is a member of
M m (10) schemata which have few defined bits that are packed
tightly together, and the involved schema have a low-
standard deviation. That is, since GA’s search primarily
where, given a chromosome , by combining pieces of good chromosomes, codings
m , and should be devised so that similar chromosomes have
with probability similar objective function values. In this way, the GA
(11) is not deceived in creating chromosomes which are much
with probability
worse than the sum of their parts.
where is any other legal allele. In real-coded algorithms, • For most practical applications, a population size pro-
mutation is very similar to that described for allelic GA’s, portional to the chromosome length should suffice [19],
except that is, in general, a small random perturbation although there is never any harm in using a larger
of its initial value [15]. Mutation is included to prevent population. The implicit parallelism theorem implies that
premature convergence by ensuring allelic diversity at each doubling the population may more than halve the run
gene position. In general, it is not considered a search tool as, time; thus, if the objective function calculation is fast
unlike crossover, it alters chromosomes blindly. enough for the use of a large population to be feasible, a
The population resulting from the application of the muta- large population should be used.
tion operator serves as the starting point for the next generation
Because of the incomplete nature of GA theory, much
or ; hence, a single GA generation can be
knowledge about the successful implementation of GA’s
described as M CS . If the best member
comes from experience and experiment.
of is not in , an elitism operator is very often
included to insert this best member into [6]. This process • For binary-coded GA’s, the crossover and mutation rates
is iterated until no improvement is observed, a set goal is most often quoted in the literature are 0.6 0.9,
met, or a set number of generations has passed. Though the and 0.001 0.01. Generally, corresponds
algorithm described here forms the basis of all GA’s, a plethora to at most the mutation of one or two alleles per chro-
of advanced operators (see Section IV) may be included to mosome and at least a few chromosomes per generation.
ameliorate the performance of the simple GA. Real-coded GA’s often use higher mutation rates.
Despite the successes of GA’s and their facile implemen- • Problems with real design parameters are often (but not
tation, GA theory remains largely enigmatic. Currently, there always) better handled by real-coded GA’s. In general,
is no general GA theory to prove population convergence to codings for physical problems work best if they resemble
global or even local optima, nor is there any clear delineation the parameter they model.
between problems on which GA’s work well or poorly. Still, • The inclusion of a local optimizer to either: 1) seed
some questions about the implementation of GA’s can be the initial population; 2) optimize the intermediate pop-
answered, either based on the incomplete models developed ulations; or 3) touch up the final result, can speed the
by GA theorists, or from the experimental results garnered by optimization process by removing the local refinement
GA researchers. Therefore, we catalogue these implementation burden from the GA.
guidelines here. • Binary tournament selection generally works faster than
Most of the results of theoretical GA studies revolve around roulette-wheel selection, and it avoids convergence trou-
the concept of schemata, or similarity templates, and apply to bles [20].
binary-coded GA’s using roulette-wheel selection and single- • When the GA stagnates, raising the mutation rate often
point crossover. A schema is a set of chromosomes which are yields useful information; if the objective function value
similar in some way. For instance, the schema of the best design changes significantly, the GA may have
is the set of binary 10-b chromosomes with defined bits of one been temporarily stuck in a local optimum and continued
in the first position and zero in the fourth position, and don’t optimization is likely to be useful. If not, restarting the
cares “ ” elsewhere. There are two important theorems dealing GA with a new initial population will yield a better chance
with the processing of schemata by the GA [7], [8], [16]–[18]. for improvement than attempting to push further.
• Many difficult problems will succumb to GA optimization
• The schema theorem states that schemata with above after the inclusion of an advanced operator. Sharing
average objective function values will receive exponen- (described in Section IV) often induces the GA to produce
tially increasing representation in successive populations, novel results [21].
especially those schemata which have very few defined
bits located close to one another. III. GENETIC ALGORITHMS IN
• The implicit parallelism theorem states that the GA pro- ELECTROMAGNETICS: A RETROSPECTIVE
cesses schemata in each generation. It also Since the early part of this decade, GA’s have been applied
shows that for allelic GA’s, more schemata are pro- with growing frequency to the design of electromagnetic
cessed if the alphabet is small. Historically, this was the devices of increasing complexity. This section presents a
motivation for using binary codings in the first GA’s. review of research in this area since 1992. Because of the rapid
346 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 45, NO. 3, MARCH 1997

growth of GA applications in electromagnetics, this review Computationally more expensive than the above problems is
cannot be complete. Though the applications reviewed come the design of frequency selective surfaces (FSS’s) [55], [56].
from all areas of electromagnetic research, for the purpose The optimization can proceed identically to that of layered
of this review, we divide them into four categories: antennas, optical filters or absorbers, except with a more complicated
stratified medium structures, static devices, and miscellaneous. chromosome and time-consuming objective function evalua-
tion. Due to the variety of design parameters encountered in
A. Genetic Algorithm Antenna Design these problems and the large computation time required, [57]
Antenna design (in particular, array design) is the most uses hybrid chromosomes with both real and allelic parts to
studied application of GA’s in electromagnetics. A plethora optimize the chromosome structure as much as possible.
of studies have investigated GA-based methods of reducing
the sidelobes of an array by thinning [22]–[26], amplitude C. Applications of Genetic Algorithms in Statics
or phase tapering [27]–[35], or element position perturbations Many applications of GA’s involve shaping magnetic pole
[36]. Closely related to this problem is reducing the scattering pieces or insulators to produce a desired magnetic or electric
of strip arrays by thinning or perturbation [28], [37]. Most of field distribution in a given region of space [58], [59]. Typi-
these studies involve digital parameters and, thus, are textbook cally, chromosomes describe the location of several points in
cases for binary or other finite alphabet GA’s. For instance, space which are interpolated to produce the shape of the device
an array built with ten 3-b phase shifters can be encoded into under consideration. The objective function to be minimized is
a chromosome of 30-b, composed of a concatenation of ten some norm of the vector difference between the field produced
3-b phase shifter settings. Additionally, many of these studies by the device and the desired field at several points.
draw on the implementation guidelines listed in Section II. For One notable study of this sort is [60], in which the authors
example, in [22], Buckley combines the GA technique with a optimize a pot-core transformer not only to produce a specific
local optimizer which is used to seed the GA. GA’s have also field pattern, but to minimize the device dimensions as well.
designed arrays with specified null locations [38] or a specific The inclusion of a second design goal renders this problem
pattern shape [39]. difficult and poorly defined. As the stated goals are physically
Besides arrays, single antenna elements have been designed incommensurable, they are combined in an ad hoc manner to
as well. For example, in [40]–[44], a technique is described for achieve a palatable result. This technique is very common in
the fast GA optimization of wire antennas loaded with passive optimization, but may lead to suboptimal results or extensive
tinkering to find a suitable algebraic combination of objectives.
resonant circuits (RLC) to broaden their frequency response.
Perhaps even more interesting is the application of GA’s
Reflector antennas [45], [46] and Yagi antennas [47] have also
to magnetostatic inverse problems [61]–[63]. Such problems
been optimized by GA’s.
easily fit the GA mold; the unknowns (for instance the
Perhaps the most intriguing application of GA’s to location, orientation, and size of a crack [61]) are encoded into
antenna design is the so-called “genetic antenna” [48]. In chromosomes, and the objective function to be minimized is
this study, a wire antenna was designed to radiate circularly the difference between the field produced by the configuration
polarized (CP) waves by choosing endpoints for straight described in the chromosome and the desired distribution. A
wire segments comprising the antenna and then “connecting last application is found in [64], where coils are designed to
the dots.” The GA produces counterintuitive antennas with create fields sensitive to cracks in aircraft.
radical bends, but they do indeed radiate CP waves over
a considerable frequency band. D. Miscellaneous Applications of Genetic
Algorithms to Electromagnetics
B. Design of Layered Electromagnetic Devices
In addition to the application areas mentioned above, GA’s
Among the earliest GA studies in electromagnetics is the have been applied to the design of wireless communication
design of multilayered optical filters by minimizing the dif- networks [65], lenses [66], waveguide junctions [67], dielectric
ference between the observed and desired filter characteristic gratings [68], and RF coils [69].
[15], [49], [50]. Because [15] realizes filters by layering two Outside the realm of design, GA’s have been used for
preselected materials, the only design parameters considered solving two-dimensional inverse scattering problems [70],
are layer thicknesses which are real valued and arrayed into to extract resonances from data sets [71], [72], to produce
a real-coded chromosome. In accordance with the previous pulses for target identification [73], in scattering center
section’s guidelines, the authors make the observation that analyses of radar targets [74], and to generate super-resolved
a real-coded GA outperforms a binary coded GA for this spatial excitation patterns in two-level systems [75]. They have
problem. Very similar to the design of optical filters is the even been applied to the optimization of perfectly matched
design of absorbers consisting of a perfect electric conductor layers for finite-difference applications [76].
(PEC) coated with lossy materials to minimize reflection of
impinging waves [51]–[54]. Despite the similarities among IV. DESIGN EXAMPLES USING ADVANCED GA OPERATORS
all of these devices, the GA’s used to optimize them might Most of the studies cited in Section III follow the basic
be quite different depending on the nature of the problem. GA pattern outlined in Section II with little or no deviation.
For instance, because [52] chooses materials from a finite However, as GA’s are applied to electromagnetic design
database, an allelic chromosome is used instead of the real problems of increasing complexity, they will need to be
coding mentioned above. augmented to optimize faster or more globally. This can
WEILE AND MICHIELSSEN: GENETIC ALGORITHM OPTIMIZATION APPLIED TO ELECTROMAGNETICS 347

0
Fig. 1. A flowchart illustrating the genetic algorithm process optimizing the function f (x) = x(15 x). A 4-b chromosome is used, which decodes to
an integer x between 0 and 15. After the initial population is generated and evaluated, selection eliminates the inferior chromosome 1111 and replaces it
with 0101, which has a very high objective value. The chromosomes p2 and p4 then undergo crossover between the second and third genes, whereas the
other chromosomes are not chosen to be crossed. Finally bits in the first three chromosomes are mutated, and the process begins anew by evaluating
the objective function values of the new population.

be accomplished by using advanced GA techniques such the GA is not misled. We illustrate this by applying a simple
as dominance and diploidy, community structure, sharing, GA to the design of a wire antenna situated in an arbitrary,
or knowledge-based operators that can vastly improve GA possibly complex, environment for operation over a broad
performance at the expense of very small modifications of the range of frequencies [41]. The wire is to be augmented with
algorithm [7]. Thus, we present three examples of advanced a given number of parallel RLC loads whose locations and
GA’s that further tap the potential of the GA technique and
element values are to be determined by the GA to produce
also serve to give concrete examples of GA optimization at
work. All of the examples presented use the standard binary maximum gain over the frequency band of interest. In addition,
coding [including the decoding of real parameters given in the GA simultaneously designs a matching network to achieve
(3)] and one point crossover. an acceptable standing wave ratio (SWR) over the band. As
GA’s typically require many objective function evaluations,
A. Genetic Algorithm Design of Broadband Loaded a generalization of the Sherman–Morrisson updating scheme
Wire Antennas in a Complex Environment [44] using numerical Green’s functions was developed, which
The most critical issue in creating a successful GA involves permits the fast analysis of loaded structures residing on a
the coding of the chromosomes which must be designed so complex platform [41].
348 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 45, NO. 3, MARCH 1997

Fig. 2. An illustration of the broadband antenna design problem. The top inset is a general matching network topology, as described in Section III-A,
and the bottom inset is an actual matching network optimized by the GA.

At first glance, it appears that this antenna can be opti- similarly. This requirement is a direct consequence of the
mized in a straightforward manner after incorporating load schema theorem described in Section II. A design in the
locations and element values into a chromosome using (3). population with a good objective function value may be
The difficulty, however, lies in the design of the matching rendered useless if mutation happens to change a key capacitor
network. For a given network topology, encoding parameter in the matching network to an inductor.
values is straightforward, but often the topology itself requires A more GA friendly approach assumes a general matching-
optimization. network template, as shown in Fig. 2. By allowing the most
A simple way to optimize network topology is to encode extreme values of the groups of genes describing the capac-
all matching network component values with two extra control itances and inductances in the network to take on values of
bits. The first of these bits could indicate whether the element zero and infinity, elements may slowly evolve into shorts or
in question is a capacitor or an inductor, and the second opens as needed. The number of elements in the network may
whether the element is series or shunt connected. The matching be kept to a minimum by penalizing designs with an excessive
network could then be built element by element from the number of elements. The matching network part of the gene
antenna back to the source. This would lead to a chromo- would now require
some structured as follows. The first part of the chromosome
would contain descriptions of the antenna loads and locations (14)
spanning the first bits. If the antenna has possible bits where is the number of matching-network sections,
load locations, and RLC loads with components that and is the number of bits added to allow for elements
can be adequately described by bits with zero and infinity values. The antenna loading part of the
(12) chromosome is unaltered.
With this coding, the problem becomes a straightforward
Describing the matching network would require application of a simple GA. The objective function to be
(13) maximized involves a sum of the system gain and penalties
described more fully in [41] and [77]. As a demonstration of
bits, where is the number of elements in the matching the technique, a whip residing on top of a tank was designed
network. A full antenna-matching network system would then to radiate between 9 and 60 MHz with maximum gain at the
have a chromosome of bits. This approach horizon on the side of the tank (Fig. 2). This problem involves
fails, however, as it violates an important requirement of the a 7 : 1 band, and the antenna is only 3/20 of a wavelength
GA—namely, that designs with similar descriptions behave long at 9 MHz. Antennas for use over bandwidths of up to
WEILE AND MICHIELSSEN: GENETIC ALGORITHM OPTIMIZATION APPLIED TO ELECTROMAGNETICS 349

Fig. 3. The gain and SWR of the GA optimized antenna.

TABLE I Fig. 4. A schematic of the broadband microwave absorber synthesis prob-

THE MATCHING NETWORK, LOAD LOCATIONS, lem. The chromosome layout is also shown for a problem of three layers with
AND COMPONENT VALUES SELECTED BY THE GA a database of eight materials and 16 possible thickness values.

15 : 1 have also been optimized with this technique [44]. The

specific GA implemented to accomplish this used stochastic
binary tournament selection with , ,
, and , resulting in a 148-b chromosome.
The GA parameters used were , ,
and 0.0001 0.005. Capacitance, inductance, and
resistance values were allowed to vary in between 1 and 110
pF, 3 and 5000 nH, and 1 and 1500 , respectively. The Fig. 5. The Pareto front constructed by the GA for the absorber problem.
optimized antenna gain and SWR after 100 generations are Inset is the frequency response of a particular design on the front for TE
waves incident at 30 .
shown in Fig. 3. The matching network, load locations, and
component values selected by the GA are shown in Fig. 3 and
Table I, respectively. balance these competing goals, one might attempt to minimize
an algebraic combination of goals , where represents
the total thickness of the absorber. Unfortunately, finding a
B. Multiobjective Optimization of Broadband Absorbers suitable algebraic combination of goals for use as an objective
Often engineers face the task of satisfying conflicting goals function can be a monumental task.
in a single design. The problem then is not finding a design Pareto GA’s avoid this problem entirely by returning not a
that is optimal with respect to a single design goal, but single optimal design, but the set of designs that represents
a design which makes tradeoffs between conflicting goals. all of the best tradeoffs [78]–[80]. Within the context of the
Whenever this occurs, a multiobjective optimization technique absorber design, this set can be defined as follows: 1) a design
is required. GA’s are ideal for such problems because their is termed dominated if there exists another design in the space
reliance on a population allows them to return a wide variety of all possible designs which is thinner and less reflective; 2)
of designs representing optimal tradeoffs. a design which is not dominated is known as nondominated,
A number of studies have focused on the design of mi- efficient, or Pareto optimal; and 3) the Pareto front is the set
crowave absorbers comprised of layered lossy materials with of all nondominated designs, each of which is considered to
PEC backing (Fig. 4) [51], [52]. Most of these studies employ be equally optimal. The simple GA, described in Section II,
simple binary-coded GA’s with chromosomes consisting of must be altered in two fundamental ways to return the Pareto
two groups of bits per layer describing each layer’s material front instead of a single point: 1) designs must be evaluated
and thickness. Materials are often selected from a database on the basis of their proximity to the Pareto front and 2) the
[52], and the objective function is coded to minimize reflection GA must prevent the population from converging to a small
coefficients of TE and TM waves over a range of frequencies subset of the front instead of the whole front.
and angles of interest. But, besides suppressing reflection, Proper design evaluation is achieved using a scheme known
absorbers often also have to be thin or light to be useful. To as nondomination ranking [7], [81]. The population is ranked
350 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 45, NO. 3, MARCH 1997

Fig. 6. The FSS design problem, along with a fairly general-shape database and a chromosome structure. The example pictured uses 4 b to specify the
cell size (a global design parameter). The design parameters characterizing each layer are shape number from a 16-element database, shape dimension,
and dielectric layer material and thickness.

by first removing all nondominated members and assigning simultaneous optimization of beamwidth and sidelobe level in
them a rank of one. This process is repeated on the population antenna arrays as well [84].
remaining—its nondominated members are assigned a rank
of two, and so on. To force the population members to C. Community GA’s for Periodic Structure Design
spread out uniformly over the front, Pareto GA’s use the Wherever a GA is applied to a problem of great computa-
sharing technique based on competition between organisms tional expense, it becomes necessary to accelerate either the
for environmental resources. In sharing, the objective function device analysis, the GA convergence, or both. An example
value of each population member is divided by a niche count of a technique that achieves both is the community GA for
which is a measure of the population density in its vicinity; optimizing the periodic structure shown in Fig. 6 [85]. This
that is, the niche count is high for designs that are very much structure is not unlike the absorber described in the previous
like many others in the population. This disallows population section, except that the materials used here are assumed
bunching since chromosomes tending to group will be forced lossless and a periodic metallization pattern is included in
to die off [7], [82]. the layer interfaces. The GA not only chooses layer materials
The Pareto GA combines nondomination ranking and shar- and thicknesses as before, but also the shape and size of the
ing. In each generation, rank-one designs are assigned a periodic metallizations, the periodic cell size, and the relative
common objective value and then shared. The lowest shared offset between metallization in different layers to achieve a
value of the rank-one designs is taken as the unshared objective desired transmission or reflection characteristic.
function value for rank-two designs, and so on. This algorithm, The analysis in this study proceeds via a spectral Galerkin
described in more detail in [79], [81], was applied to the technique [86] and, therefore, involves Floquet harmonics with
synthesis of five-layer absorbers with materials from the transverse wave vectors given by
database in [52]. Tradeoffs were found between the maximum
reflection for both polarizations between 0.2 and 2 GHz
and 0 and 30 and the thickness of the absorber. For the
(15)
results shown in Fig. 5, , , and
. Because of well-known problems in GA’s using where and are the size of the periodic cell, and
sharing with tournament selection, roulette-wheel selection are the transverse components of the incident field wavevector,
was used [83]. This Pareto GA has since been applied to the and and are integers. Now, envision a community of
WEILE AND MICHIELSSEN: GENETIC ALGORITHM OPTIMIZATION APPLIED TO ELECTROMAGNETICS 351

TABLE II
DIPOLE TWIST POLARIZER DESIGN

TABLE III
L-DIPOLE TWIST POLARIZER DESIGN
Fig. 7. Polarization response of two twist polarizers. The narrow band
polarizer was designed with dipoles for an incident wave of  = 5o and
 = 45o , and the broadband polarizer was designed with L-dipoles for and
incident wave of  = 5o and  = 0o .

designs, all of which are identical, except in the thickness of

their dielectric layers which are all nonetheless greater than
some value . For those modes for which

(16)

the interactions between different metallized layers can be to the synthesis of novel antennas, inexpensive arrays, filters,
safely ignored, and the intralayer interactions can be computed absorbers, and FSS’s, and have been used to tackle inverse
assuming that the metal is located between two dielectric scattering problems. GA’s have been employed as electro-
half-spaces with dielectric constants equal to those of the magnetic optimizers because of their natural advantages over
surrounding layers. More importantly, interactions for which DOM’s—they are highly suited for optimizing multimodal
condition (2) is true can also be reused for the entire com- problems, and operate with ease on discrete parameters. They
munity. require no initial guesses to start the optimization process
This suggests a community structure for the GA which and often give rise to counterintuitive designs. Though often
begins by creating a population of community chro- slower than DOM’s, the speed of GA’s is increasing due to
mosomes which describe element shapes, positions, etc. Each several factors. GA’s are increasingly being hybridized with
community chromosome is associated with a community of other optimization techniques, either stochastic [87] or deter-
individual chromosomes which describe only dielectric ministic [88] in nature. Furthermore, as parallel computing
slab thicknesses. The GA proceeds by first applying all of resources become more widespread, more researchers will take
its operators to the community chromosomes (assuming for advantage of the innate parallelism of the GA structure as
crossover that the individual chromosomes are attached to objective function values of whole populations of designs can
the right of the community ones). Then, it runs several be calculated in parallel [56]. Finally, the application of messy
generations operating on the individual chromosomes within GA’s that overhaul the basic structure of the simple GA cannot
the communities. Community structures are not unheard of be far off [89].
in GA optimization [7], but they are usually used only to
accelerate the convergence of the GA. Here, communities both ACKNOWLEDGMENT
accelerate analysis and convergence.
Using this technique, two different twist polarizers were The authors would like to thank J. Veihl for his contribution
designed to reflect incident TE waves as TM waves. GA of the mesh for Fig. 2.
parameters for this problem included , ,
REFERENCES
, and . The relatively high
mutation rates result from the computational difficulty of this [1] C. Darwin, The Origin of Species. London, U.K.: Penguin, 1985.
problem; they tend to make the GA converge faster, even if to [2] G. Mendel, Experiments in Plant Hybridization. Cambridge, MA:
Harvard Univ. Press, 1950.
a slightly suboptimal solution. Shapes were picked from the [3] R. Fletcher, Practical Methods of Optimization. Chichester, U.K.:
database shown in Fig. 6, the responses of the two designs Wiley Intersci., 1980.
[4] L. C. W. Dixon, Global Optima Without Convexity, in Design and Im-
are shown in Fig. 7, with the designs themselves described in plementation of Optimization Software. Sijnoof Noordhof, The Nether-
Tables II and III. lands: Aalphen aan den Rijn, 1978.
[5] L. Davis, Genetic Algorithms and Simulated Annealing. London, U.K.:
Pittman, 1987.
V. CONCLUSION [6] K. A. DeJong, “An analysis of the behavior of a class of genetic adaptive
systems,” Ph.D. dissertation, Univ. Michigan, 1975.
GA’s show great promise in electromagnetic device design [7] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Ma-
and problem solution. They have been applied successfully chine Learning. Reading, MA: Addison–Wesley, 1989.
352 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 45, NO. 3, MARCH 1997

[8] J. H. Holland, Adaptation in Natural and Artificial Systems. Ann [38] A. Tennant, M. M. Dawoud, and A. P. Anderson, “Array pattern nulling
Arbor, MI: Univ. Michigan, 1975. by element position perturbations using a genetic algorithm,” Electron.
[9] , “Genetic algorithms,” Sci. Amer., vol. 267, pp. 66–72, July 1992. Lett., vol. 30, no. 3, pp. 174–176, 1994.
[10] S. Forrest, “Genetic algorithms: Principles of natural selection applied [39] D. Marcano, M. Jimenez, and O. Chang, “Synthesis of linear array using
to computation,” Sci., vol. 261, pp. 872–878, Aug. 1993. Schelkunoff’s method and genetic algorithms,” in Proc. IEEE Antennas
[11] S. Begley, “Software au naturel,” Newsweek, pp. 70–71, May 8, 1995. Propagat. Soc. Int. Symp., Baltimore, MD, July 1996, pp. 814–817.
[12] R. A. McIntyre, “Bach in a box: The evolution of four part Baroque [40] E. Michielssen, A. Boag, J. M. Sajer, and R. Mittra, “Design of electri-
harmony using the genetic algorithm,” in Proc. IEEE Conf. Evolutionary cally loaded wire antennas using massively parallel genetic algorithms,”
Computat., Orlando, FL, June 1994, pp. 852–857. in Proc. URSI Radio Sci. Meet., Seattle, WA, June 1994.
[13] J. Ventrella. “Disney meets Darwin—The evolution of funny animated [41] D. S. Weile, E. Michielssen, and A. Boag, “Optimization of broad-band
figures,” in Proc. Comput. Animat., Geneva, Switzerland, Apr. 1995, loaded wire antennas in real environments using genetic algorithms,” in
pp. 35–43. Proc. 12th Ann. Rev. Progress Appl. Computat. Electromagn., Monterey,
[14] D. E. Goldberg, “Real-coded genetic algorithms, virtual alphabets and CA, Mar. 1996, pp. 726–733.
blocking,” Complex Syst., vol. 5, no. 2, pp. 129–167, 1991. [42] Z. Altman, R. Mittra, J. Philo, and S. Dey, “New designs of ultra-
[15] E. Michielssen, S. Ranjithan, and R. Mittra, “Optimal multilayer filter broadband antennas using genetic algorithms,” in Proc. IEEE Antennas
design using real coded genetic algorithms,” Proc. Inst. Elect. Eng., pt. Propagat. Soc. Int. Symp., Baltimore, MD, July 1996, pp. 2054–2057.
J., vol. 139, no. 12, pp. 413–420, 1992. [43] M. Bahr, A. Boag, E. Michielssen, and R. Mittra, “Design of ultra-
[16] K. Deb and D. E. Goldberg, “Sufficient conditions for deceptive and easy broadband loaded monopoles,” in Proc. IEEE Antennas Propagat. Soc.
binary functions,” Ann. Mathem. Artificial Intell., vol. 10, pp. 385–408, Int. Symp., Seattle, WA, June 1994, pp. 1290–1293.
1994. [44] A. Boag, A. Boag, E. Michielssen, and R. Mittra, “Design of electrically
[17] , “Analyzing deception in trap functions,” in Foundations of loaded wire antennas using genetic algorithms,” IEEE Trans. Antennas
Genetic Algorithms 2, L. D. Whitley, Ed. San Mateo, CA: Morgan Propagat., vol. 44, pp. 687–695, May 1996.
Kauffman, 1993. [45] J. M. Johnson and Y. Rahmat-Samii, “Genetic algorithms in electromag-
[18] D. E. Goldberg, “Genetic algorithms and Walsh functions: Part I—A netics,” in Proc. IEEE Antennas Propagat. Soc. Int. Symp., Baltimore,
gentle introduction,” Complex Syst., vol. 3, no. 2, pp. 129–152, 1989. MD, July 1996, pp. 1480–1483.
[19] D. E. Goldberg, K. Deb, and J. H. Clark, “Genetic algorithms, noise, and [46] , “Genetic algorithm optimization and its application to antenna
the sizing of populations,” Complex Syst., vol. 6, no. 3, pp. 333–362, design,” in Proc. IEEE Antennas Propagat. Soc. Int. Symp., Seattle, WA,
1991. June 1994, pp. 326–329.
[20] D. E. Goldberg and K. Deb, “A comparative analysis of selection [47] D. S. Linden and E. E. Altschuler, “The design of Yagi antennas using
schemes used in genetic algorithms,” in Foundations of Genetic Al-
a genetic algorithm,” in Proc. USNC/URSI Radio Sci. Meet., Baltimore,
gorithms, G. J. E. Rawlins, Ed. San Mateo, CA: Morgan Kaufmann,
MD, July 1996, p. 283.
1991. [48] , “Automating wire design using genetic algorithms,” Microwave
[21] D. E. Goldberg, personal communication, 1995.
J., vol. 39, no. 3, p. 74, 1996.
[22] M. J. Buckley, “Linear array synthesis using a hybrid genetic algorithm,”
[49] T. Eisenhammer, M. Lazarov, M. Leutbecher, U. Schoffel, and R.
in Proc. IEEE Antennas Propagat. Soc. Int. Symp., Baltimore, MD, July
Sizmann, “Optimization of interference filters with genetic algorithms
1996, pp. 584–587.
[23] R. L. Haupt, J. J. Menozzi, and C. J. McCormack, “Thinned arrays using applied to silver-based mirrors,” Appl. Opt., vol. 32, no. 31, pp.
genetic algorithms,” in Proc. IEEE Antennas Propagat. Soc. Int. Symp., 6310–6315, 1993.
[50] S. Martin, J. Rivory, and M. Shoenauer, “Simulated Darwinian evolution
Ann Arbor, MI, June 1993, pp. 712–715.
[24] R. L. Haupt, “Thinned arrays using genetic algorithms,” IEEE Trans. of homogeneous multilayer systems: A new method for optical coatings
Antennas Propagat., vol. 42, pp. 993–999, July 1994. design,” Opt. Communicat., vol. 110, pp. 503–506, 1994.
[25] , “An introduction to genetic algorithms for electromagnetics,” [51] A. Tennant and B. Chambers, “Adaptive optimization techniques for
IEEE Antennas Propagat. Mag., vol. 37, pp. 8–15, Apr. 1995. the design of microwave absorbers,” in Proc. Conf. Adaptive Computing
[26] D. J. O’Neill. “Element placement in thinned arrays using genetic Eng. Design Contr., Univ. Plymouth, U.K., Sept. 1994, pp. 44–49.
algorithms,” in Proc. OCEANS ’94, Brest, France, Sept. 1994, pp. [52] E. Michielssen, J.-M. Sajer, S. Ranjithan, and R. Mittra, “Design of
II/301–306. lightweight, broad-band microwave absorbers using genetic algorithms,”
[27] F. Ares, S. R. Rengarajan, E. Villanueva, E. Skochinski, and E. Moreno, IEEE Trans. Microwave Theory Tech., vol. 41, pp. 1024–1031, June/July
“Application of genetic algorithms and simulated annealing technique in 1993.
optimising the aperture distributions of antenna array patterns,” Electron. [53] B. Chambers and A. Tennant, “Design of wideband Jaumann radar
Lett., vol. 32, no. 3, pp. 148–149, 1996. absorbers with optimum oblique incidence performance,” Electron. Lett.,
[28] R. L. Haupt and A. S. Ali, “Optimized backscattering sidelobes from vol. 30, no. 18, pp. 1530–1531, 1994.
an array of strips using a genetic algorithm,” in Proc. Appl. Computat. [54] , “Optimized design of Jaumann radar absorbing materials using a
Electromagn. Conf., Monterey, CA, Mar. 1994, pp. 266–270. genetic algorithm,” Proc. Inst. Elect. Eng. Radar, Sonar, Navigat., vol.
[29] R. L. Haupt, “Optimum quantised low sidelobe phase tapers for arrays,” 143, no. 1, pp. 23–30, 1996.
Electron. Lett., vol. 31, no. 14, pp. 1117–1118, 1995. [55] E. Michielssen, J. M. Sajer, and R. Mittra, “Design of multilayered FSS
[30] , “Comparison between genetic and gradient-based optimization and waveguide filters using genetic algorithms,” in Proc. IEEE Antennas
algorithms for solving electromagnetics problems,” IEEE Trans. Magn., Propagat. Soc. Int. Symp., Ann Arbor, MI, June 1993, pp. 1936–1939.
vol. 31, no. 3, pp. 1932–1935, May 1995. [56] E. Michielssen, A. Boag, J. M. Sajer, and R. Mittra, “Design of fre-
[31] , “Speeding convergence of genetic algorithms for optimizing quency selective surfaces using massively parallel genetic algorithms,”
antenna arrays,” in Proc. 12th Ann. Rev. Progress Appl. Computat. in Proc. URSI Radio Sci. Meet., Seattle, WA, June 1994, p. 441.
Electromagn., Moneterey, CA, Mar. 1996, pp. 742–749. [57] K. Aygun, D. S. Weile, and E. Michielssen, “Design of multilayered
[32] Y. Lu and K. K. Lan, “Genetic algorithms based pattern synthesis strip gratings by genetic algorithms,” Microwave Opt. Tech. Lett., vol.
approach for arbitrary array designs,” in Proc. 12th Ann. Rev. Progress 42, pp. 81–85, Feb. 1997.
Appl. Computat. Electromagn., Monterey, CA, Mar. 1996, pp. 734–741. [58] G. F. Uler, O. A. Mohammed, and C.-S. Koh, “Utilizing genetic
[33] D. Marcano, F. Duran, and O. Chang, “Synthesis of multiple beam linear algorithms for the optimal design of electromagnetic devices,” IEEE
arrays using genetic algorithms,” in Proc. IEEE Antennas Propagat. Soc. Trans. Magn., vol. 30, pp. 4296–4298, Nov. 1994.
Int. Symp., Newport Beach, CA, June 1995, pp. 938–941. [59] J. A. Vasconcelos, L. Karahenbuhl, L. Nicholas, and A. Nicholas,
[34] M. Shimizu, “Determining the excitation coefficients of an array using “Design optimization using the BEM coupled with genetic algorithm,”
genetic algorithms,” in Proc. IEEE Antennas Propagat. Soc. Int. Symp., in Proc. 2nd Int. Conf. Computat. Electromagn., London, U.K., Apr.
Seattle, WA, June 1994, pp. 530–533. 1994, pp. 60–63.
[35] D. Marcano and A. Nieto, “Genetic algorithms for the synthesis of [60] O. A. Mohammed and G. F. Uler, “Genetic algorithms for the optimal
planar arrays,” in Proc. USNC/URSI Radio Sci. Meet., Baltimore, MD, design of electromagnetic devices,” in Proc. 11th Annu. Rev. Progress
July 1996, p. 11. Appl. Computat. Electromagn., Monterey, CA, Mar. 1995, pp. 384–392.
[36] R. Haupt. “Optimization of subarray amplitude tapers,” in Proc. IEEE [61] A. A. Arkadan, T. Sareen, and S. Subramaniam, “Genetic algorithms
Antennas Propagat. Society Int. Symp., Newport Beach, CA, June 1995, for nondestructive testing in crack evaluation,” IEEE Trans. Magn., vol.
pp. 1830–1833. 30, pp. 4320–4322, Nov. 1994.
[37] , “Optimization of aperiodic conducting grids,” in Proc. 11th [62] M. Enokizono and Y. Akinari, “Estimation of current distribution by
Annu. Rev. Progress Appl. Computat. Electromagn., Monterey, CA, Mar. a hybrid of genetic algorithm and sampled pattern matching method,”
1995, pp. 211–215. IEEE Trans. Magn., vol. 31, pp. 2012–2015, May 1995.
WEILE AND MICHIELSSEN: GENETIC ALGORITHM OPTIMIZATION APPLIED TO ELECTROMAGNETICS 353

[63] D. McNay, E. Michielssen, R. L. Rogers, F. A. Taylor, M. Akhtari, [84] D. S. Weile and E. Michielssen, “Integer coded Pareto genetic algorithm
and W. W. Sutherling, “Multiple source localization using genetic design of antenna arrays,” Electron. Lett., vol. 19, pp. 1744–1745, Sept.
algorithms,” J. Neurosci. Methods., vol. 64, pp. 163–172, Feb. 1996. 1996.
[64] F. Thollon and N. Burais, “Geometrical optimization of sensors for eddy [85] D. S. Weile, E. Michielssen, and A. Boag, “Community-based evolution-
currents—Nondestructive testing and evaluation,” IEEE Trans. Magn., ary optimization of frequency selective surfaces,” in Proc. USNC/URSI
vol. 31, pp. 2026–2031, May 1995. Radio Sci. Meet., Baltimore, MD, July 1996, p. 345.
[65] J. M. Johnson and Y. Rahmat-Samii, “Genetic algorithm optimization of [86] R. Mittra, C. H. Chan, and T. Cwik, “Techniques for analyzing frequency
wireless communication networks,” in Proc. IEEE Antennas Propagat. selective surfaces—A review,” Proc. IEEE, vol. 76, pp. 1593–1615,
Soc. Int. Symp., Newport Beach, CA, June 1995, pp. 1964–1967. Dec. 1988.
[66] E. Betensky, “Postmodern lens design,” Opt. Eng., vol. 38, no. 2, pp. [87] A. E. Eiben, E. H. L. Aarts, and K. M. V. Hee, “Global convergence
1750–1756, 1993. of genetic algorithms: On infinite Markov chain analysis,” in Proc. 1st
[67] J. Kim, H.-B. Lee, C. Cheon, H.-S. Kim, H. K. Jung, and S.-Y. Hahn, Int. Conf. Parallel Problem Solving Nature, Dortmund, Germany, Oct.
H
“Numerical design technique for waveguide T-junctions in -plane,” 1991, pp. 4–12.
in Proc. IEEE Antennas Propagat. Soc. Int. Symp., Newport Beach, CA, [88] J. Bosworth, N. Foo, and B. P. Zeigler, “Comparison of genetic
June 1995, pp. 1562–1565. algorithms with conjugate gradient methods,” NASA-CR-2093; NTIS
[68] E. G. Johnson and M. A. G. Abushagur, “Microgenetic algorithm N72–30526, 1972.
optimization methods applied to dielectric gratings,” J. Opt. Soc. Amer. [89] D. E. Goldberg, K. Deb, and B. Korb, “Messy genetic algorithms
A, vol. 12, no. 5, pp. 1152–1160, 1995. revisited: Studies in mixed size and scale,” Complex Syst., vol. 4, pp.
[69] N. Lu, J. Jin, E. Michielssen, and R. L. Magin, “Optimization of RF 415–444, 1990.
coil design using genetic algorithm and simulated annealing method,”
in Proc. 3rd Meet. Soc. Magn. Resonance, Nice, France, Aug. 19–25,
1995, p. 1002.
[70] C.-C. Chiu and P.-T. Lu, “Image reconstruction of a perfectly conducting
cylinder by the genetic algorithm,” Proc. Inst. Elect. Eng. Microwaves, Daniel S. Weile received the B.S. degree in electri-
Antennas Propagat., vol. 143, no. 3, pp. 249–253, June 1996. cal engineering and the B.S. degree in mathematics
[71] P. Ilavarasan, E. J. Rothwell, R. Bebermeyer, K. M. Chen, and D. P. from the University of Maryland, College Park,
Nyquist, “Natural resonance extraction from multiple data sets using a in 1994, and the M.S. degree in electrical engi-
genetic algorithm,” in Proc. IEEE Antennas Propagat. Soc. Int. Symp., neering from the University of Illinois at Urbana-
Seattle, WA, June 1994, pp. 576–579. Champaign, in 1995. He is currently working toward
[72] P. Ilavarasan, E. J. Rothwell, K.-M. Chen, and D. P. Nyquist, “Natural the Ph.D. degree at the University of Illinois at
resonance extraction from multiple data sets using a genetic algorithm,” Urbana-Champaign.
IEEE Trans. Antennas Propagat., vol. 43, pp. 900–904, Aug. 1995. From 1990 to 1994, he was a Technical Aide with
[73] G. Turhan-Sayan, K. Leblebicioglu, and S. Inan, “The use of genetic the Army Space Technology and Research Office
algorithms in input signal shaping for target identification,” in Proc. (ASTRO), Gaithersburg, MD, where he was offi-
USNC/URSI Radio Sci. Meet., Baltimore, MD, July 1996, p. 21. cially commended three times for his work on reflection from inhomogeneous
[74] Q. Li, E. J. Rothwell, K.-M. Chen, and D. P. Nyquist, “Scattering center media. During the summer of 1994, he worked at the Institute of Plasma
analysis of radar targets using fitting scheme and genetic algorithm,” Research, University of Maryland, College Park, working on developing
IEEE Trans. Antennas Propagat., vol. 44, pp. 198–207, Feb. 1996. interactive software for the design of depressed collectors for gyrotron beams.
[75] E. Michielssen and D. Brady, “Control of spatial excitation patterns in Currently, he is a National Science Foundation Fellow and Teaching Assistant
two-level systems using time domain fields,” Opt. Lett., pp. 1931–1933, at the University of Illinois at Urbana-Champaign.
Dec. 1994. Mr. Weile is a member of Eta Kappa Nu, Tau Beta Pi, and Phi Beta Kappa.
[76] E. Michielssen, W. C. Chew, and D. S. Weile. “Genetic algorithm
optimized perfectly matched layers for finite difference frequency do-
main applications,” in Proc. IEEE Antennas Propagat. Soc. Int. Symp.,
Baltimore, MD, July 1996, pp. 2106–2109.
[77] D. S. Weile, “Genetic algorithm applications in electromagnetics,” Eric Michielssen (M’95) received the M.S. degree
Master’s thesis, Univ. Illinois, Urbana-Champaign, Aug. 1995. in electrical engineering from Katholieke Univer-
[78] D. S. Weile, E. Michielssen, and D. E. Goldberg, “Multiobjective siteit, Leuven, Belgium, in 1987, and the Ph.D.
synthesis of electromagnetic devices using genetic algorithms,” in Proc. degree from the University of Illinois at Urbana-
IEEE Antennas Propagat. Soc. Int. Symp., Baltimore, MD, July 1996, Champaign, in 1992.
pp. 592–596. He is an Assistant Professor of electrical and
[79] , “Genetic algorithm design of Pareto optimal broad band mi- computer engineering at the University of Illinois at
crowave absorbers,” IEEE Trans. Electromagn. Compat., vol. 38, pp. Urbana-Champaign and the Associate Director of its
518–524, Aug. 1996. Center for Computational Electromagnetics. From
[80] E. Michielssen, J. M. Sajer, and R. Mittra, “Pareto-optimal design of 1987 to 1988, he was a Research Assistant in the
broadband microwave absorbers using genetic algorithms,” in Proc. Microwaves and Lasers Laboratory at Katholieke
IEEE Antennas Propagat. Int. Symp., Ann Arbor, MI, June 1993, pp. Universiteit, Leuven, Belgium (KUL). He joined the faculty of the Department
1167–1170. of Electrical and Computer Engineering at the University of Illinois as
[81] N. Srinivas and K. Deb, “Multiobjective optimization using nondom- a Visiting Assistant Professor in 1992, and was appointed an Assistant
inated sorting in genetic algortihms,” Evolutionary Computat., vol. 2, Professor of electrical and computer engineering in 1993. He serves as an
no. 3, pp. 221–248, 1995. Associate Editor for Radio Science. His research interests include all aspects of
[82] D. E. Goldberg and J. J. Richardson, “Genetic algorithms with sharing computational electromagnetics, with a focus on fast multilevel algorithms, the
for multimodal function optimization,” in Genetic Algorithms Applicat. application of combinatorial stochastic optimization techniques to the design
Proc. 2nd Int. Conf. Genetic Algorithms, Cambridge, MA, July 1987, of electromagnetic and optical components, and computational photonics.
pp. 41–49. Dr. Michielssen serves as the Technical Chair for the 13th Annual Review
[83] C. K. Oei, D. E. Goldberg, and S. J. Chang, “Tournament selection, of Progress in Applied Computational Electromagnetics, Monterey, CA. He
niching, and the preservation of diversity,” Univ. Illinois, Urbana- was appointed Belgian American Educational Foundation Fellow in 1988 and
Champaign, IlliGAL Tech. Rep. 91011, 1991. is a 1995 NSF CAREER Award recipient.