IEEE TRANSACTIONS ON MAGNETICS, VOL. 39, NO. 5, SEPTEMBER 2003 2163
Hybrid Genetic Algorithm for Electromagnetic
Topology Optimization
Chang-Hwan Im, Hyun-Kyo Jung, Senior Member, IEEE, and Yong-Joo Kim
Abstract—This paper proposes a hybrid genetic algorithm
(GA) for electromagnetic topology optimization. A two-dimen-
sional (2-D) encoding technique, which considers the geometrical
topology, is first applied to electromagnetics. Then, a 2-D ge-
ographic crossover is used as the crossover operator. A novel
local optimization algorithm, called the on/off sensitivity method,
hybridized with the 2-D encoded GA, improves the convergence
characteristics. The algorithm was verified by applying it to
various case studies, and the results are presented herein.
Index Terms—Genetic algorithm (GA), geographic crossover,
local optimization, topology optimization, two-dimensional (2-D)
encoding.
I. INTRODUCTION
S
TOCHASTIC optimization algorithms have been widely
used for various electromagnetic optimization problems,
in particular to optimize the geometrical dimensions of various
electromagnetic devices [1]–[4]. The genetic algorithm (GA),
simulated annealing (SA), evolution strategy (ES), and tabu
search algorithm, etc. are all well-known examples of the sto-
chastic algorithms. Unfortunately, they have not been applied
much to direct shape optimization or topology optimization
problems [5], because they carry a severe computational cost
and have difficulty in dealing with large numbers of design
variables. Instead, sensitivity analysis has attracted much
interest because it can deal with large amounts of design
variables quickly and effectively [6]–[8]. However, it has the
problem that a solution may converge on a local minimum, due
to its natural deterministic characteristic that stems from the
derivatives of the objective functions.
In this paper, a hybrid GA for electromagnetic topology opti-
mization is proposed. The GA adopts a two-dimensional (2-D)
encoding [9] to represent the geometrical topology effectively.
A multidimensional geographic crossover is applied to increase
the diversity of the population [10]–[12]. The effectiveness of
this approach has already been verified through its application
to computer science [10], [11] and very large scale integration
(VLSI) circuit design [12]. However, its application to topology
optimization is totally new, not only in electromagnetic opti-
mizations, but also in mechanical ones.
In this paper, to improve the accuracy and effectiveness of the
“conventional” GA, a novel local optimization algorithm called
the on/off sensitivity method is proposed and hybridized with
a 2-D encoded GA. The proposed local optimization method
Manuscript received July 23, 2002; revised May 19, 2003.
C.-H. Im and H.-K. Jung are with the School of Electrical Engineering and
Computer Science, Seoul National University, Seoul 151-742, Korea (e-mail:
ichich2@snu.ac.kr; hkjung@snu.ac.kr).
Y.-J. Kim is with the Korea Electrotechnology Research Institute (KERI),
Kyungsangnam-do 641-120, Korea (e-mail: yjkim@keri.re.kr).
Digital Object Identifier 10.1109/TMAG.2003.817094
(a)
(b)
Fig. 1. Comparison of 1-D and 2-D encoding from the viewpoint of
schemata. (a) 2-D encoding and its schemata. (b) Equivalent 1-D encoding and
its schemata. (b) is a 1-D representation of (a). We can see from the figures that
the 1-D encoding may lose its neighboring information, as shown in (b).
assists the main GA to generate superior solutions, and at the
same time, the main GA perturbs the local solutions and pre-
vents them from converging to a local minimum too prema-
turely. An appropriate combination of the two algorithms can
accelerate the convergence speed and minimize the possibility
of being trapped in a local optimum.
To verify the proposed algorithm, three case studies were
considered, including a simplified magnetoencephalography
(MEG) source localization problem, a current source op-
timization problem, and a brushless DC motor (BLDCM)
optimization for reduced cogging torque. Through the case
studies, it will be shown that the proposed method can yield
reasonable solutions with high efficiency.
II. M
ETHODOLOGY
A. 2-D Encoding for GA
A linear string is a symbolic feature of GAs, and most GA im-
plementations have been based on linear encodings [13]. To fit
into linear strings, solutions are encoded into one-dimensional
(1-D) chromosomes, as shown in Fig. 1(b). However, if 2-D ge-
ometry is embedded into a 1-D array, considerable distortion
of neighboring information is unavoidable, as we can see from
Fig. 1(a) and (b). Due to the broken schemata, the 1-D encoding
has always suffered from its poor convergence characteristic.
Cohoon and Paris first proposed 2-D encodings and demon-
strated moderate success in solving such a problem [9]. This
approach led to the development of various crossover methods,
which will be described in Section II-B.
B. 2-D Crossover Methods
Crossover is the most representative and important feature of
GAs, which is also true of 2-D GAs. Cohoon and Paris, who
first proposed the 2-D encoding, used a rectangular crossover
0018-9464/03$17.00 © 2003 IEEE
2164 IEEE TRANSACTIONS ON MAGNETICS, VOL. 39, NO. 5, SEPTEMBER 2003
(a) (b)
(c)
Fig. 2. Crossover methods. (a) Cohon and Paris—rectangular boxes were
used. (b) Anderson—parallel lines were used. (c) Bui and Moon—more natural
lines were used. The white region represents genes from parent 1 and the gray
region represents those from parent 2. The recent method (c) can generate more
diverse solutions than the others.
[9]. In their approach, given two parent chromosomes, the algo-
rithm chooses a small rectangle from one parent and copies the
genes in the rectangle into the offspring; the remainder of the
genes are copied from the other parent, as shown in Fig. 2(a).
Anderson et al. suggested a block-uniform crossover on a 2-D
matrix chromosome that tessellated the chromosome into
blocks; for each block, the genes in the block are copied from a
uniformly selected parent, as shown in Fig. 2(b) [14]. It is evi-
dent that the block uniform crossover has a greater opportunity
to generate more diverse offspring.
Most recently, Bui and Moon proposed a geographic
crossover to increase the diversity of offspring [10]. It general-
ized the conventional block uniform crossover and introduced
natural lines, as shown in Fig. 2(c). It was proven that the whole
encoding space could always be divided into two separated
regions, as shown in the figure. The offspring can be generated
by alternatively copying the intervals of two parent strings. It
was also proven that the geographic crossover could generate
more diverse offspring than any of the conventional crossovers
[10]–[12]. Because it is somewhat difficult to implement
the method practically, we will introduce a relatively simple
process used in this paper.
Step 1) First, one should bear in mind that there can be
six possible cutting lines: a) top-to-right; b) top-to-
left; c) bottom-to-left; d) bottom-to-right; e) top-to-
bottom; and f) left-to-right. The whole space is sepa-
rated into two regions by one cutting line. Then, one
should define a rule to assign 0 or 1 to each region,
for all cases a)–f). Fig. 3(a)–(d) shows the examples
of 0/1 assignment.
Step 2) After generating all the cutting lines, the assigned
numbers are summed up for every gene as in
Fig. 3(e).
Step 3) If the summed number has an odd value, the off-
spring copies the gene from parent 1, and vice versa.
This process is applied to all genes. Then, as shown
in Fig. 3(f), we can see that the whole space is sepa-
rated into two regions very effectively.
Fig. 3. Example of regional separation process: (a)–(d) show examples
of basic cutting lines and rules for assigning region number. The assigned
numbers for each region are summed up as shown in (c). Then, the whole
region can be separable into two groups by checking if the summed numbers
are odd or even, as shown in (f).
C. New Local Optimization Algorithm—On/Off Sensitivity
In this paper, a new local optimization algorithm for electro-
magnetic topology optimization is proposed. In essence, elec-
tromagnetic topology optimizations optimize the distribution
of materials such as current sources, magnetic materials, etc.
Hence, binary encoding is appropriate to represent whether such
materials are present or not. In this paper, 1 represents the pres-
ence of a material and 0 represents the absence of it. The pro-
posed local optimization algorithm adopts the concepts of both
a sensitivity analysis and SA. The algorithm procedure is as
follows.
Step 1) Calculation of On/Off Sensitivity. Change the state
of each gene—if the state is 0 (off), replace it by 1
(on), and vice versa. Then, check the variation (sen-
sitivity) of cost value. Higher sensitivity of a gene
implies that the change of state can improve the cost
value much more.
Step 2) Change State of Genes. After calculating the sensi-
tivity of all genes, change the state of some genes,
selectively. More sensitive genes are selected for
IM et al.: HYBRID GA FOR ELECTROMAGNETIC TOPOLOGY OPTIMIZATION 2165
mutation and the number of genes selected for mu-
tation is defined as follows:
(1)
where
is the number of genes that change their
states
, is the number of genes
that have positive sensitivity values, and
is a prob-
ability between 0 and 1, which will be explained
again at the end of this section.
Step 3) Annealing. Check the cost value after the mutation.
If the cost value is not improved at all, decrease the
value of
using the following annealing scheme:
new old (2)
Any value can be used instead of 0.85. In this paper,
0.85 is selected because it is usually used for conven-
tional annealing processes. Repeat Step 2)–Step 3)
until some improvementofthe cost valueis detected.
Step 4) Repeat Step 1)–Step 3) until the cost value is not
improved any more.
Fig. 4 shows an example of the process of the local optimiza-
tion algorithm. Assume that Fig. 4(a) is an optimal solution and
Fig. 4(b) is an initial one. The calculated sensitivity profile is
shown in Fig. 4(c). The positive sensitivity implies that cost
value is improved when a gene changed its state, and vice versa.
If the most sensitivegene changes its state, as shownin Fig. 4(d),
it is evident that the cost value is improved. However, it seems
somewhat inefficient because the cost value can be more im-
proved by changing two or more genes’ states. That is why we
adopted a concept from SA. If the probability
in (1) is set
to be 1, the solution is changed, as shown in Fig. 4(e). From
several simulations, it was observed that the cost value usually
becomes worse (not improved at all) when too large numbers of
genes are changed, as shown in Fig. 4(e). When the probability
is decreased using (2), an improved solution can be obtained,
as shown in Fig. 4(f). Then, the sensitivity computation is per-
formed again, of which the result is shown in Fig. 4(g). From the
sensitivity profile, we can find the optimal solution as shown in
Fig. 4(h). In Step 2), the initial
should be a large value close
to 1 because there is no process to increase the probability. The
initial
is fixed throughout the whole iteration.
The proposed local optimization algorithm sometimes
yielded very good characteristics, even without GA. In most
cases, however, it converged to a local optimum, unless a good
approximation of the initial solution was given.
D. Whole Procedure of the Hybrid GA
Usually, there are two main classes in hybrid GAs: Lamar-
ckian and Baldwinian GAs. The Lamarckian GAs update the
chromosome after the local optimizations, whereas the Bald-
winian GAs do not update the chromosome after the local op-
timizations, and the local optimizations are only used as a fit-
ness evaluation. Generally,the Lamarckian GAs havebeen more
widely accepted and the Baldwinian GAs’ use has been some-
what restricted [15], [16]. In this paper, we adopted the concepts
of the Lamarckian GAs.
The proposed local optimization algorithm requires many
function calls and consumes extensive computational time.
Fig. 4. An example to illustrate the process of the proposed local optimization
algorithm (see text).
Moreover, excessive use of local optimization can degrade
the diversity of the population. Therefore, in this paper, the
proposed local optimization algorithm was not applied to every
iteration or population. If a random number generated between
0 and 1 is larger than a predetermined probability, the on/off
sensitivity is applied to some superior population. Fig. 5 shows
the whole procedure of the proposed GA. In the procedure,
a general mutation method is applied only when a random
number between 0 and 1 is smaller than a predetermined prob-
ability, which perturbs a small fraction of the offspring. The
predetermined probability of 0.015 is generally used; however,
in this paper, larger values are used to alleviate premature
convergence due to local optimization. After the mutation
or crossover, worst solutions in the original population are
replaced with newly generated ones. The GA procedure is
terminated when 80% of the solutions have the same value.
2166 IEEE TRANSACTIONS ON MAGNETICS, VOL. 39, NO. 5, SEPTEMBER 2003
Fig. 5. Flowchart of the hybrid GA.
Fig. 6. Description of simulation model. Case study I: MEG source
reconstruction.
III. SIMULATION AND RESULTS
The proposed algorithm was applied to three case studies: 1)
a MEG source localization problem; 2) a current source opti-
mization problem; and 3) a BLDCM optimization for reduced
cogging torque.
A. Case Study I—MEG Source Localization Problem
Reconstructing electric activity inside a brain using magnetic
measurements outside the head has attracted a great deal of in-
terest and is usually referred to as the MEG source reconstruc-
tion problem [17], [18]. A simplified MEG model was selected
for the verification of the optimization algorithm. Fig. 6 shows
the schematic view of the simulation model. Sixteen
-direc-
tional superconducting quantum interference device (SQUID)
magnetometers, capable of measuring parallel magnetic fields,
were assumed. The magnitudes of the current dipole moments
at 144 (12
12) rectangular meshes needed to be reconstructed.
The magnitudes of the moments were fixed at a constant value
and their directions were perpendicular to the source surface.
Fig. 7 shows the exact (forward) solution that should have been
reconstructed.
The main reason in selecting this model was that the forward
solutions could be easily obtained using an analytic approach.
1
1
Some kinds of reconstruction problems such as a crack detection problem
(materials to be reconstructed:air and conductor) can be classified as a topology
optimization problem. Generally, the MEG source reconstruction problem is
not such a problem, because the magnitude of each dipole cannot be determined
a priori. Therefore, we simplified the problem to verify the optimization
algorithm.
Fig. 7. Exact (forward) solution that is to be reconstructed.
TABLE I
C
ONDITIONS FOR EXECUTING GA
Fig. 8. Best reconstructed solution after20 repeated executions of GA without
local optimization.
The magnetic field induced by a current dipole moment can be
calculated using Biot–Savart’s law
(3)
where
is the current dipole moment vector, is the positional
vector of the sensor, is the positional vector of the dipole,
and
is the measured magnetic field at the sensor.
To apply GA, the objective function (cost) is defined to mini-
mize the difference between the measured (forward) and recon-
structed magnetic field at each sensor position.
Although the problem may seem very simple, GA has hardly
been applied to such a problem because the total number of pos-
sible cases is so high; in this case 2
. For this reason, a general
GA using a 1-D linear array and a general GA using 12 separated
arrays failed to find the exact solution. However, the GA using
a 2-D geographic crossover did succeed in finding a very close
solution—of course, it was not exact. Table I shows the con-
ditions required to execute the GA without local optimization.
Fig. 8 shows the best reconstructed solution after 20 repeated
applications of the GA without local optimization.
IM et al.: HYBRID GA FOR ELECTROMAGNETIC TOPOLOGY OPTIMIZATION 2167
Fig. 9. Best reconstructed solution after 20 repeated executions of GA with
local optimization.
TABLE II
C
OMPARISONS OF CAPABILITY OF GAS WITH AND WITHOUT
LOCAL
OPTIMIZATION ALGORITHM AFTER 20 REPEATED EXECUTIONS (CASE I)
All conditions, which were applied to the proposed GA using
local optimization, are given in Table I. The local optimization
started to work after the 50th GA iteration; not from the start,
because impatient use of the local optimization may have been
harmful to the creation of the schemata. If a random number
between 0 and 1 was larger than 0.8, then local optimization
was applied to 100 superior solutions. Fig. 9 shows the best
reconstructed solution after 20 repeated applications of the GA
with local optimization.
To verify the results quantitatively, best, worst, and average
values of the objective function were compared. The results are
given in Table II. From the table, the superiority of the GA with
local optimization can be easily confirmed.
B. Case Study II—Source Optimization Problem
Magnetic resonance imaging (MRI) coil design is a typical
application of the source optimization problem [19], [20].
Fig. 10 shows the numerical model under consideration and the
initial magnetic field distribution. The problem was to make
the
-directional flux density in a measuring box constant.
A12
14 (totaling 168 variables) 2-D matrix was used to
represent the distribution of the source coils. The objective
function that needed to be minimized was defined as
(4)
where
represents -directional magnetic flux density at a
measuring point [in Fig. 10(b)], and is the standard devia-
tion of the
-directional flux density at all measuring points.
Both GAs—with and without local optimization—were ap-
plied to the topology optimization, and the operating condi-
tions were the same as in Table I. An analytic solution based on
Fig. 10. Initial magnetic field distribution (a) in whole numerical model and
(b) in measuring domain. The FEM was used to illustrate the flux distribution.
Fig. 11. Best reconstructed solution after 20 repeated executions of GA
without local optimization.
Biot–Savart’s law was used for the magnetic field calculation.
Figs. 11 and 12 show the best reconstructed source distributions
after 20 repeated executions of GA with and without local op-
timization, respectively. Fig. 13 shows the magnetic field dis-
tribution in a measuring box by the GA with local optimiza-
tion. Table III shows the comparison of best, worst, and average
values of the objective function (4) after 20 repeated executions.
