1, June 2007
Brazilian Microwave and Optoelectronics Society-SBMO Received 10 Aug., 2006; revised 12 Feb., 2007; accepted 23 March, 2007
ISSN 1516-7399 2007 SBMO
167
Abstract The aim of this work is to compare and investigate
mathematical methodologies to deal with multiobjective
optimization problems. Four methods are considered: the Weight
Sum Approach, the Nondominated Sorting Genetic Algorithm
(NSGA), the Strength Pareto Evolutionary Algorithm (SPEA) and
the Pareto Archive Evolutionary Strategy (PAES). The work
proposes some indexes in order to compare the performance of the
algorithms. An analytical benchmark problem was investigated to
validate the methods and their implementations. For real-world
problems, the electromagnetic field computation was performed by
the finite element method and Kriging models were used as
surrogate functions to approximate the objective functions of the
problems. Three real-world problems were analyzed: the first one
consists in the design of a die press model, the second problem deals
with the design of superconducting magnetic energy storage
(SMES) and the third one dealt with the design of a special
reluctance motor.
Index Terms Optimization Methods, Finite Element Method
I. INTRODUCTION
Multiobjective optimization methods deal with the conflict of the requirements of an
electromagnetic design, instead of mono-objective optimization method, that deals only with a
maximization (or minimization) of one design requirement. In this paper, we will compare some
multiobjective optimization methods and propose indexes to measure the performance of a
multiobjective optimization method on the Pareto set achievement. These indexes must reflect two
important attributes: the Pareto set must converge to an optimal-Pareto set and the diversity of the
Pareto set must be as high as possible.
II. THE MULTIOBJECTIVE OPTIMIZATION METHODS
Four multiobjective optimization methods will be analyzed: the Weighted Sum Method, the Non-
dominated Sorting Genetic Algorithm (NSGA) [1], the Strength Pareto Evolutionary Algorithm
(SPEA) [2] and Pareto Archived Evolutionary Strategy (PAES) [3]. In this section, we will briefly
describe these methods.
A. Weighted Sum Method
The Weighted Sum Method is, by far, the most intuitive approach to solve a multiobjective
optimization. Instead of deal with the conflict of the requirements, a weighted sum of the functions is
performed and a standard mono-optimization is performed. For each set of weights one point of the
Multiobjective Optimization Applied to
Electromagnetic Devices
Fbio Augusto Baria Perdiz and Luiz Lebensztajn
Escola Politcnica da Universidade de So Paulo
perdiz@gmail.com, leb@pea.usp.br
Journal of Microwaves and Optoelectronics, Vol. 6, No. 1, June 2007
Brazilian Microwave and Optoelectronics Society-SBMO Received 10 Aug., 2006; revised 12 Feb., 2007; accepted 23 March, 2007
ISSN 1516-7399 2007 SBMO
168
Pareto set is achieved.
B. Nondominated Sorting Genetic Algorithm
The selection operator on NSGA [1] differs from the usual genetic algorithm. It ranks the
population and identifies the nondominated individuals. They are assumed to be the first
nondominated front and a large dummy fitness value is assigned to them. To preserve diversity in the
population, a fitness sharing process is applied to these individuals.
The process divides the fitness by a quantity proportional to the number of individuals around it.
This process punishes the individuals located in high populated areas.
The fitness value degradation of neighbor individuals may be executed using (1) e (2), where the
parameter d
ij
is the variable distance between two individuals i e j, and
shared
is the maximum distance
allowed between any two individuals to become members of the same niche. The term df
i
is the
dummy fitness value assigned to individual i. The df
'i
is its corresponding shared value. N
pop
is the
size of the population.
( )
<
=
shared ij
shared ij
shared
ij
ij
d if
d if
d
d Sh
, 0
, 1
2
(1)
and
1
1
) (
=
=
pop
N
j
ij i i
d Sh df df
(2)
The dominated individuals are now assumed to constitute a new population who will suffer the
same process. The new nondominated front is assigned a dummy fitness value, which is kept smaller
than the minimum shared dummy fitness of the previous front. The process continues until the entire
population is classified into fronts.
The individuals assigned share values are used as parameter for the stochastic remainder
proportional selection for the crossover and mutation operations. During the selection, the best ones
(higher share values) get more copies. This results in quick convergence of the population towards
nondominated regions and sharing helps to distribute it over this region.
C. Strength Pareto Evolutionary Algorithm - SPEA
The SPEA [2] stores the nondominated solutions found so far externally. A fitness assignment
method is used, where one population serves as basis for anothers evaluation using the concept of
Pareto dominance. Clustering is performed to reduce the Pareto-optimal front to the allowed maximal
size, preserving its original characteristics. A Pareto-based niching method is provided in order to
preserve diversity in the population.
SPEA starts updating the Pareto set. All nondominated solutions in the population are copied to the
Pareto set, and possible dominated ones are removed from it. This set is called the extended Pareto
set, and if its size exceeds a given maximum, a reduction clustering base method is applied. The
Journal of Microwaves and Optoelectronics, Vol. 6, No. 1, June 2007
Brazilian Microwave and Optoelectronics Society-SBMO Received 10 Aug., 2006; revised 12 Feb., 2007; accepted 23 March, 2007
ISSN 1516-7399 2007 SBMO
169
population is used to calculate the strength (fitness). In a first step each solution in the Pareto set is
assigned a value s [0,1) called strength. S is proportional to the number of population members
covered. In the second step, for each individual in the population, the strengths of all external Pareto
solutions by which it is covered are summed up. Adding one to this result, the fitness value is
obtained. This guarantees that Pareto solutions are most likely to be reproduced.
Finally, both population and external Pareto set are considered for reproduction. A binary
tournament selection is applied: the smaller the fitness, the higher the probability to be selected.
Crossover and mutation are then used to generate the new population.
D. Pareto Archived Evolution Strategy - PAES
The algorithm proposed by Knowles and Corne [3] begins with the initialization of a single
chromosome (the current solution), which is evaluated using the objective functions. A mutation
operator is then applied to current solution to obtain a copy. This new individual is evaluated and is
called the new candidate solution.
The acceptance criterion is straightforward, when one solution dominates the other, but when there
is no dominance between the solutions, the candidate is compared to a reference population, called the
archive. On the acceptance criterion, the algorithm takes into account the location of the solution:
there is a preference to solutions located on the least crowded regions.
So the archive stores all the nondominated generated solutions and helps the accurate selection
between the candidate and the current solution. The result of this process is a front closer to the
optimal solution and more diverse. This diversity is achieved thanks to a d-dimensional grid, where d
is the number of objectives.
III. THE INDEXES AND THE PROBLEMS
The first index was created to show the convergence of the Pareto-set to an optimal-Pareto set. It
measures the non-dominated area under the Pareto surface. The second index is linked with the Pareto
front diversity. The individuals that belong to the Pareto set are ordered according to one function.
Then the distance between neighbors is performed and the variance of the distances is calculated. If
the distribution on the Pareto set is uniform then a low variance is achieved. The third one is
associated to the coverage concept and shows that the outcomes of one algorithm dominate the
outcomes of another algorithm, although it does not tell how much better it is.
A. The Analytical Function
The analytical problem was proposed by Deb [4] and deals with the minimization of two functions,
1 1 1
x ) x ( f = (3)
( ) ( ) g / f 1 ) 1 m /( x 9 1 ) x ,..., x , x ( f
1
m
2 i
i m 2 1 2
+ =
=
(4)
where m=30 e x
i
[0,1].
Journal of Microwaves and Optoelectronics, Vol. 6, No. 1, June 2007
Brazilian Microwave and Optoelectronics Society-SBMO Received 10 Aug., 2006; revised 12 Feb., 2007; accepted 23 March, 2007
ISSN 1516-7399 2007 SBMO
170
B. TEAM Workshop Problem 25
The geometry of TEAM Workshop Problem 25 is shown on Fig.1. The goal is to obtain a
homogeneous radial magnetic induction distribution on the path e-f. Four parameters (R1, L2, L3 and
L4) could be changed in a specified range without any constraint. The complete description of the
problem is presented in [5], so it is omitted here.
Fig. 1. Team Workshop problem 25: die press model
The homogeneity of the magnetic distribution along the path, if the index p shows calculated points
and the index o prescribed points, could be written as:
( ) ( ) { } + =
=
10
1 i
2
yxio yip
2
xio xip
B B B B W
(5)
It is possible to define the local quality of the induction magnetic distribution, as proposed on [6].
The error on the amplitude is defined as 100 / ) ( max
max
=
p o p B
B B B E and the error on the angle of
magnetic induction vector is computed by
0 max
max
B Bp
E