IEEE TRANSACTIONS ON MAGNETICS, VOL. 42, NO.

4, APRIL 2006 1207

Multiobjective Approaches for Robust

Electromagnetic Design
Frederico G. Guimarães1 , David A. Lowther2 , and Jaime A. Ramírez1
Department of Electrical Engineering, Federal University of Minas Gerais, Belo Horizonte, MG 31270-010, Brazil
Department of Electrical and Computer Engineering, McGill University, Montreal, QC, H3A-2A7 Canada
Robust optimization problems are inherently multiobjective, because the designer searches for compromise solutions between the
mathematical model and possible uncertainties when constructing the physical device. We propose a novel formulation that consists
of the employment of a multiobjective approach for robust design. The proposed methodology is applied to two practical problems:
the design of a loudspeaker and the design of a superconducting magnetic storage device. The results provide more alternatives to the
decision maker, confirming the usefulness and benefit of the approach.
Index Terms—Electromagnetic design, multiobjective optimization, robust optimization.

I. INTRODUCTION The classical formulation does not take into account problem
uncertainties. Thus, small perturbations in the design may lead
R OBUST optimization is a methodology for treating opti-
mization problems in the presence of uncertainties in the
design parameters [1]–[3]. In this situation, a solution less sen-
to unacceptable performance. This characteristic may only be
detected in a posteriori analysis, for example, with sensitivity
sitive to variations around the nominal design is often preferable analysis [2].
to a more sensitive solution, even though its nominal design is
somewhat better. B. Robust Formulation
The treatment of the demand for robustness in electromag- Recent works in the literature are devoted to the robust ap-
netic design is relatively recent. Some approaches involve the proach to tackle uncertainties in optimization problems [4]–[8].
a posteriori analysis of the solution obtained by the optimiza- A robust solution may be characterized as 1) a solution that is
tion process [2]. In order to perform an online robustness anal- insensitive to variations of the nominal design and 2) is feasible
ysis, some works adopt the estimation of the worst-case de- within an uncertainty range around the nominal design. In this
sign [4], [6], [7], and other works employ statistical analysis section, only unconstrained problems are considered. We dis-
[5], [8]. However, these approaches do not treat the problem as cuss robust constraints further on.
one of conflicting design. Robust optimization problems are, in Typical methodologies for robust optimization aim to mini-
essence, multiobjective because we search for compromise so- mize a given robust criterion. This robust criterion may be de-
lutions between the mathematical and computational model and fined with respect to statistical quantities as the mean value
the physical implementation of the final device. and the variance of the objective function.
This paper discusses a novel approach for robust electromag- For example, one may minimize the expected mean and de-
netic design, which consists of the employment of a multiob- fine a constraint for the variance
jective point of view for robust design. The main advantage of
the multiobjective approach is to provide more robust solutions
to the designer than typical methodologies, for example, using (2)
the worst-case analysis. The suggested formulation includes the
where is a level defined by the designer.
worst-case approach as one of the robust solutions achieved.
Another approach is to minimize a combination of the mean
and the variance
II. MATHEMATICAL FORMULATION
(3)
A. Classical Formulation
A general optimization problem may be stated as with .
Finally, a more flexible approach is to treat the problem in a
multiobjective way
(1)
(4)
where is the vector of optimization variables,
is the objective function, and
In fact, the formulations (2) and (3) are particular solutions of
are the problem constraints.
(4). Nevertheless, the computation of the mean and variance of
the objective function in electromagnetic problems is computa-
tionally very expensive. Due to this characteristic, some alter-
Digital Object Identifier 10.1109/TMAG.2006.871573 native formulations are adopted. In [1], is represented
0018-9464/$20.00 © 2006 IEEE
1208 IEEE TRANSACTIONS ON MAGNETICS, VOL. 42, NO. 4, APRIL 2006

Fig. 1. Worst vertex prediction. Fig. 2. Illustration of a robust feasible point.

by the nominal value of the objective function and is where is a specified level. However, the statistical analysis
approximated by may be prohibitive in electromagnetic design. For this reason,
the worst-case approach may be taken
(5)
(10)
in which is the maximum perturbation along the th variable. The worst vertex prediction may be used to get the worst value
Another robust formulation in electromagnetic design is for the constraint. Fig. 2 depicts the effect of the worst-case
given by the worst-case analysis within the uncertainty set [4]. value in the feasibility of a point. All the rectangular uncertainty
Therefore, the robust criterion is set must be feasible to consider that the point is feasible.
(6)
III. MULTIOBJECTIVE APPROACHES
where is the uncertainty set. In practice, it is a rectangular This section details the multiobjective approach for dealing
set centered at . Again, due to the computational burden in with the requirement for robustness: the nominal value and de-
electromagnetic optimization problems, only the vertices of the viation approach.
uncertainty set are considered and the worst vertex may be pre-
dicted using [4] A. Nominal Value and Deviation Approach
The multiobjective problem stated in (4) may be solved by
.. (7) adopting (5), but this equation needs the calculation of the gra-
.
dient at the nominal point.
Another way to view the problem is to consider the following
where is the th column of the identity matrix, and is a
multiobjective optimization problem:
nominal value (see Fig. 1).
Thus, the evaluation of the robust criterion consumes (11)
evaluations of the objective function, which is less than the
number of evaluations required by statistical analysis. The Notice that this formulation is closely related to that in (4).
problem (6) may be rewritten as The nominal value of the objective function may be viewed as
the mean value. The second objective is the devi-
(8)
ation from the nominal value, which is related to the variance.
in which is the worst value predicted with (7). Thus, this formulation is similar to that in (4) and it is compu-
The robust criterion is not continuously differentiable, tationally acceptable.
even if this characteristic exists in the original problem. Further-
more, the robust solution—the solution of —will differ B. Extension to Multiobjective Problems
from the solution of only in some multiminima problems. If the original problem is already a multiobjective problem,
Therefore, a multimodal optimization method that does not rely and the robust formulation is needed, the multiobjective ap-
on derivatives is recommended. proach may also be applied. For example, consider the problem

C. Robust Constraints
With the statistical analysis, the original constraints of (1) (12)
need to be substituted for with , and is the vector of objectives
(9) to be minimized.
GUIMARÃES et al.: MULTIOBJECTIVE APPROACHES FOR ROBUST ELECTROMAGNETIC DESIGN 1209

TABLE I
LOUDSPEAKER PARAMETERS, BEST NOMINAL
SOLUTION, AND MOST STABLE SOLUTION

Fig. 3. Two-dimensional axisymmetric visualization of the loudspeaker

configuration.

Fig. 5. Pareto front found by a multiobjective genetic algorithm using the

nominal value and deviation approach.

TABLE II
BEST NOMINAL SOLUTION AND MOST STABLE SOLUTION
FOR NOMINAL VALUE AND DEVIATION APPROACH
Fig. 4. Pareto front found by a multiobjective genetic algorithm. Remember
that the flux density must be maximized and the variance minimized. The best
value for the objective function is associated with a higher sensitivity.

The robust formulation using the nominal value and deviation

approach is given by

(13)

where is the vector with the worst-case values. However,

this robust formulation doubles the number of objective func- minimizes the variance. For , we get T and esti-
tions. mated deviation . On the other hand,
the most robust solution gives T and
IV. RESULTS .
In order to avoid the computation of derivatives in (5), one
A. Loudspeaker Design may employ the nominal value and deviation approach defined
The loudspeaker design investigated in [9] is used here as an in (11). The value is used again, and
example of a manufactured device subjected to variations of the the Pareto front obtained is shown in Fig. 5.
design variables. The problem contains seven design variables The solutions belonging to the extremum points of each
and consists of maximizing the magnetic flux density through Pareto front are detailed in Table II. In this Table, and
the air gap (see Fig. 3). The problem has no constraints. are as before. and are the extremum points of the Pareto
We assume and use the multiob- front for the nominal value and deviation approach. is the
jective approach with the nominal value and variance estimated solution that maximizes the flux density and is the most
from (5). The final Pareto front obtained by a multiobjective ge- stable solution. However, none of these extrema may be useful
netic algorithm is shown in Fig. 4. for the designer and the other solutions shown in Figs. 4 and 5
Table I summarizes the information about each design vari- provide alternative designs. In this manner, the decision maker
able. It also illustrates the best nominal solution , which max- may choose the best compromise between the best nominal
imizes the flux density, and the most robust solution , which design and the most robust one.
1210 IEEE TRANSACTIONS ON MAGNETICS, VOL. 42, NO. 4, APRIL 2006

TABLE III also of the objective function. These values are used for com-
PARAMETERS FOR THE PROBLEM 22 puting the worst penalized value. In this manner, we can employ
the nominal value and deviation approach associated with a
multiobjective genetic algorithm. Using this methodology, we
have obtained the Pareto front shown in Fig. 6, in which we can
see that the best nominal designs are characterized by higher
deviations from the nominal value.
The computational cost involved here is less expensive than
what would be required by statistical analysis and is in accor-
dance with the cost taken by stochastic methods applied to mul-
tiobjective problems, as discussed in [2]–[9].

V. CONCLUSION
Robust optimization problems are inherently multiobjective
ones and the formulation proposed in this work incorporates
the multiobjective approach to tackle the demand for robustness
in electromagnetic design. The proposed methodology has the
benefit of providing more alternatives to the decision maker,
not only the worst-case analysis, but also all the compromise
solutions between the nominal value and solution stability.

ACKNOWLEDGMENT
Fig. 6. Pareto front obtained using the nominal value and deviation approach.
This work was supported in part by the National Council
of Scientific and Technologic Development—CNPq—under
B. SMES Design Grants 300 353/1997-9 and 141 731/2004-4 and in part by
Finally, we consider the design of the SMES system in the Fapemig under Grant TEC-941/05, Brazil.
TEAM workshop benchmark problem 22, in order to illustrate
the proposed methodology with a constrained problem. This is REFERENCES
a well-known problem and the details are given in [10]. We have [1] S. B. Yoon, I. S. Jung, and D. S. Hyun, “Robust shape optimization
examined the full version of the problem with eight parameters of electromechanical devices,” IEEE Trans. Magn., vol. 35, no. 3, pp.
1710–1713, May 1999.
and with uncertainties in the current densities as in [4], [7]. See [2] R. H. C. Takahashi, J. A. Ramírez, J. A. Vasconcelos, and R. R. Sal-
Table III for further details. danha, “Sensitivity analysis for optimization problems solved by sto-
Additionally, we transform the equality constraint for the pre- chastic methods,” IEEE Trans. Magn., vol. 37, no. 5, pp. 3566–3569,
Sep. 2001.
scribed stored energy value into a new inequality by specifying [3] K. H. Lee and G. J. Park, “Robust optimization considering tolerances
that the energy must be within the interval %. This of design variables,” Comput. Struct., vol. 79, pp. 77–86, 2001.
is because we can not guarantee the satisfaction of an equality [4] P. Alotto, C. Magele, W. Renhart, A. Weber, and G. Steiner, “Robust
target functions in electromagnetic design,” COMPEL, vol. 22, no. 3,
constraint under parameter uncertainty. Therefore, the formula- pp. 549–560, 2003.
tion adopted for the problem is given by [5] V. Cavaliere, M. Cioffi, A. Formisano, and R. Martone, “Robust design
of high field magnets through Monte Carlo analysis,” COMPEL, vol. 22,
(14) no. 3, pp. 589–602, 2003.
[6] G. Spagnuolo, “Worst case tolerance design of magnetic devices by evo-
lutionary algorithms,” IEEE Trans. Magn., vol. 39, no. 5, pp. 2170–2178,
subject to Sep. 2003.
[7] G. Steiner, A. Weber, and C. Magele, “Managing uncertainties in elec-
tromagnetic design problems with robust optimization,” IEEE Trans.
(15a) Magn., vol. 40, no. 2, pp. 1094–1099, Mar. 2004.
[8] M. Cioffi, A. Formisano, and R. Martone, “Stochastic handling of toler-
ances in robust magnets design,” IEEE Trans. Magn., vol. 40, no. 2, pp.
(15b) 1252–1255, Mar. 2004.
(15c) [9] K. Rashid, J. A. Ramírez, and E. M. Freeman, “A general approach for
extracting sensitivity analysis from a neuro-fuzzy model,” IEEE Trans.
Magn., vol. 36, no. 3, pp. 1066–1070, Jul. 2000.
with MJ. [10] TEAM Workshop Problem 22: SMES Optimization Benchmark [On-
The constraints are treated with the penalty approach, in line]. Available: http://www-igte.tu-graz.ac.at/archive/team/index.htm
which the objective function value is penalized by each con-
straint violation. The worst vertex prediction is utilized for Manuscript received June 20, 2005 (e-mail: [email protected];
obtaining the worst-case value of each inequality constraint and [email protected]).