Engineering Optimization

ISSN: 0305-215X (Print) 1029-0273 (Online) Journal homepage: https://www.tandfonline.com/loi/geno20

Reliability-based robust multi-objective

optimization applied to engineering system design

Fran Sérgio Lobato, Márcio Aurelio da Silva, Aldemir Ap Cavalini Jr & Valder
Steffen Jr

To cite this article: Fran Sérgio Lobato, Márcio Aurelio da Silva, Aldemir Ap Cavalini Jr & Valder
Steffen Jr (2019): Reliability-based robust multi-objective optimization applied to engineering
system design, Engineering Optimization

To link to this article: https://doi.org/10.1080/0305215X.2019.1577413

Published online: 14 Mar 2019.

Submit your article to this journal

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=geno20
ENGINEERING OPTIMIZATION
https://doi.org/10.1080/0305215X.2019.1577413

Reliability-based robust multi-objective optimization applied to

engineering system design
Fran Sérgio Lobatoa , Márcio Aurelio da Silvab , Aldemir Ap Cavalini Jrb and Valder Steffen Jrb
a NUCOP—Laboratory of Modeling, Simulation, Control, and Optimization, School of Chemical Engineering, Federal

University of Uberlândia, Uberlândia, Brazil; b LMEst—Structural Mechanics Laboratory, Federal University of

Uberlândia, School of Mechanical Engineering, Uberlândia, Brazil

ABSTRACT ARTICLE HISTORY

Probabilistic and non-probabilistic methods have been proposed to deal Received 4 June 2018
with design problems under uncertainties. Reliability-based design and Accepted 11 January 2019
robust design are probabilistic strategies traditionally used for this purpose. KEYWORDS
In the present contribution, reliability-based robust design optimization Reliability-based design;
(RBRDO) is formulated as a multi-objective problem considering the inter- robust design;
action of both approaches. The proposed methodology is based on the multi-objective optimization;
differential evolution algorithm associated with two strategies to deal with engineering systems;
reliability and robustness, respectively, namely inverse reliability analysis differential evolution
and the effective mean concept. This multi-objective optimization prob-
lem considers the maximization of reliability and robustness coefficients
as additional objective functions. The effectiveness of the methodology
is illustrated by two classical test cases and a rotor-dynamics application.
The results demonstrate that the proposed methodology is an alternative
method to solve RBRDO problems.

1. Introduction
Traditionally, the simulation, control and optimization of engineering systems are performed by con-
sidering deterministic models, variables and parameters. This assumption is not valid in real-world
applications. Observations and measurements clearly show, for instance, variability and random-
ness in parameters and design variables (Schuëller and Jensen 2008; Ritto, Sampaio, and Cataldo
2008). Thus, the solution of these problems often requires sensitivity, uncertainty and reliability
analyses associated with the design variables. The variables, mathematical models, loads, geometry,
boundary conditions, material properties, manufacturing processes and operational environment are
commonly affected by uncertainties (Choi, Grandhi, and Canfield 2006).
Various strategies based on the concepts of reliability and robustness have been used to handle
engineering system design under uncertainties (Gu et al. 2013). Robustness and reliability method-
ologies are classified according to probabilistic and non-probabilistic methods. The former class of
methods has been widely used to deal with uncertainties in the context of reliability-based design
(RBD) and robust design (RD) (Carter 1997; Melchers 1999; Wang, Wang, and Almaktoom 2013;
Prigent et al. 2015; Doh, Kim, and Lee 2018).
As mentioned by Du and Chen (2004) and Du, Sudjianto, and Chen (2004), RBD emphasizes
reliability in the design by ensuring the probabilistic achievement of constraints at predefined lev-
els. According to Poles and Lovison (2009), RBD can be understood as the ability of a system or

CONTACT Aldemir Ap Cavalini Jr [email protected]

© 2019 Informa UK Limited, trading as Taylor & Francis Group

2 F. S. LOBATO ET AL.

component to perform its required functions under certain conditions. In the context of engineering
systems, RBD is based on the probability of the desired system performance failing. In contrast, RD
determines solutions presenting small sensitivity to changes in the design variables (Deb et al. 2009).
In both approaches, either Monte Carlo simulation (MCS) or the Latin hypercube design (LHD)
method is used to perform analyses around the deterministic solution (Du, Sudjianto, and Chen
2004). Consequently, these approaches are time consuming. Non-probabilistic methods have been
used to quantify uncertainty (i.e. fuzzy sets and interval sets), as they are more efficient than MCS
and LHD (Poles and Lovison 2009). Non-probabilistic methods can quantify uncertainties without
using probability distributions.
Both RD and RBD approaches have been applied to engineering system design. Deb and Gupta
(2006), considering the effective mean concept (EMC), studied the influence of probabilistic con-
straints in an RD problem devoted to mathematical and engineering problems. In Khakhali et al.
(2010), a multi-objective optimization approach was proposed to deal with robustness. In this case,
two polynomial surrogate models based on the group method of data handling were obtained to rep-
resent conflicting objectives related to independent geometric and material design variables. Sun et al.
(2011) presented a multi-objective particle swarm optimization (PSO) to evaluate the effects of para-
metric uncertainties on multiple crashworthiness criteria, where the energy absorption and reduction
of structural weight were considered as objective functions. Martowicz and Uhl (2012) evaluated the
performance of a multi-criteria RD optimization approach dedicated to microelectromechanical sys-
tems. Sun et al. (2014) proposed a RD methodology to address the effects of parametric uncertainties
on a foam-filled thin-walled structure. The methodology is based on the kriging modelling technique,
sequential quadratic programming and MCS.
Fang et al. (2015) proposed a multi-objective optimization to simultaneously enhance the perfor-
mance and robustness of the fatigue life of a truck cab. Different dual surrogate modelling methods
were used to approximate the mathematical model. The problem was solved using a multi-objective
PSO algorithm. Li et al. (2015) studied design problems that present parametric uncertainties in both
random and interval variables of a foam-filled thin-walled tube. The proposed methodology consid-
ers two loops: in the outer loop, the non-dominated sorting genetic algorithm II (NSGA-II) is used
to generate a robust Pareto solution; in the inner loop, MCS is applied to evaluate the responses of
the robustness criterion of the optimized design. Sun et al. (2015) proposed a methodology based on
the successive Taguchi approach to maximize the crashworthiness of hybrid tailor welded structures
under uncertainties. The resulting peak force was taken as the objective function, while the specific
energy absorption was considered as a constraint function. Sun et al. (2018) proposed a novel multi-
objective discrete robust optimization algorithm dedicated to engineering structures design under
uncertainties. In this case, the original multi-objective problem was rewritten as an equivalent mono-
objective problem using the grey relational analysis associated with the principal component analysis.
The proposed method was applied to a sophisticated design of a full-scale vehicle for crashworthiness
criteria.
Du, Sudjianto, and Chen (2004) proposed the inverse reliability analysis (IRA) to deal with relia-
bility in non-concave and non-convex performance functions of random variables described by any
continuous distributions. Song, Sun, and Li (2016) performed experimental and numerical analyses
devoted to the optimum design of high-strength steel tailor-welded structures for crashworthiness.
Finite element analyses were adopted to simulate the crashworthiness process, which was validated
by a set of experimental tests. The associated mathematical model was approximated using surro-
gate models and the optimization procedure was performed using global sensitivity analysis based
on Sobol’s method. Sun et al. (2017) proposed a multi-objective reliability-based design optimiza-
tion (RBDO) of a double-hat thin-walled structure. In this case, surrogate-modelling techniques were
associated with a multi-objective PSO and MCS to perform the optimum design. Sun et al. (2017) pro-
posed a multi-objective and multi-case RBDO devoted to a tailor-rolled blank hat-shaped structure.
The radial basis function metamodel was adopted to approximate the responses of the objective func-
tions and constraint functions in association with the NSGA-II, coupled with MCS, to find optimal
 ENGINEERING OPTIMIZATION 3

reliability solutions. Zhang et al. (2018) studied the crashworthiness optimization for a foam-filled
tapered structure by considering RBDO. Uncertainty in the design variables and noise factors were
taken into account in both single- and multiple-objective contexts.
RBD and RD approaches can deal with uncertainties. As mentioned by Tang, Chen, and Wei
(2012), in realistic problems the discrepancy between real and calculated optimal values may cause
violations of the probabilistic constraints in an RBD problem. In this scenario, it is important to mini-
mize the variation of the probabilistic constraints, associated with the design variables at the optimum.
This is achieved by applying a robustness criterion to the reliability problem. Thus, the development
of RBD- and RD-based methodologies is a natural way to deal with uncertainties. In the literature,
these methods are known as reliability-based robust design optimization (RBRDO).
Different strategies have been proposed to solve RBRDO problems. Youn, Choi, and Yi (2005) eval-
uated the performance moment integration (PMI) method to estimate product quality loss based on
mean and variance values. The authors demonstrated the efficiency of this approach using both math-
ematical and engineering examples to estimate product quality loss for even non-normally distributed
statistical responses with high skewness and kurtosis. In addition, the authors presented three new
formulations of RBRDO for various types of robustness criterion to represent the objective function.
The considered formulations were demonstrated to be effective for RBRDO. In this case, the robust
objective function was minimized and the probabilistic constraints were satisfied. For the reliabil-
ity part of the optimization problem, the enriched performance measure approach (PMA) and the
enhanced hybrid-mean value (HMV) method were used.
Lee et al. (2008) developed an RBRDO methodology based on the dimension reduction method
(DRM). This approach was applied to one-dimensional and two-dimensional functions. The per-
formance of the mentioned approach was also evaluated in a large-scale engineering problem. The
obtained results were compared with those determined by other strategies for accuracy and efficiency
purposes. Tang, Chen, and Wei (2012) developed a sequential algorithm based on a sequence of cycles.
The approach is composed of a deterministic optimization followed by an inverse robustness and reli-
ability evaluation to reduce the computational cost used to solve the RBRDO problem. In this strategy,
the epistemic uncertainty of design variables is quantified using an info-gap model and the feasible
robustness is characterized by a robustness index. The performance and feasibility of the proposed
methodology were successfully tested for two case studies.
Du, Sudjianto, and Chen (2004) proposed the percentile performance difference (PPD) approach
as an alternative strategy to formulate the RBRDO problem. In this case, PPD is used to represent
variations in performance. The objective function was represented either by two reliability levels or
using cumulative distribution functions. The multi-objective problem was transformed into an equiv-
alent mono-objective problem using the weighted sum method. Lagaros, Plevris, and Papadrakakis
(2007) presented an integrated reliability and robustness design methodology devoted to optimiz-
ing steel structures. In this approach, the probabilistic constraints were incorporated into the RD
optimization problem, which was formulated as a multi-criteria optimization problem. Gu et al.
(2013) compared the reliable, robust and deterministic designs of a typical vehicle structure sub-
jected to offset frontal crashing scenarios. Experimental data, metamodels and the NSGA-II were
used. Shahraki and Noorossana (2013) combined RD and RBD approaches and proposed a genetic
algorithm strategy combined with a reliability analysis code to solve mathematical functions. Rathod
et al. (2013) presented a comparative study of different RBRDO approaches. Zhuang, Pan, and Du
(2015) developed a multi-objective optimization strategy with two objective functions to solve a real-
istic engineering problem. Gholaminezhad, Jamali, and Assimi (2016) proposed a multi-objective
reliability-based RD approach using the sensitivity-assisted MCS method applied to the design of
a robot gripper mechanism. In this application, the mean and standard deviation of the difference
between the maximum and minimum gripping forces and the transmission ratio of actuated and
experienced gripping forces at the gripper ends were considered as the objective functions. To achieve
the desired reliability level, the mechanism functional constraints were converted into probabilistic
constraints.
4 F. S. LOBATO ET AL.

The RBRDO approach is inherently a multi-objective optimization problem. As mentioned

already, this kind of problem has been treated by considering two alternative approaches: (1) transfor-
mation of the original multi-objective optimization problem into a similar mono-objective problem
using the weighted sum method; and (2) solution of the original problem by performing multi-
objective techniques (Pareto’s dominance criterion). In general, the RBRDO problem is formulated
in terms of the objective function mean and variance values associated with the probabilistic con-
straints. As an alternative strategy, in the present contribution, a new mathematical formulation and
methodology to solve RBRDO problems are proposed. In this methodology, the original optimiza-
tion problem (mono-objective problem) is rewritten as an equivalent multi-objective optimization
problem. The original objective function is associated with two new objectives, namely the maxi-
mization of the robustness parameter (necessary to apply the EMC strategy) and the maximization
of the reliability coefficient (necessary to apply the IRA strategy). The proposed methodology can
be summarized as follows: (1) application of the differential evolution (DE) algorithm (Storn and
Price 1995) to generate potential candidates; (2) evaluation of each candidate generated accord-
ing to the EMC approach (Deb and Gupta 2006); (3) evaluation of each candidate considering
the IRA strategy (Du, Sudjianto, and Chen 2004); and (4) determination of non-dominated solu-
tions according to the Pareto criterion. Thus, the multi-objective optimization differential evolution
(MODE) algorithm is used (Lobato and Steffen 2011). In the present contribution, the proposed
methodology, namely IRA + EMC + MODE, was successfully tested in two classical case studies. In
addition, a rotor-dynamics application is presented to demonstrate the effectiveness of the proposed
methodology.
This research work is organized as follows. Sections 2 and 3 provide a brief review of the RD and
RBD approaches, respectively. The proposed methodology is discussed in Section 4. The numerical
results are presented in Section 5. Finally, the conclusions are outlined in Section 6.

2. Robust design
Engineering systems are sensitive to small variations in their design variables. Thus, robust optimiza-
tion can be used to minimize the influence of these variations on the performance of the system.
Figure 1 illustrates this concept by considering a mono-objective optimization problem. In this case,
A and B represent a local and global solution, respectively. If B is subject to a small perturbation δ,
the objective function f (x) can reach higher values. However, a small variation in f (x) is obtained if
the same perturbation δ is applied in A. Thus, A is considered more robust than B.
Commonly, new constraints and/or objective functions are required to deal with robustness issues
in both mono- and multi-objective design problems. As an alternative method, Deb and Gupta (2006)

Figure 1. Global versus robust solutions in a mono-objective optimization problem (Deb and Gupta 2006).
 ENGINEERING OPTIMIZATION 5

proposed an extension of the EMC approach to avoid the introduction of new constraints into the
original problem, which is rewritten to consider the mean value of the original objective functions.
The robustness measure is based on the following definition.
Definition: f :  ⊂ Rn → R being an integrable objective function, the EMC of f associated with
the neighbourhood δ of the solution x is the function feff (x,δ) given by:

1
f eff (x, δ) = f (y)dy (1)
|Bδ (x)|
y∈Bδ (x)

subject to:
eff
gj (x, δ) ≤ 0 j = 1, . . . , k x ∈  ⊂ Rn (2)

xl ≤ x ≤ xu (3)
where gj represents the jth inequality constraint (j = 1, . . . , k; k is the number of inequality con-
straints), |Bδ (x)| is the hypervolume of the neighbourhood δ, and xl and xu represent the lower and
upper limits of the design variables, respectively. N samples are generated randomly using the LHD
approach to estimate Equation (1).
This procedure was successfully used in engineering systems design (Souza, Lobato, and Gedraite
2015; Moreira et al. 2016). However, it is important to emphasize that the computational cost required
to solve this kind of problem is high owing to additional evaluations of the objective functions (Deb
and Gupta 2006).

3. Reliability-based design
According to Deb et al. (2009), the true optimum result has to be penalized and an interior solution
can be chosen within the feasible region to find a reliable solution, which presents a small probability
of producing an infeasible solution. From a given measure of reliability R (desired reliability value), it
is necessary to find a solution ensuring the probability of having an infeasible solution given by 1 – R.
Thus, the RBD problem can be formulated as follows:

min f (xd , xr ) (4)

P(Gj (xd , xr ) ≤ 0) ≥ Rj (5)

xld ≤ xd ≤ xud (6)

where Gj represents the jth inequality constraint, xd is the vector of design variables, xr is the vector
of random variables, and P is the probability of Gj being less than or equal to zero (Gj ≥ 0 indicates
failure). The vector of design variables contains deterministic values associated with the lower and
upper limits xd l and xd u , respectively. The probability of failure is defined by the following cumulative
distribution function (Deb et al. 2009):
 
P(Gj (xd , xr ) ≤ 0) = . . . fX (xr )dxr (7)
Gj ≤0

where fX is a joint probability density function.

The integral defined by Equation (7) has to be evaluated along the design space defined by the vec-
tor of inequality constraints to determine the probability of failure. Either the analytical or numerical
evaluation of Equation (7) is computationally expensive owing to the domain specified. Thus, the
6 F. S. LOBATO ET AL.

so-called Rosenblatt transformation is used to rewrite the original problem with random variables
xr as a similar problem with a new vector of random variables u (Rosenblatt 1952). Considering the
new search space (u-space), the most probable point (MPP) for failure is determined by locating the
minimum distance between the origin and a given constraint function, which is defined by the reli-
ability coefficient β. Equation (7) can be expressed through an inverse transformation for a given
probabilistic constraint, as presented by Equation (8):

P(G(xd , u) ≤ 0) = (β) (8)

where  is the standard normal distribution function. This transformation is used by different
approaches to measure reliability and avoid the evaluation of Equation (7). In the reliability coefficient
approach,  is used to describe the probabilistic constraint in Equation (7).
It is important to mention that the probability distribution of the design variables has to be known.
Consequently, probability tools are required to model uncertainties (Soize 2005). In this way, random
variables are associated with the uncertain parameters and probability density functions are deter-
mined. Regarding RBD applications, the stochastic variables are converted from the physical space
to a standard normal space. The associated optimization problem is reformulated to compute the
largest probability of failure, which is simultaneously equated to the desired value (Deb et al. 2009).
Therefore, RBD procedures can be classified according to four main categories (Agarwal 2004): (1)
simulation methods, (2) double-loop methods, (3) decoupled methods, and (4) single-loop methods.

3.1. Inverse reliability analysis

The IRA approach, proposed by Du, Sudjianto, and Chen (2004), arose as an alternative to traditional
RBD techniques. This approach is based on the following probability equation for a given constraint:

P(G(xd , u) ≤ Gp ) = R (9)

where Gp must be determined through the solution of a typical reliability inverse problem and R is
defined by the user.
Equation (9) indicates the probability of G(xd ,xr ) being equal to a given reliability measure R if
G(xd ,u)≤Gp , in which Gp is estimated using the well-known reliability approach FORM (first order
reliability method) (Zhao and Ono 1999). FORM is applied by considering G’(xd ,u) and the MPP u*
for P(G’(xd ,u) ≤ 0) = P(G’(xd ,u) ≤ GP ), as given by Equation (10):

G (xd , u) = G(xd , u) − Gp (10)

Figure 2 shows that u* is a tangent point of the circle with radius β and performance function
G(xd ,u) − Gp = 0. In addition, u* is a point associated with the minimum value of G(xd ,u) on the
circle. Therefore, the MPP search for an IRA problem becomes a search to find the minimum value of
G(xd ,u) in the β-circle (or β-sphere for a three-dimensional problem or β-hypersphere for a higher
dimensional problem).
The procedure can be summarized according to the following steps (Du, Sudjianto, and Chen
2004):

Step 1. Inform the starting point (k = 0, u0 = 0, β, and distribution type);

Step 2. Evaluate Gj (xd , uk ) and ∇Gj (xd , uk ). Compute a (see a* in Figure 2) as follows:

∇Gj (xd , uk )
ak = (11)
||∇Gj (xd , uk )||

Step 3. Update uk+1 = −βak and k = k + 1;

 ENGINEERING OPTIMIZATION 7

Figure 2. Inverse most probable point search (Du, Sudjianto, and Chen 2004).

Step 4. Repeat Steps 2 and 3 until convergence (|uk+1 − uk | < tolerance) is achieved.

As can be observed, if the reliability coefficient and the distribution are known, the point associated
with the minimum value of G(xd ,u) can be determined.

4. Methodology
Traditionally, RBRDO deals with robustness by considering the optimization of both the mean (μf )
and variance (σ 2 f ) of the objective function f. Reliability is considered in the probabilistic constraints.
Mathematically, this problem can be formulated as follows (Du, Sudjianto, and Chen 2004; Lee et al.
2008):
min (μf , σf2 ) (12)

P(Gi (xd , xr ) ≥ 0) ≤ (−β) (13)

xld ≤ xd ≤ xud (14)

where (μf , σ2
f ) is the new objective function in terms of the mean and variance values.
The IRA + EMC + MODE methodology proposed in the present contribution starts by rewriting
the original problem (mono-objective – min f ) into an equivalent multi-objective optimization prob-
lem with two new objective functions: the maximization of the robustness parameter (max δ) and the
maximization of the reliability coefficient (max β). Mathematically, the RBRDO problem is defined as:

min f (xd , xr ) (15)

max β (16)

max δ (17)

P(Gi (xd , xr ) ≥ 0) ≤ (−β) (18)

xld ≤ xd ≤ xud (19)

8 F. S. LOBATO ET AL.

Figure 3. Flowchart of the inverse reliability analysis (IRA) + effective mean concept (EMC) + multi-objective optimization dif-
ferential evolution (MODE) strategy to solve reliability-based robust design optimization (RBRDO) problems. DE = differential
evolution.

The main advantage of this formulation is the possibility of treating the original problem by con-
sidering the Pareto dominance criterion. In this case, an optimal solution is found for all the objective
functions simultaneously. In each point belonging to the obtained Pareto curve, it is possible to choose
the best configuration in terms of one or more objective functions. Thus, it is possible to define the
most interesting design after obtaining the optimal solution.
The proposed methodology starts by evaluating the candidates generated by MODE according to
EMC, in which N neighbours related to each candidate are determined using the LHD technique.
Then, for each candidate, IRA is applied by taking into account the reliability coefficient to find the
point exhibiting the highest probability value, i.e. the point that corresponds to the minimum value of
G(xd , xr ). Finally, the integral associated with the original objective function and the constraints of the
problem are evaluated by considering the robustness parameter. Then, the original objective function
is evaluated by taking into account the influence of reliability and robustness. In addition, both the
reliability coefficient and the robustness parameter are simultaneously maximized. The Pareto curve
is constructed for each generation. This procedure continues until the maximum number of genera-
tions is achieved (stop criterion). Figure 3 presents the flowchart regarding the IRA + EMC + MODE
methodology to solve RBRDO problems.
The IRA + EMC + MODE strategy is implemented as follows:

Step 1. Initially, the MODE parameters [population size (NP), maximum number of generations
(GEN), crossover rate (CR), weight factor (F), reduction rate (r), number of pseudo-curves
(Nf ) and the strategy considered to generate potential candidates to solve the optimization
problem] are defined by the user;
Step 2. The multi-objective optimization problem is defined by the minimization of the original
objective and the maximization of the reliability index and the robustness parameter. Thus,
the original problem is converted into an equivalent multi-objective function that presents
 ENGINEERING OPTIMIZATION 9

two new objectives, namely maximizations of the reliability index and the robustness
parameter;
Step 3. The population of candidates (vector of design variables xd ) is generated using MODE
considering the strategy proposed in DE;
Step 4. For each candidate generated by DE, N neighbours (perturbations) are obtained consider-
ing the range delimited by the robustness parameter δ;
Step 5. For each neighbour, IRA is applied to determine the value of the inequality constraint from
the vector of random variables (uncertain variables) considering a reliability coefficient β;
Step 6. The values of xd and r u (the vector xr is transformed into an equivalent vector u) are used to
evaluate the original objective function defined by EMC and the inequality constraints (the
treatment of the constraints is performed through the static penalization method proposed
by Castro (2001));
Step 7. The evaluated population is organized according to the Pareto dominance criterion;
Step 8. The iteration process is repeated (Steps 3 and 7) until convergence is achieved (maximum
number of generations).

It is worth mentioning that the robustness concept is related to feasibility robustness in the context
of the present contribution. Thus, the formulation of the robust optimization problem is a special case
of RBDO.

5. Numerical applications
To evaluate the IRA + EMC + MODE strategy, two classical engineering test cases and a rotor-
dynamics application (rotating machine design) are considered. The following conditions were
considered in the context of these applications.

1. A mono-objective is proposed and solved using DE to evaluate the IRA approach. This problem
consists of minimizing the original objective functions using fixed values for the reliability
coefficient and for the robustness parameter. The obtained results are compared with those
obtained by other well-known techniques, except for the rotating machine design since this
case study has not been previously solved by any approach found in the literature.
2. DE was used to solve the more realistic problem (a rotating machine design) in the deter-
ministic, robust and reliability contexts. For each context, a mono-objective problem was
solved considering the following parameters: NP = 20, GEN = 100, F = 0.8, CR = 0.8 and
DE/rand/1/bin strategy to generate potential candidates using DE.
3. In the multi-objective context, the following parameters are used by MODE to solve all test
cases: NP = 50, GEN = 50, F = 0.8, CR = 0.8, r = 0.9, Rf = 10 and DE/rand/1/bin strategy
to generate potential candidates using DE.
4. The following ranges for the reliability coefficient and robustness parameter were considered:
1% ≤ δ ≤ 20% and 0.1 (53.9828%) ≤ β (R) ≤ 4 (99.9968%).
5. The number N of samples considered to generate neighbours in EMC is 50 for two first test
cases and 25 for the rotor design.
6. The vector of random variables u is considered equal to zero in the first iteration of IRA.
7. The derivatives of the constraints Gj required by IRA were obtained analytically owing to their
simplicity.
8. The stopping criterion used to finish IRA is associated with the Euclidean norm of the vector
u. If ||u|| < 1 × 10−5 along two consecutive iterations, the process is finished.
9. The number of evaluations of the objective function using the proposed methodology is
NP + NP × GEN × N × N IRA , where N IRA is the number of iterations required by IRA for each
neighbour in each generation.
10 F. S. LOBATO ET AL.

Figure 4. Cantilever beam problem (Qu and Haftka 2004).

Table 1. Statistical data of the random variables used in the

cantilever beam problem (Liao and Ha 2008).
Mean value Standard deviation
xr 1 (lb) 500 100
xr 2 (lb) 1000 100
xr 3 (psi) 29 × 106 1.45 × 106
xr 4 (psi) 40,000 2000

10. The stopping criterion used to finish optimization algorithms is the maximum number of
generations.
11. All case studies were run 10 times to obtain the upcoming average values.
12. It is important to emphasize that the two classical engineering test cases were proposed origi-
nally in the reliability context. In the present contribution, these problems were adapted to the
RBRDO context. The rotating machine design is a new engineering test case that was solved
in deterministic, robust and reliability contexts.

5.1. Cantilever beam problem

Several authors have previously studied this classical test case in the reliability context (Liao and Ha
2008; Wu et al. 2001; Ramu et al. 2004; Qu and Haftka 2004; Kuo-Wei and Gautama 2014). The
objective of this example is to find the minimum cross-sectional area of the cantilever beam presented
in Figure 4, following Equation (20):
min xd1 xd12 (20)
xd

where xd1 and xd2 are the width and height of the beam, respectively. These parameters are defined
as design variables (deterministic values) with lower and upper limits equal to 0 and 10, respectively.
The two loads xr1 and xr2 that are applied at the free end of the beam, Young’s modulus xr3 and the
yield strength xr4 are considered as random variables with normal distribution. The associated mean
values and standard deviations are shown in Table 1.
The length of the cantilever beam is L = 254 cm and the tip displacement has to be smaller than
the allowable displacement do = 7.724 cm. The target reliability is defined as 99.8651% (β = 3). The
stress at the fixed end of the beam must be smaller than the yield strength u4 . Therefore,
600xr2 600xr1
GS = 2
+ 2 − xr4 ≤ 0 (21)
xd1 xd2 xd1 xd2
The second failure mode GD is defined as the tip displacement exceeding the allowable displacement
do , as given by Equation (22):
   2

4L 3  xr2 2 xr1
GD =  + − do ≤ 0 (22)
xr3 xd1 xd2 2
xd2 2
xd1
 ENGINEERING OPTIMIZATION 11

Table 2. Reliability-based design results and performance measure obtained by solving the cantilever beam problem considering
δ = 0% (Kuo-Wei and Gautama 2014).
Strategy f (cm2 ), Equation (20) xd 1 (cm) xd 2 (cm)
Reliability index approach (Ramu et al. 2004) 9.5202 2.446 3.8922
Performance measure approach (Ramu et al. 2004) 9.5202 2.446 3.8920
Exact (Wu et al. 2001) 9.5204 2.4484 3.8884
First order reliability method (Kuo-Wei and Gautama 2014) 9.5202 2.44599 3.89219

Table 3. Results associated with the cantilever beam problem considering different values of δ using the inverse reliability analysis
(IRA) approach considering β = 3.
δ (%) f (cm2 ), Equation (20) xd 1 (cm) xd 2 (cm) neval
9.52025a /1.11 × 10−5 2.44838a /2.37 × 10−5 3.88838a /3.64 × 10−5
b b b
0 2,500,025a /8c
9.74023a /2.24 × 10−5 2.55312a /1.16 × 10−5 3.81515a /2.99 × 10−5
b b b
2.5 3,750,025a /12c
9.90947a /1.21 × 10−5 2.53019a /2.34 × 10−5 3.94261a /3.85 × 10−5
b b b
5 2,187,525a /7c
10.10964a /3.09 × 10−5 2.59731a /1.65 × 10−5 3.94082a /2.44 × 10−5
b b b
7.5 3,125,025a /10c
10.26731a /2.03 × 10−5 2.66355a /3.88 × 10−5 3.91947a /2.32 × 10−5
b b b
10 3,437,525a /11c
15 10.68800 /3.99 × 10
a −5 b
2.83320 /2.02 × 10
a −5 b
3.87016 /3.29 × 10
a −5 b
3,437,525a /11c
11.02776a /3.91 × 10−5 2.98477a /3.95 × 10−5 3.82826a /2.01 × 10−5
b b b
20 4,062,525a /13c
Note: a Mean value; b standard deviation; c average value required by IRA for each candidate in each run (NIRA ).

Table 2 presents the solutions obtained by various authors considering different strategies and the
associated analytical solution for β = 3 (R = 99.8651%). Table 3 shows the results obtained using
IRA (present work) with β = 3 and different values for the robustness parameter. Note that the results
obtained by considering δ = 0% (disregarding robustness) are similar to those determined by clas-
sical approaches (see Table 2), evidencing the quality of the solution obtained by IRA. As expected,
the value of the objective function increases according to the robustness parameter δ (β = 3). Note
that, owing to the evolutionary optimization approach, the total number of evaluations of the objec-
tive function neval required by the proposed methodology was higher than that used by the classical
approaches.
As mentioned, the value of the objective function increases according to the robustness parameter
for the same reliability coefficient. Thus, it is interesting to propose a unified strategy that considers the
influence of robustness and reliability simultaneously. In this case, Figure 5 presents the Pareto curve
obtained using a triple-objective formulation (minimization of the original objective function and
maximizations of the reliability and robustness coefficients) using the IRA + EMC + MODE strategy.
Note that the value of the objective function increases with β (Figure 5a) since more reliable solutions
are found. For a given value of β in the Pareto curve represented by Figure 5(b), it can be observed
that the maximum value of f increases with δ (more robust solutions are found). A similar behaviour
can be observed in Figure 5(c) for the robustness parameter δ. The relationship between β and δ is
presented in Figure 5(d). In this case, the user can choose the most interesting design configuration
among the solutions along the Pareto curve.
Table 4 presents selected points of the trade-off frontier observed in Figure 4. Note that the min-
imum value of the objective function is found for low values of β and δ. Therefore, as expected, the
best results are obtained when reliability and uncertainty are disregarded (deterministic solution).
Similar results were obtained in Deb and Gupta (2006) and Deb et al. (2009). The average number of
objective function evaluations required by IRA + EMC + MODE in each run is 7,250,100.
In this case, the obtained results could not be compared with other methodologies presented in the
literature owing to the new mathematical formulation of the RBRDO problem. However, the results
presented in Table 3 for the RBDO problem are consistent with those expected by considering only
the reliability analysis and robustness parameter. Thus, the results are qualitatively consistent with
those expected, i.e. the value of the objective function increases according to β and δ.
12 F. S. LOBATO ET AL.

Figure 5. Trade-off frontier between the optimal solution, the robustness parameter and the reliability coefficient considering the
cantilever beam problem: (a) f, β and δ; (b) f and β; (c) f and δ; (d) δ and β.

Table 4. Selected points of the Pareto curve for the cantilever beam problem.
xd 1 (cm) xd 2 (cm) β δ f (cm2 ), Equation (20)
3.86972 3.46936 3.97075 0.12750 13.18128
2.78185 3.90386 1.26356 0.19505 10.44066
2.53606 3.19853 0.14697 0.01081 8.11156

5.2. Short column design

This classical engineering test case considers a short column with rectangular cross-section and
dimensions xr5 and xr6 , subjected to the yield stress xr1 , the biaxial bending moments xr2 and xr3 ,
and the normal force xr4 (Aoues and Chateauneuf 2009). This model is based on the elastic–plastic
constitutive law of the material and contains six random variables (xri , i = 1, . . . , 6). The original
RBD problem is formulated to minimize the cross-sectional area of the column under the reliabil-
ity coefficient β = 3 (R = 99.8651%), in terms of the mean values (μ) of xr5 and xr6 , as given by
Equation (23):

min μxr 5 μxr 6

μxr
⎧
⎪
⎨ P(G ≤ 0) ≥ R
2
4xr2 4xr3 xr4
G=1− − − (23)
⎪
⎩ μxr5 μxr6 2 xr1 μxr5 2 μxr6 xr1 μxr5 μxr6 xr1
0 ≤ μxr5 /μxr6 ≤ 2
 ENGINEERING OPTIMIZATION 13

Table 5. Statistical data of the random variables used in the short

column design problem for the reliability-based design problem
(Aoues and Chateauneuf 2009).
Mean value Variation coefficient
xr 1 (MPa) 40 0.10
xr 2 (kNm) 250 0.30
xr 3 (kNm) 125 0.30
xr 4 (kN) 2500 0.2
xr 5 (m) µxr 5 0.15
xr 6 (m) µxr 6 0.15

Table 6. Reliability-based design results associated with the short column design problem considering
δ = 0%.
Strategy f (m2 ) – Equation (23) neval
Reliability index approach 0.3015 945
Performance measure approach 0.3015 684
Karush–Kuhn–Tucker 0.3077 567
Single-loop approach 0.2970 46
Sequential optimization and reliability assessment 0.3014 346
Note: a Mean value; b standard deviation.

Table 7. Results associated with the short column design problem considering different values of δ.
δ (%) f (m2 ), Equation (23) µxr 5 (m) µxr 6 (m) neval
0.30151a /2.33 × 10−5 0.38827a /3.04 × 10−5 0.77654a /3.44 × 10−5
b b b
0 3,437,525a /11c
2.5 0.30855/2.01 × 10 −5 0.43134/2.02 × 10 −5 0.71541/1.37 × 10 −5 3,437,525/11
5 0.31568/4.20 × 10−5 0.43211/1.32 × 10−5 0.73093/4.12 × 10−5 1,875,025/6
7.5 0.32419/1.32 × 10−5 0.43128/2.44 × 10−5 0.75220/3.09 × 10−5 1,875,025/6
10 0.32880/1.43 × 10−5 0.45666/2.03 × 10−5 0.72101/2.22 × 10−5 2,187,525/7
15 0.34929/5.11 × 10−5 0.48013/3.38 × 10−5 0.72961/3.30 × 10−5 2,500,025/8
20 0.36601/4363 × 10−5 0.51186/1.30 × 10−5 0.71808/1.53 × 10−5 2,812,525/9
Note: a Mean value; b standard deviation; c average value required by inverse reliability analysis for each candidate in each run (NIRA ).

The statistical information associated with the random variables are presented in Table 5. This prob-
lem was solved by Aoues and Chateauneuf (2009) in the RBD context by considering the initial
design (0.5, 0.5) and the classical optimization strategies presented in Table 6. As observed by Aoues
and Chateauneuf (2009), the Karush–Kuhn–Tucker (KKT) method converges to an optimal solution
slightly higher than the minimum values obtained by the other methods. The RBD method con-
verged to (μxr5 ,μxr6 ) = (0.61,0.31) considering μxr5 and μxr6 as deterministic values (optimal area of
0.1920 m2 ). The KKT approach converged to (μxr5 ,μxr6 ) = (0.42,0.47), corresponding to the opti-
mal area of 0.2003 m2 . A different number of objective function evaluations were obtained from the
application of the mentioned strategies.
Table 7 shows the results obtained using IRA associated with DE for β = 3 (R = 99.8651%) and
different values for the robustness parameter. Note that the results obtained using IRA + DE dis-
regarding robustness are similar to those determined by the classical approaches (see Table 6). In
addition, it can be observed that the value of the objective function increases according to δ. Conse-
quently, it is necessary to consider the robustness and reliability contexts simultaneously to obtain a
more realistic solution.
Figure 6 presents the influence of β (0.1 (53.9828%) ≤ β (R) ≤ 4 (99.9968%)) using the
EMC + IRA + MODE strategy in the short column design problem. In this case, δ = 0, 2.5, 5, 7.5, 10,
15 and 20%. As observed in the last test case, note that the value of the objective function increases
according to the reliability coefficient and the robustness parameter. The best results, as expected,
were obtained when reliability and uncertainty are disregarded (deterministic solution). Thus, a
14 F. S. LOBATO ET AL.

Figure 6. Trade-off frontier between the optimal solution, the robustness parameter and the reliability coefficient for the short
column design problem.

Figure 7. Finite element model of the rotor.

robust and reliable solution leads to an increase in the value of the objective function. Finally, the user
can choose the best combination of reliability coefficient and robustness parameter for the specific
application.

5.3. Rotating machine design

The methodology proposed in the present contribution was applied to a horizontal rotating machine
modelled using 16 Timoshenko’s beam elements, which is shown in Figure 7.
The finite element model is composed by a flexible steel shaft with 780 mm length (LS ) and 6 mm
radius (E = 2.07 × 1011 Pa, ρ = 7800 kg/m3 and υ = 0.3), two rigid discs D1 (node 8) and D2 (node
11), both made of steel with 100 mm diameter and 20 mm thickness (ρ = 7800 kg/m3 ), and two cylin-
drical hydrodynamic bearings (B1 and B2 , located at nodes 4 and 14, respectively), each one with
10 mm radius (RB ), 20 mm length (LB ) and radial clearance of 100 μm. The oil viscosity is equal to
0.02 Pa.s. The hypothesis of short sleeve bearing is justified from the geometry of the bearings. The
hydrodynamic supporting forces of the bearings were determined by following the approach pro-
posed by Capone (1986). More details about the rotor model can be found in Cavalini et al. (2015).
The rotating parts model takes into account a proportional damping (Dp = β K) with the coefficient
β = 2 × 10−4 . An unbalance of 10 kg m at 0o applied to each disc of the rotor was included in the
model.
Figure 8 illustrates the vibration responses (displacements) obtained along the horizontal and ver-
tical directions of the disc D1 following a linear run-up condition (100–2300 rev/min in 5 s). It is
possible to observe that the first critical speed of the rotating machine is approximately 1584 rev/min.
Similar vibration amplitudes are obtained at the disc D2 .
 ENGINEERING OPTIMIZATION 15

Figure 8. Rotor performing a linear run-up condition: (a) horizontal direction; (b) vertical direction.

The objective of this example is to find optimum values of the shaft radius xd1 , oil viscosity (xd2 and
xd4 ; bearings B1 and B2 , respectively) and radial clearance (xd3 and xd5 ; bearings B1 and B2 , respec-
tively). This problem consists of minimizing the maximum vibration amplitude (Vib) measured at
the discs (dimensionless value), as given by Equation (24):

min Vib + g12 + g22 (24)

x

The rotor design problem is subjected to the constraints g 1 and g 2 , as given by:

g1 ≡ −S1 + 0.085 ≤ 0 (25)

g2 ≡ −S2 + 0.085 ≤ 0 (26)

where the value of 0.085 is the Sommerfeld number associated with the design that results in the
minimum allowed friction coefficient in the bearings (Shigley, Mischke, and Budynas 2003). S1 and
S2 are the Sommerfeld numbers associated with the bearings B1 and B2 , respectively, as given by
Equations (27) and (28):
 
RB 2 xd2 
S1 = (27)
xd3 60P1
 2
RB xd4 
S2 = (28)
xd5 60P2
where  is the rotation speed of the rotor. P1 = F 1 /(2RB LB ) and P2 = F 2 /(2RB LB ) are the projected
pressure on the bearings’ surface (B1 and B2 , respectively). F 1 and F 2 are the forces acting along
the vertical direction of B1 and B2 due to the static load generated by the rotating parts (shaft and
discs). Consequently, F 1 and F 2 have to be updated according to xd1 (shaft radius), as demonstrated
by Equations (29)–(32):
FD = 9.81 π(0.052 − xd1 2 )LB ρ (29)

FS = 9.81 πxd1 2 LS ρ (30)

F1 = 0.8771FD + 0.339FS (31)

F2 = 1.1229FD + 0.661FS (32)

in which FD and FS are the static forces provided by each disc and the shaft, respectively.
16 F. S. LOBATO ET AL.

In this test case, the maximum vibration response of the rotor system (Vib in Equation 24)
was obtained for the rotor following a linear run-up condition (100–2300 rev/min in 5 s; see
Figure 6). However, the Sommerfeld numbers (Equations 27 and 28) were determined by consid-
ering  = 2300 rev/min (operational rotation speed of the rotor). The rotor design was performed
by considering different scenarios: deterministic, robust, reliability-based and reliability-based robust
optimization problems. The last scenario corresponds to the approach proposed in this work.
This optimization problem was proposed by the authors to evaluate the performance of the
IRA + EMC + MODE strategy. For this purpose, the problem was solved considering deterministic
optimization, robust optimization, reliability-based optimization and RBRDO.

5.3.1. Deterministic optimization

The deterministic optimization considers that the design variables are not affected by uncertainties,
i.e. the variables are deterministic quantities (xdi ). The minimization problem of Equations (24)–(32)
considers the following design space: 0.0054 m ≤ xd1 ≤ 0.009 m, 0.018 Pa.s ≤ xd2 ≤ 0.03 Pa.s,
90 μm ≤ xd3 ≤ 150 μm, 0.018 Pa.s ≤ xd4 ≤ 0.03 Pa.s and 90 μm ≤ xd5 ≤ 150 μm. Table 8 shows
the deterministic results obtained using the canonical DE approach, disregarding robustness and reli-
ability. Note that the value obtained for xd1 (shaft radius) was close to the upper limit of the design
space. In contrast, the values determined for xd2 and xd4 (oil viscosity of B1 and B2 , respectively)
were close to the associated lower limits. Intermediate values were obtained for xd3 and xd5 (radial
clearance of B1 and B2 , respectively).

5.3.2. Robust optimization

The robust optimization considers that all the design variables are affected by uncertainties. There-
fore, Table 9 presents the solutions obtained using EMC and DE approaches for different values
of δ. Note that, as expected, the value of the objective function increases according to the robust-
ness parameter. In addition, the shaft diameter (variable xd1 ) decreases as δ increases. A different
behaviour is observed for the remaining variables. The average number of evaluations of the objective
function required by this methodology for each run was 50,020 (NP + NP × GEN × N evaluations).

Table 8. Deterministic results associated with the

rotating machine design problem.

0.00017a /6.19×10−7
b
f, Equation (24)
0.00899a /1.46 × 10−7
b
xd 1 (m)
0.01809a /2.22 × 10−7
b
xd 2 (Pa.s)
xd 3 (μm) a
139.99998 /0.356 b

0.01895a /2.49 × 10−7

b
xd 4 (Pa.s)
xd 5 (μm) a
119.99998 /0.409 b
neval 2020a /0b
Note: a Mean value; b standard deviation.

Table 9. Robust results associated with the rotating machine design problem.
δ (%)

2.5 5 10 20
0.00033a /1.43 × 10−7 0.00039a /2.43 × 10−7 0.00107a /1.48 × 10−7 0.00259a /2.43 × 10−7
b b b b
f, Equation (24)
0.0090a /6.02 × 10−7 0.00839a /1.33 × 10−7 0.00723a /1.11 × 10−7 0.0066a /2.39 × 10−7
b b b b
xd 1 (m)
xd 2 (Pa.s) 0.01800 /5.65 × 10
a −4 b
0.01800 /2.77 × 10
a −7 b
0.01924 /2.38 × 10
a −7 b
0.01800 /1.37 × 10
a −7 b

xd 3 (μm) a
139.99998 /1.49 b a
149.99998 /0.188 b a
149.99998 /0.138 b a
149.99998 /0.339 b

0.03000a /6.09 × 10−4 0.03000a /1.37 × 10−7 0.02289a /1.35 × 10−7 0.02659a /2.22 × 10−7
b b b b
xd 4 (Pa.s)
xd 5 (μm) a
149.99998 /23.5 b a
149.99999 /0.436 b a
139.99998 /0.365 b a
149.99998 /0.131 b
neval 50020a /0b 50020a /0b 50020a /0b 50020a /0b
Note: a Mean value; b standard deviation.
 ENGINEERING OPTIMIZATION 17

Table 10. Statistical data of the random variables used in the rotating
machine design problem.
Mean value Coefficient of Variation
xr 1 (m) µxd 1 0.10
xr 2 (Pa.s) µxd 2 0.10
xr 3 (μm) µxd 3 0.10
xr 4 (Pa.s) µxd 4 0.10
xr 5 (μm) µxd 5 0.10

Table 11. Reliability results associated with the rotating machine design problem.
β (R)

0.1 (53.9828%) 2 (97.7249%) 3 (99.8650%) 4 (99.9968%)

0.00017a /1.70 × 10−7 0.00234a /1.23 × 10−7 0.00947a /2.42 × 10−10 0.02999a /5.65 × 10−10
b b b b
f, Equation (24)
0.00900a /2.25 × 10−7 0.00900a /2.98 × 10−7 0.00891a /2.64 × 10−10 0.0090a /1.87 × 10−10
b b b b
xd 1 (m)
0.01889a /2.45 × 10−7 0.01800a /3.87 × 10−8 0.01800a /2.45 × 10−9 0.0180a /1.99 × 10−10
b b b b
xd 2 (Pa.s)
xd 3 (μm) a
149.99998 /0.975 b a
149.99998 /0.0144 b 149.99998a /0.00187b a
149.99998 /0.000264 b

0.02739a /9.29 × 10−7 0.01800a/2.63 × 10−7 0.0180a /2.41 × 10−9 0.0180a /3.65 × 10−10
b b b b
xd 4 (Pa.s)
xd 5 (μm) a
149.99998 /0.835 b a
149.99998 /0.0142 b a
149.99998 /0.00142 b a
149.99998 /0.000336 b
neval 22020a /0b 26020a /0b 18020a /0b 20020a /0b
Note: a Mean value; b standard deviation.

5.3.3. Reliability-based optimization

In this case, only reliability is considered and the variables are uncertain quantities. Thus, the mean
values of the design variables should be obtained (μxdi ; i = 1, . . . , 5). The statistical data associated
with the random variables are presented in Table 10.
The constraints presented in Equations (25) and (26) are considered in the reliability context,
which are defined by the following equation:
P(gi (x) ≤ 0) = (β) i = 1, 2 (33)
Table 11 shows the results obtained using the IRA and DE approaches by considering different
values of reliability coefficient and a normal distribution for the design variables. Note that the limit
of the design space was achieved by all design variables for β = 2, β = 3 and β = 4. However, as
expected, the value of the objective function increases according to β. This result is associated with
the considered probabilistic constraints, in which identical values for the design variables can result in
different values of the objective function due to the probabilistic constraints. The average number of
evaluations of the objective function required by the proposed methodology for each run was 21,520
(NP + NP × GEN × N evaluations).

5.3.4. Reliability-based robust optimization

The results obtained using the IRA + EMC + MODE approach proposed in the present work for
the rotating machine design are shown next. The original optimization problem (minimization of
Equation 24) is rewritten by defining two additional objective functions, namely the maximization
of the reliability index and the maximization of the robustness parameter. Figure 9 presents the pro-
jection of the objective function versus the reliability coefficient and the Pareto curve obtained using
a three-objective formulation (minimization of the original objective function and maximizations
of the reliability and robustness parameters). Note that the value of the objective function increases
according to β (Figure 9a). In addition, for β ≤ 2.5 (approximately), no influence of the robustness
parameter on the value of the objective function is verified. The influence of the robustness param-
eter can be observed for higher values of β (Figure 9b). As expected, a similar behaviour can be
observed for the robustness parameter in Figure 9(c). Figure 9(d) presents the relationship between
the parameters δ and β.
18 F. S. LOBATO ET AL.

Figure 9. Trade-off frontier between the optimal solution, the robustness parameter and the reliability coefficient considering the
rotating machine design problem: (a) f, β and δ; (b) f and β; (c) f and δ; (d) δ and β.

Table 12. Selected points of the Pareto curve for the rotating machine design problem.
µxd 1 (m) µxd 2 (Pa.s) µxd 3 (μm) µxd 4 (Pa.s) µxd 5 (μm) β δ f, Equation (24)
0.00698 0.02549 149.99998 0.02665 139.99998 3.99318 0.17138 0.13093
0.00678 0.02758 129.99997 0.02331 139.99998 3.79887 0.19976 0.27905
0.00691 0.01841 129.99998 0.02105 119.99999 0.128561 0.02744 0.09348

Table 12 presents some points of the trade-off frontier observed in Figure 9. As expected, the min-
imum value of the objective function is found for low values of β and δ. The average number of
evaluations of the objective function required by the proposed methodology for each run was 687,550
(NP + NP × GEN × N × N IRA ).

6. Conclusions
This article investigated the combination of the EMC and IRA approaches with the MODE algorithm
to solve RBRDO problems. The proposed IRA + EMC + MODE approach is based on the definition
of a new optimization problem with additional objective functions, which are the maximizations of
the reliability and robustness parameters. The methodology performs a three-loop coding, namely an
external loop associated with MODE, an outer loop with EMC and an inner loop with IRA. Three test
cases with different complexities were studied to evaluate the effectiveness of the proposed algorithm.
The first two cases represent classical problems used in the literature to evaluate reliability-based
 ENGINEERING OPTIMIZATION 19

optimization techniques. The last test case is a rotor-dynamics application, which consists of the min-
imization of the rotor vibration amplitudes considering a minimum allowed friction coefficient in the
bearings.
Considering the proposed IRA + EMC + MODE strategy, the following points should be
highlighted:

1. All test cases considered in this contribution (except for the rotating machine design problem)
were, originally, proposed in the reliability context. Then, these case studies were rewritten in
the RBRDO context.
2. Owing to the new mathematical formulation proposed for the RBRDO problem, the results
could not be compared directly with other RBRDO methodologies. However, the IRA approach
was compared with other strategies to validate the RBDO results. The robustness and reliability
strategies were combined and the obtained results were demonstrated to be consistent, i.e. the
value of the objective function increased with the values of β and δ.
3. It is important to emphasize that the proposed methodology consists of the combination of IRA
and EMC in a multi-objective optimization context. In this case, the MODE algorithm was used.
However, other multi-objective optimization techniques could also be used successfully.
4. A normal probability distribution was chosen to represent all the uncertain design variables.
However, the proposed methodology is able to handle problems that consider different proba-
bility distributions. In this sense, the Rackwitz–Fessler method should be used.
5. It is worth mentioning that the number of evaluations of the objective function required by the
proposed approach is larger than that required by cases in which either robustness or reliability
techniques were applied separately.

Acknowledgements
The authors are thankful for the financial support provided to the present research effort by CNPq, FAPEMIG and
CAPES through the INCT-EIE.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This work was supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico [grant numbers
574001/2008-5 and 304546/2018-8]; Coordenação de Aperfeiçoamento de Pessoal de Nível Superior; and Fundação
de Amparo à Pesquisa do Estado de Minas Gerais [grant numbers TEC-APQ-3076-09, TEC-APQ-02284-15 and
TEC-APQ-004].

References
Agarwal, H. 2004. “Reliability-Based Design Optimization: Formulations and Methodologies.” PhD diss., University of
Notre Dame.
Aoues, Y., and A. Chateauneuf. 2009. “Benchmark Study of Numerical Methods for Reliability-Based Design.”
Structural and Multidisciplinary Optimization 41: 277–294.
Capone, G. 1986. “Orbital Motions of Rigid Symmetric Rotor Supported on Journal Bearings.” La Meccanica Italiana
199: 37–46.
Carter, D. S. 1997. Mechanical Reliability and Design. New York: Wiley.
Castro, R. E. 2001. “Optimization of Structures with Multi-objective Through Genetic Algorithms.” MSc thesis,
COPPE/UFRJ, Portuguese.
Cavalini, Jr A. A., F. A. Lara-Molina, T. P. Sales, E. H. Koroishi, and V. Jr Steffen. 2015. “Uncertainty Analysis of a Flexible
Rotor Supported by Fluid Film Bearings.” Latin American Journal of Solids and Structures 12 (8): 1487–1504.
Choi, S. K., R. Grandhi, and R. A. Canfield. 2006. Reliability-based Structural Design. London: Springer
Science+Business Media.
20 F. S. LOBATO ET AL.

Deb, K., and H. Gupta. 2006. “Introducing Robustness in Multi-objective Optimization.” Evolutionary Computation
14 (4): 463–494.
Deb, K., D. Padmanabhan, S. Gupta, and A. K. Mall. 2009. “Handling Uncertainties Through Reliability-Based
Optimization Using Evolutionary Algorithms.” IEEE Transactions on Evolutionary Computation 13 (5): 1054–1074.
Doh, J., Y. Kim, and J. Lee. 2018. “Reliability-Based Robust Design Optimization of Gap Size of Annular Nuclear Fuels
Using Kriging and Inverse Distance Weighting Methods.” Engineering Optimization 1 (1): 1–16.
Du, X., and W. Chen. 2004. “Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic
Design.” Journal of Mechanical Design 126 (2): 225–233.
Du, X., A. Sudjianto, and W. Chen. 2004. “An Integrated Framework for Optimization Under Uncertainty Using Inverse
Reliability Strategy.” Journal of Mechanical Design 126 (4): 562–570.
Fang, J., Y. Gao, G. Sun, C. Xu, and Q. Li. 2015. “Multi-objective Robust Design Optimization of Fatigue Life for a Truck
Cab.” Reliability Engineering & System Safety 135: 1–8.
Gholaminezhad, I., A. Jamali, and H. Assimi. 2016. “Multi-objective Reliability-Based Robust Design Optimization
of Robot Gripper Mechanism with Probabilistically Uncertain Parameters.” Neural Computing and Applications
1: 1–12.
Gu, X., G. Sun, G. Li, L. Mao, and Q. Li. 2013. “A Comparative Study on Multi-objective Reliable and Robust Opti-
mization for Crashworthiness Design of Vehicle Structure.” Structural and Multidisciplinary Optimization 48 (3):
669–684.
Khakhali, N., A. Nariman-Zadeh, A. Darvizeh, A. Masoumi, and B. Notghi. 2010. “Reliability-Based Robust Multi-
objective Crashworthiness Optimization of S-Shaped Box Beams with Parametric Uncertainties.” International
Journal of Crashworthiness 15 (4): 443–456.
Kuo-Wei, L., and I. Gautama. 2014. “A Single Loop Reliability-Based Design Optimization Using EPM and MPP-Based
PSO.” Latin American Journal of Solids and Structures 11 (5): 826–847.
Lagaros, N. D., V. Plevris, and M. Papadrakakis. 2007. “Reliability-Based Robust Design Optimization of Steel
Structures.” International Journal for Simulation and Multidisciplinary Design Optimization 1: 19–29.
Lee, I., K. K. Choi, L. Du, and D. Gorsich. 2008. “Dimension Reduction Method for Reliability-Based Robust Design
Optimization.” Computers and Structures 86: 1550–1562.
Li, F., G. Sun, X. Huang, J. Rong, and Q. Li. 2015. “Multi-objective Robust Optimization for Crashworthiness Design of
Foam Filled Thin-Walled Structures with Random and Interval Uncertainties.” Engineering Structures 88: 111–124.
Liao, K. W., and C. Ha. 2008. “Application of Reliability-Based Optimization to Earth-Moving Machine: Hydraulic
Cylinder Components Design Process.” Structural and Multidisciplinary Optimization 36 (5): 523–536.
Lobato, F. S., and V. Steffen Jr. 2011. “A New Multi-objective Optimization Algorithm Based on Differential Evolu-
tion and Neighborhood Exploring Evolution Strategy.” Journal of Artificial Intelligence and Soft Computing Research
1: 1–12.
Martowicz, A., and T. Uhl. 2012. “Reliability- and Performance-Based Robust Design Optimization of MEMS Struc-
tures Considering Technological Uncertainties.” Mechanical Systems and Signal Processing 32: 44–58.
Melchers, R. E. 1999. “Structural Reliability Analysis and Prediction.” Chichester: Wiley.
Moreira, F. R., F. S. Lobato, A. A. Jr Cavalini, and V. Jr Steffen. 2016. “Robust Multi-objective Optimization Applied to
Engineering Systems Design.” Latin American Journal of Solids and Structures 13: 1802–1822.
Poles, S., and A. Lovison. 2009. “A Polynomial Chaos Approach to Robust Multiobjective Optimization, Hybrid and
Robust Approaches to Multiobjective Optimization.” In Proceedings of Dagstuhl Seminar, edited by K. Deb, S. Greco,
K. Miettinen, and E. Zitzler, 1–15. Dagstuhl: Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik.
Prigent, S., P. Maréchal, A. Rondepierre, T. Druot, and B. Belleville. 2015. “A Robust Optimization Methodology for
Preliminary Aircraft Design.” Engineering Optimization 48 (5): 883–899.
Qu, X., and R. T. Haftka. 2004. “Reliability-Based Design Optimization Using Probabilistic Sufficiency Factor.”
Structural and Multidisciplinary Optimization 27 (5): 314–325.
Ramu, P., X. Qu, B. D. Youn, R. T. Haftka, and K. K. Choi. 2004. “Safety Factor and Inverse Reliability Measures.” In
Proceedings of 45th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, AIAA
2004-1670. Reston, VA: AIAA.
Rathod, V., O. P. Yadav, A. Rathore, and R. Jain. 2013. “Optimizing Reliability-Based Robust Design Model Using Multi-
objective Genetic Algorithm.” Computers and Industrial Engineering 66: 301–310.
Ritto, T. G., R. Sampaio, and E. Cataldo. 2008. ““Timoshenko Beam with Uncertainty on the Boundary Conditions.”
Journal of the Brazilian Society of Mechanical Sciences and Engineering XXX (4): 295–303.
Rosenblatt, M. 1952. “Remarks on a Multivariate Transformation.” The Annals of Mathematical Statistics 23: 470–472.
Schuëller, G. I., and H. A. Jensen. 2008. “Computational Methods in Optimization Considering Uncertainties – An
Overview.” Computer Methods in Applied Mechanics and Engineering 198: 2–13.
Shahraki, F., and R. Noorossana. 2013. “A Combined Algorithm for Solving Reliability-Based Robust Design Optimiza-
tion Problems.” Journal of Mathematics and Computer Science 7: 54–62.
Shigley, J., C. Mischke, and R. Budynas. 2003. Mechanical Engineering Design. 7th edn. New York: McGraw-Hill.
Soize, C. 2005. “A Comprehensive Overview of a Non-Parametric Probabilistic Approach of Model Uncertainties for
Predictive Models in Structural Dynamics.” Journal of Sound and Vibration 288 (3): 623–652.
 ENGINEERING OPTIMIZATION 21

Song, X., G. Sun, and Q. Li. 2016. “Sensitivity Analysis and Reliability Based Design Optimization for High-Strength
Steel Tailor Welded Thin-Walled Structures Under Crashworthiness.” Thin-Walled Structures 109: 132–142.
Souza, D. L., F. S. Lobato, and R. Gedraite. 2015. “Robust Multi-objective Optimization Applied to Optimal Control
Problems Using Differential Evolution.” Chemical Engineering and Technology 1: 1–8.
Storn, R., and K. V. Price. 1995. “Differential Evolution: A Simple and Efficient Adaptive Scheme for Global Optimiza-
tion over Continuous Spaces.” International Computer Science Institute 12: 1–16.
Sun, G., J. Fang, X. Tian, G. Li, and Q. Li. 2015. “Discrete Robust Optimization Algorithm Based on Taguchi Method
for Structural Crashworthiness Design.” Expert Systems with Applications 42 (9): 4482–4492.
Sun, G., G. Li, S. Zhou, H. Li, S. Hou, and Q. Li. 2011. “Crashworthiness Design of Vehicle by Using Multi-objective
Robust Optimization.” Structural and Multidisciplinary Optimization 44 (1): 99–110.
Sun, G., X. Song, S. Baek, and Q. Li. 2014. “Robust Optimization of Foam-Filled Thin-Walled Structure Based on
Sequential Kriging Metamodel.” Structural and Multidisciplinary Optimization 49 (6): 897–913.
Sun, G., H. Zhang, J. Fang, G. Li, and Q. Li. 2017. “Multi-objective and Multi-case Reliability-Based Design Optimiza-
tion for Tailor Rolled Blank (TRB) Structures.” Structural and Multidisciplinary Optimization 55 (5): 1899–1916.
Sun, G., H. Zhang, J. Fang, G. Li, and Q. Li. 2018. “A New Multi-objective Discrete Robust Optimization Algorithm for
Engineering Design.” Applied Mathematical Modelling 53: 602–621.
Sun, G., H. Zhang, R. Wang, X. Lv, and Q. Li. 2017. “Multi-objective Reliability-Based Optimization for Crashworthy
Structures Coupled with Metal Forming Process.” Structural and Multidisciplinary Optimization 56 (6): 1571–1587.
Tang, Y., J. Chen, and J. Wei. 2012. “A Sequential Algorithm for Reliability-Based Robust Design Optimization Under
Epistemic Uncertainty.” Journal of Mechanical Design 134 (1): 1–10.
Wang, P., Z. Wang, and A. T. Almaktoom. 2013. “Dynamic Reliability-Based Robust Design Optimization with Time-
Variant Probabilistic Constraints.” Engineering Optimization 246 (6): 784–809.
Wu, Y. T., Y. Shin, R. Sues, and M. Cesare. 2001. “Safety-Factor Based Approach for Probability-Based Design Optimiza-
tion.” In Proceedings of the 19th AIAA Applied Aerodynamics Conference, Fluid Dynamics and Co-located Conferences,
AIAA 2001-1522. Reston, VA: AIAA.
Youn, B. D., K. K. Choi, and K. Yi. 2005. “Performance Moment Integration (PMI) Method for Quality Assessment in
Reliability-Based Robust Design Optimization.” Mechanics Based Design of Structures and Machines 33 (2): 185–213.
Zhang, Y., X. Xu, G. Sun, X. Lai, and Q. Li. 2018. “Nondeterministic Optimization of Tapered Sandwich Column for
Crashworthiness.” Thin-Walled Structures 122: 193–207.
Zhao, Y. G., and T. Ono. 1999. “A General Procedure for First/Second-Order Reliability Method (FORM/SORM).”
Structural Safety 21 (1): 95–112.
Zhuang, X., R. Pan, and X. Du. 2015. “Enhancing Product Robustness in Reliability-Based Design Optimization.”
Reliability Engineering and System Safety 138: 145–153.