See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/332256296

A sequential constraints updating approach for Kriging surrogate model-

assisted engineering optimization design problem

Article  in  Engineering With Computers · April 2019

DOI: 10.1007/s00366-019-00745-w

CITATIONS READS

15 334

5 authors, including:

Jiachang Qian Jiaxiang Yi

11 PUBLICATIONS   46 CITATIONS   
Huazhong University of Science and Technology
7 PUBLICATIONS   15 CITATIONS   
SEE PROFILE
SEE PROFILE

Yuansheng. Cheng Liu Jun

Huazhong University of Science and Technology Huazhong University of Science and Technology
86 PUBLICATIONS   1,283 CITATIONS    47 PUBLICATIONS   362 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Research on failure mechanisms of foam-filled hybrid sandwich structure under air blast loading View project

Multi-fidelity Surrogate Model View project

All content following this page was uploaded by Qi Zhou on 07 May 2019.

The user has requested enhancement of the downloaded file.

Engineering with Computers
https://doi.org/10.1007/s00366-019-00745-w

ORIGINAL ARTICLE

A sequential constraints updating approach for Kriging surrogate

model‑assisted engineering optimization design problem
Jiachang Qian1,2 · Jiaxiang Yi1 · Yuansheng Cheng1,3 · Jun Liu1,3 · Qi Zhou4

Received: 14 January 2019 / Accepted: 30 March 2019

© Springer-Verlag London Ltd., part of Springer Nature 2019

Abstract
Kriging surrogate model has been widely used in engineering design optimization problems to replace computational cost
simulations. To facilitate the usage of the Kriging surrogate model-assisted engineering optimization design, there are still
challenging issues on the updating of Kriging surrogate model for the constraints, since there exists prediction error between
the Kriging surrogate model and the real constraints. Ignoring the interpolation uncertainties from the Kriging surrogate
model of constraints may lead to infeasible optimal solutions. In this paper, general sequential constraints updating approach
based on the confidence intervals from the Kriging surrogate model (SCU-CI) are proposed. In the proposed SCU-CI
approach, an objective switching and sequential updating strategy is introduced based on whether the feasibility status of
the design alternatives would be changed because of the interpolation uncertainty from the Kriging surrogate model or not.
To demonstrate the effectiveness of the proposed SCU-CI approach, nine numerical examples and two practical engineer-
ing cases are used. The comparisons between the proposed approach and five existing approaches considering the quality
of the obtained optimum and computational efficiency are made. Results illustrate that the proposed SCU-CI approach can
generally ensure the feasibility of the optimal solution under a reasonable computational cost.

Keywords  Kriging surrogate model · Sequential constraint updating · Prediction interval · Feasibility

1 Introduction [1–3]. Just take the airfoil simulation case as an example, it

is reported that it takes the designer about 7–10 h to produce
In practical engineering design, computer simulation models a single output using the CFD model [4]. It is unrealistic to
have often been employed to replace the costly controlled directly use these simulation models to evaluate a large num-
real-life experiments. However, running simulation models, ber of design solutions when optimizing the design space. A
e.g., computational fluid dynamics (CFD) and finite-element promising way to address this issue is to adopt the surrogate
analysis (FEA), for complex engineering system with mul- model, also called the metamodel or approximation model,
tiple inputs and outputs can be computationally prohibitive to replace the computational expensive simulation model
[5–8]. There are several commonly used surrogate models,
such as Kriging model [9–11], Support Vector Regression
* Qi Zhou (SVR) model [12–14], Polynomial Response Surface (PRS)
qizhouhust@gmail.com; qizhou@hust.edu.cn model [15], and Radial Basis Function (RBF) model [16,
1
School of Naval Architecture and Ocean Engineering, 17]. Among these surrogate models, Kriging model is the
Huazhong University of Science and Technology, most thoroughly studied surrogate model to improve the
Wuhan 430074, Hubei, People’s Republic of China computational efficiency of engineering design optimization,
2
Wuhan Second Ship Design and Research Institute, because it has several interesting characteristics compared to
Wuhan 430064, Hubei, People’s Republic of China other surrogate models [18, 19]. First, the Kriging model can
3
Collaborative Innovation Center for Advanced Ship provide the prediction bias of each un-sampled point. The
and Deep-Sea Exploration (CISSE), Shanghai 200240, prediction bias obtained can reflect the level of prediction
People’s Republic of China confidence at the un-sampled point. The lower the prediction
4
School of Aerospace Engineering, Huazhong University bias, the higher the degree of confidence in the prediction,
of Science and Technology, Wuhan 430074, Hubei, and vice versa. Second, although different surrogate models
People’s Republic of China

13
Vol.:(0123456789)
 Engineering with Computers

perform well under different situations, preliminary results handling constraints and made a comparative among them,
illustrate that in most cases, Kriging model performs better concluding that further improvements are needed, since these
than other surrogate models regarding modeling accuracy, methods may discard solutions if the PoF of design alter-
especially when the number of sample points is small and natives close or equal to zero. Sasena et al. [42] illustrated
the black-box functions exhibit strong nonlinearity [20–23]. that the expected violation (EV) method could address the
As reported in the literature, the Kriging model-assisted shortcoming of the PoF approach. Bichon et al. [43] proposed
engineering design optimization can be classified into two an efficient global reliability analysis (EGRA) approach,
types: off-line type and on-line type. In the off-line type, a which shares a similar idea with EV, for updating constraints
pre-specified amount of sample points is employed to build with the value likely to be zero. Furthermore, Li et al. [44]
a Kriging model, which is used for replacing the simulation adjusted it from the global design region to the local one in a
analysis in the engineering optimization subsequently [24]. honeycomb material design problem. However, it is reported
The main shortcoming of the off-line type is that it is dif- that EV method evaluates design points from a finite set,
ficult to predetermine the proper sample size for obtaining which may not be able to locate the optimum of the original
an accurate Kriging model [25, 26]. On the other hand, the problem as accurately as solving the constrained one [42].
on-line type generates an initial Kriging model first and then Recently, Shu et al. [45] developed a weighted accumulative
adaptively updates the Kriging model following certain met- error sampling (WAE) strategy for updating metamodels and
rics, the distance criteria [27], prediction variance [28, 29], then applied it to improve the quality of global optimization.
cross-validation error [30], etc. Compared with the off-line Liu et al. [46] proposed a DIRECT-type constraint-handling
type, the on-line type can make use of the knowledge from technique that can separately handle feasible and infeasible
the previous iterations and is reported to be more efficient for cells. Dong et al. [47] developed a novel multi-start space
engineering design optimization. So far, most of the existing reduction (MSSR) algorithm for solving computationally
on-line Kriging model-assisted engineering design optimiza- expensive black-box global optimization problems with
tion focus on sequential updating of the objective function. bound constrains and nonlinear constrains. Shi et al. [48]
The most famous method, called efficient global optimization developed a filter-based adaptive Kriging method, in which
(EGO), was proposed by Jones et al. [31], where the Krig- a probability of constrained improvement (PCI) criterion is
ing surrogate model is updated by selecting the point that developed based on the notion of filter to sequentially gener-
maximizes the expected improvement (EI) over the present ate new samples for updating Kriging metamodels of objec-
best solution. Furthermore, Sóbester et al. [32] proposed a tive and constraints. Wu et al. [49] extended the adaptive
weighted expected improvement criterion, which is designed metamodel-based global optimization algorithm (EAMGO)
to allow a more flexible way to balance the exploration and to handle constrained global optimization problems, in which
exploitation in global searching. Wang et al. [33] developed a the updating strategy for the constraints can be used for dif-
boundary and best neighbor searching (BBNS) approach for ferent surrogate models. However, because the interpolation
selecting sampling points that are not only in the neighbor- uncertainties are not considered in all updating iterations,
hood of the present best solution but also near the boundary the feasibility of the design alternatives that changed due
of the design space. Zhan et al. [34] proposed a parallel EGO to these interpolation uncertainties will mislead the global
algorithm, where multiple updating points were selected in searching direction.
each cycle by dynamically controlling the number of EI max- To address the issue-mentioned above, a general sequential
ima and points around each EI maximum. Dong et al. [35] constraint-updating approach based on the confidence inter-
proposed a hybrid surrogate-based optimization algorithm, vals from the Kriging surrogate model (SCU-CI) is proposed.
in which the update points need to satisfy a defined distance In the developed SCU-CI approach, an objective switching
criterion. While a limited number of approaches consider and sequential updating strategy is introduced to determine,
the updating strategy for the constraints [36]. Schonlau et al. (1) whether the actual simulation models or the Kriging
[37] proposed a probability of feasibility (PoF) approach to metamodel should be used to evaluate the feasibility of the
identify regions of feasibility, i.e., the probability of the pre- design alternatives and (2) which design alternatives would
diction at design alternatives will be larger or smaller than be selected to improve the prediction accuracy of the Kriging
a constant limit. Li et al. [38] applied the PoF approach to metamodel. The developed updating strategy is mainly based
finding initial feasible sample points in a Kriging-based on whether the feasibility status of the design alternatives
constrained global optimization algorithm. Wang et al. [39] would be changed because of the interpolation uncertainty
extended the PoF approach for the design optimization of the from the Kriging surrogate model or not. The performance
black-box stochastic systems. Zhang et al. [40] adopted the of the proposed approach is illustrated using nine numerical
PoF approach to address the global constraint optimization, cases and two real-world engineering examples. The com-
where multi-fidelity analysis models are available. Parr et al. parisons between the proposed approach and some existing
[41] provided a review of three typical updating strategies for approaches considering the quality of the obtained optimum

13
Engineering with Computers

and computational efficiency are made. The merits of the pro- ̂ −1 (y − 1𝜇)
(y − 1𝜇)R ̂
posed approach are analyzed and summarized. 𝜎̂ 2 = . (5)
N
The rest of this paper is organized as follows. In Sect. 2,
the background of the Kriging model and several typical Substituting Eqs. (4) and (5) to Eq. (3), an expression can
constraint-updating strategies are presented. Details of the be got for parameters are 𝜃 and p
proposed are introduced in Sect. 3. In Sect. 4, the compari- n ̂ −1 (y − 1𝜇)
(y − 1𝜇)R ̂ 1
son results between the proposed approach and some exist- ln(L) = − ln( ) − ln |R|. (6)
2 N 2
ing approaches on a numerical benchmark case and two real-
world engineering design problems are presented. Finally, The maximum likelihood estimation should be maxi-
the concluding remarks are given in Sect. 5. mized, so that the hyper-parameters can be got, which is the
best fit to the Kriging model. It is difficult to get an analyti-
cal solution of Eq. (6); therefore, a global search algorithm,
2 Background e.g., genetic algorithm, is used for finding the optimum of
𝜃 and p.
2.1 Kriging model Then, the predicted value at an un-sampled point x can
be calculated as
Kriging model is a kind of interpolation model, which was
f̂ (x) = 𝜇̂ + rT R−1 (y − 1𝜇),
̂ (7)
developed by Danie G. Krige for forecasting the mining
holes. Then, it was introduced to approximate the com- where y is the response value of the sample points, R is the
puter experiments by Sacks et al. correlation matrix of the sample points, and r is a vector,
{ [50]. Supposing
} that whose dimension is]N × 1. The elements of the vector r are
there are N sample points X = x(1) , x(2) , … , x(N) with [
{ } ri = Cor y(x), y(x(i) ) .
x(1) = x1(1) , x2(1) , … , xk(1) and the corresponding response The advantage of the Kriging model compared with other
{ }
is y = y(1) , y(2) , … , y(N)  . The Kriging model can be surrogate models is that it can provide the variance value of
expressed as the predict points, which can be expressed as
[ ]
(1 − 1T R−1 r)2
f̂ (x) = p(x) + Z(x), (1) 2 2 T −1
s (x) = 𝜎̂ 1 − r R r + . (8)
1T R−1 1
where p(x) denotes the mean of the Gaussian process, indi-
cating the global property, Z(x) represents the local property The variance function can be calculated in every pre-
with the mean value being equal to zero, and the variance dicted point, so that a confidence interval can be got to fore-
is 𝜎. cast the range of forecasting. A one-dimensional example is
The variables of the Kriging model were considered to presented in Fig. 1. The function which is approximated has
be correlated with each other through the basis function: an expression of f (x) = (6x − 2)2 sin(12x − 4) . The sample
[ ] [ ] points are X = [0, 0.25, 0.5, 0.75, 1]T . In Fig. 1, the solid line
R y(x(i) ), y(x(j) ) = Cor y(x(i) ), y(x(j) )
( k ) represents the real response value of the function, and the
∑ | (2) dashed line represents the predicted value through the Krig-
j | m
p
= exp − 𝜃m |xm − xm |
i
,
| | ing surrogate model; the dot lines denote the upper bound
m=1
and lower bound. The predict response at sample points
[ ] equal to the real response and the errors equal to zero. The
where R is the correlation coefficient, R y(x(i) ), y(x(j) )
bounds consist of the confidence interval of the Kriging
denotes the correlation between x(i) and x(j) , 𝜃m , and pm are
model, of which has 95.5% probability that the points would
the hyper-parameters that influence the correlation of the
fall into it.
sample points and the smoothness of the surrogate model.
The hyper-parameters can be obtained by the maximum
2.2 Descriptions of five typical constraint‑updating
likelihood estimation:
strategies
� � � ∑ (i) �
(N) � 1 (y − 𝜇)2
L y , y , … , y �𝜎, 𝜇 =
(1) (2)
� n exp − . A general formulation of a constraint optimization problem
(2𝜋𝜎 2 ) ∕2 2𝜎 2
is given as:
(3)
The value of 𝜇 and 𝜎 can be obtained by setting the deri- min f (x)
vate of Eq. (3) to be zero x
s.t. gj (x) ≤ 0, j = 1, 2 … J (9)
1T R−1 y xlb ≤ x ≤ xub ,
𝜇̂ = T −1 , (4)
1 R 1

13
 Engineering with Computers

Fig. 1  Illustration of the Krig-

ing model with its confidence
interval

where f (x) is the objective function, x = (x1 , x2 , … xN )T (2) PoF method

is the design variable vector, xlb and xub are the lower and The PoF approach is based on the feasible probability of
upper bounds of x , respectively, and g = (g1 , g2 , … , gJ ) are the candidate design points, and the constraint functions are
the constraints. classified into two types (i.e., “inexpensive” and “expen-
In this work, we focus on the constraints, which involve sive”) according to computing time. The feasible probability
expensive simulations, i.e., we assume that the evaluation of of inexpensive constraints is 0 or 1, because it is quite easy
the objective is cheap. This is common in the design of engi- to verify whether the candidate points satisfy the constraints
neering structure, i.e., the minimum weight design of struc- or not. It can be expressed as [41]
tures. Therefore, in this section, brief introductions of five {
1, if g(x) ≤ 0
typical constraint-updating strategies, including EV, PoF, Pinexp =
0, otherwise
. (11)
maximum MSE (MAX-MSE) method, WAE, and EAMGO
methods, are introduced. As to expensive constraint functions, the feasible prob-
ability needs to be estimated from the statistical models.
(1) EV method The estimated the feasible probability could be expressed as
EV method is one of the most commonly used constraint- ( )
0 − ĝ (x)
handling strategies. The typical characteristic of this method Pexp = 𝛷 , (12)
is that the constraints are checked before the sampling crite- ŝ (x)
rion is taken into consideration. For every candidate point, where 𝛷 refers to the Gaussian cumulative distribution func-
the expected violation is calculated as [42] tion. ĝ (x) and ŝ i (x) are Kriging predicted mean and stand-
{ ( ) ( ) ard deviation, which are obtained from Eqs. (7) and (8),
ĝ −0 ĝ −0
(̂gi − 0)𝛷 iŝ + ŝ i 𝜙 iŝ , if ŝ i > 0 respectively.
EVi = i i , (10)
0, if ŝ i = 0
(3) MAX-MSE method
where ĝ i refers to the predicted mean of constraint i  , and The MAX-MSE method is a general updating strategy not
ŝ i is the standard deviation. It is worth mentioning that the only for handling constraints but also for objective functions.
expected violation plays a similar role as the EI function. The main idea of The MAX-MSE method is that adding the
When the constraint has prominent uncertainty or is likely to sample point with the largest variance is the most beneficial
be violated, the expected violation trends to be large. As to to improve the prediction accuracy of the Kriging model.
the problem with several constraints, the EV can be a vector. Since the Kriging model assumes that the correlations
between the design points are dependent on their distances,

13
Engineering with Computers

the accuracy of the Kriging model for the constraints in the where k is the current iteration number, dmax denotes the
neighborhood of the selected samples will also increase maximum distance between two sample points in the current
obviously. The formula can be expressed as [51] sampling set, 𝜁iter is a distance coefficient obtained from a
constant array, and xi denotes the ith sample in the current
max s2 (x), (13) sample set.
where s (x) is the variance value, which can be obtained
2

by Eq. (9). The MAX-MSE value is found through an opti-

mization method such as Direct Search when dealing with 3 Proposed approach
practical problems.
In this work, a genetic algorithm (GA) from Coello [52] is
(4) WAE method used to solve the optimization problems. If the constraints
In the WAE method, the optimization objective is to find are calculated by running the computational expensive sim-
a sample point with the maximum weighted accumulative ulation models, it will be unrealistic to directly use these
predicted error obtained by analyzing data from the previous simulation models to evaluate a large number of design
iterations, and a space-filling criterion is also introduced and solutions when optimizing the design space. To improve
treated as a constraint to avoid generating clustered sample the effectiveness of solving this problem, the Kriging surro-
points. To some extent, the WAE method is an extension of gate model instead of the actual constraints can be used. As
the MAX-MSE method, which can be expressed as mentioned in Sect. 2.1, there are prediction errors between
√ the Kriging surrogate model and the real constraint values,
√ n
√∑
eWAES (x) = √ wi (̂y−i (x) − ŷ (x))2 (14) which may lead to infeasible optimal solutions. The purpose
i=1 of the proposed approach is to sequentially update the Krig-
ing surrogate models to ensure the feasibility of the opti-
where ŷ (x) denotes the Kriging model constructed using the mization solutions. The flowchart of SCU-CI approach is
current sample set, while ŷ −i (x) is the Kriging model con- demonstrated in Fig. 2. The Step 4 and Step 5 are the cores
structed using the current sample set without the ith sample.
The value of wi reflects the influence of different sample
points on the error on x. Begin
In addition, the WAE method limits the minimum dis- Step 1
tance between a new point and existing points by the fol- Generate initial sample points and obtain the response values
lowing criterion: Step 2

min‖ ‖ Construct initial Kriging model

‖x − xi ‖ ≥ d, (15)
Step 3
where d denotes the threshold distance, which is calculated Initial the population of GA, Set N=1,evaluate fitness values
by the average value of the minimum distance between each of initial individuals by Kriging model
pair of the sample points. Step 4
Select the individuals whose feasibility may change due to the
(5) EAMGO method Kriging model uncertainty
EAMGO is different from adaptive metamodel-based Step 5
global optimization (AMGO) for its constrains updat- Determine the new sample points by a distance criterion
ing strategy. During each iteration, the updating point is Step 6
obtained by minimizing the summation of predicted con- Update the Kriging model. Generate new population and
straint values subject to the approximate constraints and the update the count number N=N+1
distance limitation. The constraints updating strategy in the Step 7
EAMGO method can be described as Evaluate the responses of the new individuals by Kriging
model
∑
J
Step 8 No
min ĝ kj (x) Satisfy the Convergence condition?
x
j=1
Step 9 Yes
s.t. ĝ kj (x) ≤ 0, j = 1, 2 … J (16)
Output the optimization solutions
‖x − x ‖ ≥ 𝜁 d
‖ i‖ iter max
Stop
xlb ≤ x ≤ xub ,

Fig. 2  Flowchart of the proposed approach

13
 Engineering with Computers

of SCU-CI and will be described in details in Sects. 3.1 and In each generation of GA, when the effects of interpola-
3.2. tion uncertainty from the Kriging surrogate model are taken
into consideration, there are four possible scenarios for a
3.1 The developed objective switching criterion design alternative, as shown in Fig. 3.
In the first scenario, as shown in Fig. 3a, point A is a
Prediction values from Kriging surrogate models have pre- feasible design alternative and the confidence interval does
diction uncertainty, which may lead to infeasible optimi- not intersect the constraint boundary. It means that point A
zation solutions. Therefore, the SCU-CI approach is pre- would be feasible at the given confidence probability. Simi-
sented to take the interpolation uncertainty of the Kriging larly, in the fourth scenario, point D would be an infeasible
surrogate model into consideration. Note that as long as the solution at the given confidence probability, as shown in
feasibilities of the design alternatives may not change due Fig. 3d. In summary, for these two scenarios, although the
to the Kriging surrogate model uncertainty, the feasibilities Kriging surrogate model has prediction errors at point A and
of the individuals can be predicted by the Kriging surrogate point D, these prediction errors do not affect the feasibilities
models instead of expensive simulation models. However, if of the two points. Therefore, there is no need to add these
the feasibilities of the design alternatives may change, then points as new sample points for updating the Kriging sur-
the feasibilities of the individuals should be evaluated by rogate model in the subsequent optimization process.
simulation models. Thus, an objective switching criterion is In the second scenario, as shown in Fig. 3b, point B is fea-
introduced in SCU-CI to determine whether the simulation sible, while its confidence interval intersects the constraint
models or the Kriging surrogate models should be used to boundary. This means that point B may be infeasible at the
evaluate the feasibilities of individuals. given confidence probability. Similarly, in the third scenario,
point C is infeasible, while it may become feasible due to the
-2σ

Feasible Region Feasible Region

A
2σ

-2σ

B
2σ

Infeasible Region Infeasible Region

(a) (b)

Feasible Region Feasible Region

-2σ

C
2σ

-2σ

D
2σ

Infeasible Region Infeasible Region

(c) (d)

Fig. 3  Four possible scenarios for a design alternative

13
Engineering with Computers

prediction error from the Kriging surrogate model. For these Table 1  Algorithm for determining new sample points satisfying the
two scenarios, the feasibilities of points B and C are uncer- distance constraint
tain. Therefore, the new sample points should be selected Input: the selected points in Sect. 3.1
from the points, whose feasibilities are uncertain. The condi-
Begin
tions that these points satisfy can be described as:
Selected individuals ← Selected points in Sect. 3.1 are
ĝ i (xj ) ≤ 0 & ĝ i (xj ) + 2s(xj ) ≥ 0 i = 1, 2, … , J SI = {SI1 SI2 … SIl } sorted in descending order of confi-
(17) Existing sample points dence interval intervals.
ĝ i (xj ) ≥ 0 & ĝ i (xj ) − 2s(xj ) ≤ 0 i = 1, 2, … , J, SP = {x1 x2 … xn }
for k = 1:l
where ĝ i (x) are the predicted constraint value of the opti- d(SIk ) ← Calculate the distance metric for SIk
mization problem; s(x) is the standard deviation of the cor- d̄ ← The individual which satisfy the
if d(SIk ) >
responding Kriging surrogate model. n
constraint is determined as a new
SP = {SP SIk }
sample point
end if
3.2 Distance metric
end for
Output: The new sample points
It is noted that it may cause unnecessary computational cost
if the points selected according to the criterion in Sect. 3.1
are all evaluated by simulation models. This is because (1)
Step 1: Generate initial sample points and obtain the
these sample points may be very close in the design space,
response values. The essential requirement of initial sam-
resulting in the oversampling for Kriging surrogate model
pling is that sample points should distribute uniformly in
[45, 53] and (2) when a new sample point is used to update
whole design space. In this paper, the optimal Latin hyper-
the Kriging surrogate model, the prediction error at an unob-
cube design (OLHD) is used for initial sampling.
served point near the new sample point will be significantly
Step 2: Construct the initial Kriging surrogate model.
reduced. It indicates that the points, which are very close to
Step 3: Initialize the population of GA, Set gen = 1, evalu-
other existing sample points, should not be selected as new
ate the fitness values of initial individuals by Kriging sur-
sample points.
rogate model.
To prevent the cluster of sample points, the points that
Step 4: Select the individuals whose feasibility may
satisfy the following constraint are selected as new sample
change due to the interpolation uncertainty from the Krig-
points:
ing surrogate model according to Eq. (17).
d(x) >
d̄ Step 5: Determine the new sample points by a distance
𝛿
(√ ) criterion according to Eq. (18).
( )T ( )
d(x) = min x − xi(n) x − xi(n) (i = 1, 2, … , N (n) ) Step 6: Update the Kriging surrogate model. Generate the
(√ ) new population and update the count number gen = gen + 1.
∑
n
( )T ( ) Step 7: Evaluate the responses of the new individuals by
d̄ = min xi − xj(n) xi − xj(n) j = 1, 2, … , i − 1, i + 1, … , n,
Kriging model
i=1
(18)
Step 8: Check whether the stopping criterion is satisfied:
If yes, go to Step 9; otherwise, go back to Step 4. The stop-
where 𝛿 is the distance threshold control parameter which is
ping criterion satisfies one of the following conditions:
set to be 2 in this work, x is the sample points to be judged,
and xi(n) is the nth sample point in the current sample points {
| | | |
|ygen − ygen−1 | ≤ 𝜀y , |ygen−1 − ygen−2 | ≤ 𝜀y
set. The algorithm for determining new sample points that | | | | , (19)
meet the requirements is listed in Table 1. gen = MAXGEN

3.3 Steps of the proposed approach where ygen−2  , ygen−1  , and ygen are the optimal solutions
obtained in the (gen − 2)th , (gen − 1)th , and genth genera-
The detailed steps of the proposed approach are described tion, 𝜀y is the relaxation factor, and MAXGEN is the maxi-
as follows. mum generation of GA.
Begin Step 9: Output the optimization solutions.
End

13
 Engineering with Computers

4 Examples and results approaches: (1) EV method [42], (2) MAX-MSE method,

(3) PoF method [37, 40], (4) WAE method [45], and (5)
In this section, nine numerical examples and two practical EAMGO method [49]. Since the true optimal solution for
engineering cases are used to demonstrate the effectiveness a numerical example can be obtained, the stopping crite-
and merits of the proposed approach. In the test cases, all ria used for the numerical example are different from those
the five updating strategies are used in the framework of GA for the engineering problems for a better comparison of the
as the proposed SCU-CI strategy. A Kriging model with the performance of different approaches. The stopping criterion
same sample points is built in the initial stage of the GA satisfying one of the following conditions:
method. Then, the Kriging model is used to evaluate the
⎧ � ygen − y ∗ �
fitness value during the evolution process. The difference ⎪ �� �≤𝜀
�
among all the approaches is the updating strategy, where ⎨� y ∗ � , (21)
the alternatives are selected. To provide a detailed descrip- ⎪ n = MAXGEN
⎩
tion of the implementation process of the proposed SCU-
CI approach, one of the numerical examples is used as an
where ygen is the optimal solution obtained in the genth gen-
illustrative example. The parameters set for the GA are as
eration; y ∗ is the true optimal solution of the numerical
follows: the population size and max iterations are 40 and
example, and MAXGEN is the maximum generation of GA.
100, respectively. The crossover probability and mutation
𝜀 is set to be 0.02 in this example.
probability are 0.80 and 0.15, respectively. The generation
For the initial Kriging surrogate model, 20 initial sample
gap is 0.95 to keep the elites in each generation.
points are generated by OLHD for different approaches. The
comparison results are summarized in Table 2.
4.1 Illustration example
As can be seen in Table 2, although the proposed SCU-
CI approach required more sample points than those of EV,
A numerical case from Zhou et al. [54] is used to describe
PoF, and WAE approaches, it can obtain the best optimal
the proposed approach steps by steps. The object is to find
solution at the same time satisfying the constraints. Com-
the minimum value of the objective function subject to two
pared to EAMGO, SCU-CI can obtain better optimal solu-
constraints. The expression of this optimization problem is
tion while requires fewer sample points. In addition, EV,
given as
PoF, and WAE approaches violate the constraint and lead to
min f = x13 sin(x1 + 4) + 10x12 + 22x1 + 5x1 x2 + 2x22 + 3x2 + 12 an infeasible optimal solution. The obtained optimal solu-
s.t. g1 = x12 + 3x1 − x1 sin(x1 ) + x2 − 2.75 ≤ 0 tions of the six methods are described in Fig. 4. From Fig. 4,
it can be seen that only the optimal solutions obtained by
g2 = − log2 (0.1x1 + 0.41) + x2 e−x1 +3x2 −4 + x2 − 3 ≤ 0
(20) the proposed SCU-CI approach and EAMGO fall into the
feasible region and are very close to the constraint boundary,
−3 ≤ x1 ≤ 1.5, − 4 ≤ x2 ≤ 1.5. while the optimal solutions obtained by other approaches
are infeasible.
In this example, it is assumed that the original nonlinear To demonstrate the sequential updating process of the
constraint g2 is an expensive cost model, while the objective proposed SCU-CI approach, Fig. 5 provides the sample
function f and g1 can be expressed explicitly. Two metrics, points and their confidence intervals of the constraint in dif-
the obtained optimal solution and the required number of ferent generations of GA.
sample points (NS), are employed in this work to evaluate As shown in Fig. 5a, the individuals in the first genera-
the accuracy and efficiency of the proposed method. Both tion are randomly distributed in the design space, and the
of these two metrics are expected to be smaller values. For confidence intervals of the Kriging model of the constraint
illustration, the proposed approach is compared with five are large for most individuals. With the optimization process

Table 2  Comparison results of x1 x2 ĝ 2 Optimal solution g1 g2 NS

the different approaches for the
illustration example EV − 1.9178 0.7269 − 0.2514 − 3.2959 − 5.9021 0.0512 36
MAX-MSE − 1.8323 0.7523 − 0.1424 − 3.3288 − 5.9072 0.0588 74
PoF − 1.9083 0.7316 − 0.0075 − 3.3096 − 5.9023 0.0607 48
WAE − 2.1004 0.7458 − 5.3995 − 3.3446 − 5.7064 0.4011 42
EAMGO − 1.7512 0.7456 − 0.0689 − 3.2219 − 5.9384 − 0.0689 77
SCU-CI − 1.9013 0.7164 − 0.0220 − 3.2514 − 5.9209 − 0.0154 62

13
Engineering with Computers

Fig. 4  Optimal solutions from

different approaches

Fig. 5  Individuals, selected
sample points, and their confi-
dence level in different genera-
tions of GA

goes on, the new sample points are added around the con- and the confidence intervals are becoming smaller. As
straint boundary. It can be seen in Fig. 5b that the individuals shown in Fig. 5c, all the individuals are near the constraint
in the sixth generation are closer to the constraint boundary boundary with small confidence intervals (most of them

13
 Engineering with Computers

Fig. 6  Added sample point dia-

gram of the proposed sequential
sampling method

close to being 0) in the last generation of GA. The newly

1
added sample points of the proposed SCU-CI approach are min f = 51.95 ((ax2 − bx12 + cx1 − d)2 + h(1 − ff ) cos(x1 ) − 44.81)
plotted in Fig. 6. In Fig. 6, the red triangles represent the s.t.
newly added sample points, the black dash line and the red g = max {gj } ≤ 0
j∈{1,2}
dash-and-dot line denote the two constraints, and the blue g1 = 1
(x (1 − x2 ) − x2 )
10 ( 1 ( ))
domain is the feasible region of the optimization problem. 1 (x −5)2
g2 = 10 1 − 1 8 + 2 4
(x −15)2

As illustrated in Fig. 6, the newly added sample points of a = 1, b = 4𝜋5.1

, c = 𝜋5 , d = 6, h = 10, ff = 1
the proposed SCU-CI approach are distributed around the
2 8𝜋
−5 ≤ x1 ≤ 10, 0 ≤ x2 ≤ 15.
constrained boundary. (23)
It can be concluded from Figs. 5 and 6 that the search-
ing space of the proposed SCU-CI approach can adaptively
New Branin
locate the regions, in where the feasibility states of the indi-
viduals could be changed due to the prediction uncertainty 1 ( )
min f = 80 −(x1 − 10)2 − (x2 − 15)2
from the Kriging surrogate model. This can result in a large
s.t.
probability to get a more desirable solution. 1
g = 51.95 ((ax2 − bx12 + cx1 − d)2 + h(1 − ff ) cos(x1 ) − 5 + h) ≤ 0
5.1
a = 1, b = 4𝜋 , c = 𝜋5 , d = 6, h = 10,
4.2 Additional numerical test cases 1
2

ff = 8𝜋 − 5 ≤ x1 ≤ 10, 0 ≤ x2 ≤ 15. (24)

In this section, eight additional numerical test cases from
Wang et al. [55] are used to illustrate the effectiveness of the Sasena
proposed SCU-CI approach. The formulations of these test
cases are described as follows. A more detailed description min f = −(x1 − 1)2 − (x2 − 0.5)2
of these examples, e.g., characteristics, actual optimal solu- s.t.
tions, can be found in Wang et al. [55]: g = max {gj } ≤ 0
j∈{1,2,3}
Gomez g1 = ((x1 − 3)2 + (x2 + 2)2 )e−x2 − 12
7
(25)
( ) ( )
g2 = 10x1 + x2 − 7
min f = 4 − 2.1x12 + 31 x14 x12 + x1 x2 + − 4 + 4x22 x22 g3 = (x1 − 0.5)2 + (x2 − 0.5)2 − 0.2
s.t. 0 ≤ xi ≤ 1 for i = 1, 2.
g = − sin(4𝜋x1 ) + 2 sin2 (2𝜋x2 ) ≤ 0
−1 ≤ xi ≤ 1 for i = 1, 2.
(22) Qcp4
Constrained Branin

13
Engineering with Computers

min f = −2x1 + x2 − x3 Angun

s.t. 5
g = max {gj } ≤ 0 min f = 34.8 (5(x1 − 1)2 + (x2 − 5)2 + 4x1 x2 )
j∈{1,2,3} s.t.
g1 = x1 + x2 + x3 − 4 g = max {gj } ≤ 0
g2 = 3x2 + x3 − 6 (26) j∈{1,2}
(29)
5
′ ′ ′ 2
g3 = −x A Ax + 2y Ax − y + 0.25b − z 2 g1 = 17.1544 ((x1 − 3)2 + x22 + x1 x2 − 4)
5
A = [0, 0, 1;0, −1, 0; − 2, 1, −1], g2 = 17.1544 (x12 + 3(x2 + 1.061)2 − 9)
b = [3;0; − 4], y = [1.5; − 0.5; − 5], z = [0; − 1; − 6] 0 ≤ x1 ≤ x3 , −2 ≤ x2 ≤ 1.
x = [x1 ;x2 ;x3 ].0 ≤ x1 ≤ 2, 0 ≤ xi ≤ 3 for i = 2, 3.

Since the actual optimal solutions for these numerical

G4 examples are known, the stopping criteria used are the same
5.3578547x32 +0.8356891x1 x5 +37.293239x1 −40792.141 as that of the illustration example, i.e., the max iterations
min f = 2000 reach or the difference between the optimal solution and the
s.t.
true solution is in the predefined domain, the algorithm will
g= max {gj } ≤ 0
j∈{1,2,3,4,5,6} stop. To consider the randomness of the GA, 30 different
g1 = 0 − u
runs are carried out for each numerical example. As men-
g2 = u − 92
g3 = 90 − v
tioned before, the obtained optimal solution may violate the
g4 = v − 110 actual constraints due to the interpolation uncertainties from
g5 = 20 − w the Kriging surrogate models. For constrained optimization
g6 = w − 25 problems, it is known that satisfying the constraints is the
u = 85.334407 + 0.0056858x2 x5 + 0.0006262x1 x4 − 0.0022053x3 x5 first condition followed by the optimal objective function
v = 80.51249 + 0.0071317x2 x5 + 0.0029955x1 x2 + 0.0021813x32 and solving efficiency. Therefore, the ratios of the feasibility
w = 9.300961 + 0.0047026x3 x5 + 0.0012547x1 x3 + 0.0019085x3 x4
in 30 runs for each numerical case are recorded and com-
78 ≤ x1 ≤ 102, 33 ≤ x2 ≤ 45, 27 ≤ xi ≤ 45 for i = 3, 4, 5.
pared. A high ratio of the feasibility denotes a greater abil-
(27) ity to guarantee the feasibility of the solutions in surrogate
G24 model-assisted optimization design problems. Table 3 sum-
maries the ratios of the feasibility of the different approaches
min f = 57 (−x1 − x2 ) for the additional numerical test cases. An intuitive conclu-
s.t. sion can be made from Table 3 that the proposed SCU-CI
g = max {gj } ≤ 0 approach can ensure the feasibility of the solution for most
j∈{1,2}
(28) numerical test cases, which is not the case for other five
g1 = 41 (−2x14 + 8x13 − 8x12 + x2 − 2)
approaches. Table 4 summarizes the average ranking results
g2 = 14 (−4x14 + 32x13 − 88x12 + 96x1 + x2 − 36) for the six approaches considering the additional numerical
0 ≤ x1 ≤ 3, 0 ≤ x2 ≤ 4. test cases. As observed in Table 4, the average ranking of

Table 3  Ratios of the feasibility Test cases EV MAX-MSE PoF WAE EAMGO SCU-CI
of the different approaches for
the additional numerical test Gomez 0.000 (4) 0.000 (4) 0.000 (4) 0.333 (3) 1.000 (1) 0.900 (2)
cases
Constrained Branin 0.967 (3) 0.933 (4) 1.000 (1) 0.367 (5) 0.000 (6) 1.000 (1)
New Branin 0.933 (2) 0.867 (3) 0.867 (3) 0.200 (6) 0.867 (3) 1.000 (1)
Sasena 0.000 (4) 0.000 (4) 0.000 (4) 0.733 (2) 1.000 (1) 0.133 (3)
qcp4 0.833 (1) 0.700 (3) 0.000 (5) 0.067 (4) 0.000 (5) 0.833 (1)
G4 0.667 (3) 0.700 (2) 0.300 (4) 0.200 (5) 0.000 (6) 0.733 (1)
G24 0.333 (4) 0.733 (2) 0.467 (3) 0.167 (5) 0.000 (6) 0.833 (1)
Angun 1.000 (1) 1.000 (1) 1.000 (1) 1.000 (1) 1.000 (1) 1.000 (1)

Table 4  Average ranking results EV MAX-MSE PoF WAE EAMGO SCU-CI

for the six approaches for the
additional numerical test cases Ranking 2.750 2.875 3.125 3.875 3.625 1.375

13
 Engineering with Computers

Table 5  Adjusted p values for ratios of the feasibility obtained in the Table 7  NS of the EV, MAX-MSE, and SCU-CI for the additional
additional numerical test cases by Bergmann–Hommel’s dynamic numerical test cases
procedure
i Hypothesis pi value
i Hypothesis pi value
1 MAX-MSE vs. SCU-CI 0.0075
1 WAE vs. SCU-CI 0.0111 2 EV vs. SCU-CI 0.0450
2 EAMGO vs. SCU-CI 0.0134
3 PoF vs. SCU-CI 0.0231
4 MAX-MSE vs. SCU-CI 0.0614
5 EV vs. SCU-CI 0.0948

Table 6  NS of the EV, MAX-MSE, and SCU-CI for the additional

numerical test cases
Test cases EV MAX-MSE SCU-CI

Gomez 49.633 100.000 64.900

Constrained Branin 100.000 63.133 61.367
New Branin 2.333 1.933 1.367
Sasena 86.800 96.700 54.700
qcp4 100.000 100.000 93.667
G4 74.300 86.700 59.533
G24 16.667 26.933 10.333
Angun 13.367 12.067 12.833

the proposed SCU-CI approach is 1.375, which is the best Fig. 7  Structure and the loading of the bracket
among all approaches. EV ranks the second, followed by the
MAX-MSE. The average rank of WAE is the worst. from Table 7 that there are significant differences in effi-
In addition, p values are used to test the differences ciency between the proposed SCU-CI approach and MAX-
between approaches over multiple data sets regarding the MSE and EV.
ratio of the feasibility. The Bergmann–Hommel procedure
is adopted to calculate adjusted p values [56], as listed in 4.3 Design optimization of lattice structure design
Table 5. The p values for WAE vs. SCU-CI, EAMGO vs. of an L‑shaped bracket
SCU-CI, and PoF vs. SCU-CI are less than 0.05, indicat-
ing that there are significant differences in the ability to The proposed SCU-CI approach is applied to the design opti-
guarantee the feasibility of the solutions between the pro- mization of lattice structure design of an L-shaped bracket.
posed SCU-CI approach and these three constraint-updating The bracket is a simplified version of a proprietary aircraft
techniques. The differences between the proposed SCU-CI component. The bracket consists of three vertical walls and a
and the rest two approaches (MAX-MSE and EV) are not base. There are four support beams between the three walls.
significant. The bracket is fixed along the back side and subjected to a
To further compare the efficiency of MAX-MSE, EV, and distributed load of 0.065 N/mm2 at the bottom surface. The
the proposed SCU-CI, the mean values of NS are recorded. structure and the loading of the bracket are plotted in Fig. 7.
Notice that the quality of the optimal solution (objective The lattice model of the bracket is established using a
function value) is considered in the convergence condi- cross lattice type with a size of 10 mm. In this paper, the
tions. The comparison results of the NS are summarized ANSYS 18.0 is used for getting the simulation results, and
in Table 6. As shown in Table 6, the proposed SCU-CI the computational platform with a 3.70 GHz AMD Ryzen 7
approach required the smallest NS in six of the eight test 2700X Eight-Core Processor and 8 GB RAM is used. The
cases, indicating that SCU-CI approach is the most effective lattice model and the simulation result can be seen in Fig. 8.
approach among these three constraint-updating techniques. The material properties are listed in Table 8.
The p values are also used to test the differences between In this design optimization problem, the objective is to
approaches over multiple data sets regarding the efficiency. minimize the total volume of the bracket while keeping the
The adjusted p values are listed in Table 7. It is concluded

13
Engineering with Computers

Fig. 8  Lattice model and the simulation results of the bracket

Table 8  Material properties of the bracket initial sample points are generated for the initial Kriging
Material property Value
metamodel. The comparison results are listed in Table 9.
As shown in Table 9, the proposed approach can obtain a
Elasticity modulus 1478 N/mm2 feasible solution, which is very close to the solution obtained
Poisson ratio 0.28 by DSBA. However, the computation cost of simulation-
Yield stress 36 MPa based approach is about 17 times more than the proposed
approach, which is unacceptable for complex engineering
problems. The EV approach and PoF approach can also
maximal stress and displacement below the threshold values. obtain feasible solutions. However, the obtained solutions
The mathematical formulation of the problem is given as are inferior to that from the proposed SCU-CI. Concern-

find d1 , d2 , d3 , d4 , d5 , d6
min V
s.t. u ≤ 5mm
(30)
Smax ≤ 36MPa
0.7 mm ≤ d1 ≤ 1.5 mm, 0.3 mm ≤ d2 ≤ 1.0 mm, 0.1 mm ≤ d3 ≤ 0.7 mm
1.0 mm ≤ d4 ≤ 2.5 mm, 0.5 mm ≤ d5 ≤ 2.0 mm, 0.1 mm ≤ d6 ≤ 1.5 mm,

where u is the maximal displacement of the bracket, and Smax ing computational efficiency, the computational cost of
is the maximum equivalent stress. d1 , d2 , d3 , d4 , d5 and d6 are these two approaches is 146 and 150% higher than the pro-
the design variables, where d1 represents the radius of the posed approach, respectively. In addition, the MAX-MSE
frame pillar, d2 represents the radius of the frame cross pil- approach, EAMGO approach, and the WAE approach failed
lar, d3 represents the radius of the cross pillar, and d4 , d5 , d6 to obtain a feasible solution. The results indicate that the
represent the radius of the frame cross, pillars cross, and SCU-CI shows a better ability to handing the constrained
upper beams, respectively. optimization problem.
In this case, the constraints are obtained from the expen-
sive simulation models, while the objective function can be 4.4 Lightweight design optimization of conical
calculated by arithmetic expression. The proposed SCU- shell with longitudinal and circumferential
CI approach and the existing five approaches are used to stiffeners
solve this design optimization problem. For the engineering
problem, we also compare the proposed SCU-CI approach In this section, the proposed approach is applied to the
with the direct simulation-based approach (DSBA), in which design optimization of a conical shell with longitudinal
simulation model instead of Kriging surrogate model is and circumferential stiffeners. The conical shell is a non-
directly used during the design optimization process. 60 pressures underwater structure. Figure 9 plots the structure
profile.

13
 Engineering with Computers

Table 9  Comparison results of D1 D2 D3 D4 D5 D6 V u Smax NS

the different approaches for the
engineering case DSBA 1.23 0.65 0.48 1.27 0.53 0.35 95,992 4.908 35.961 1840
EV 1.33 0.66 0.16 1.26 0.63 0.30 97,016 4.738 35.242 256
PoF 1.45 0.72 0.12 1.69 0.62 0.58 115,580 3.607 31.748 260
MAX-MSE 1.48 0.34 0.56 1.41 0.66 0.26 101,209 4.843 37.996 188
WAE 1.17 0.50 0.32 1.40 0.73 0.73 74,920 5.580 52.160 110
EAMGO 1.31 0.57 0.25 1.30 0.52 0.32 88,097 5.056 38.550 120
SCU-CI 1.29 0.62 0.44 1.19 0.69 0.14 96,328 4.892 35.675 104

Fig. 9  Structured profile of the conical shell

As illustrated in Fig. 9, the structure of the conical shell Fig. 10  Geometry model constructed by ANSYS
consists of an outer plate and the inner stiffened ribs. In this
work, the fixed parameters are the radius of the small end
of the conical shell R1 = 500 mm, the radius of the big end Table 10  Design variables and their ranges for the design optimiza-
of the conical shell R2 = 2000 mm, and the length of the tion of conical shell
conical shell is 600 mm. The number and the distribution of Design variable Ranges (mm)
the stiffeners are fixed. The geometry model constructed by
ANSYS 18.0 is shown in Fig. 10. The web height of the circumferential stiffeners x1 200–340
The design variables are the dimension of the longitu- The web thickness of the circumferential stiffeners x2 10–24
The panel width of the circumferential stiffeners x3 100–240
dinal and circumferential stiffeners which are T section,
The panel thickness of the circumferential stiffeners x4 10–24
the thickness of the shell plate for the former five rib space
The web height of the longitudinal stiffeners x5 100–240
t1, and the thickness of the remainder shell plate t2. In this
The web thickness of the longitudinal stiffeners x6 6–20
work, the optimizing problem is a constrained problem that
The panel width of the longitudinal stiffeners x7 40–180
minimizes the weight of the conical shell and the design
The panel thickness of the longitudinal stiffeners x8 6–20
constrains are the vibration performance of the conical shell
The thickness of the former shell plate x9 6–20
and the dimensions’ matches of the ribs. Thus, the optimiza-
The thickness of the remainder shell plate x10 6–20
tion problem of the shell can be specified as
[ ]
findx = x1 , x2 , , , x10
minimize w(x) the total weight of this conical shell, f is the frequency of
subject to g1 = 180Hz − f ≤ 0; g2 = az − 90 dB ≤ 0 (31) the first-order modal in the air, and az is the gross stage of
x x
g3 = 20x1 − 1 ≤ 0; g4 = 8x3 − 1 ≤ 0 the acceleration ranging from 100 Hz to 250 Hz at the small
g5 =
x5 2
− 1 ≤ 0; g6 =
x74
− 1 ≤ 0, end of the conical shell, which can be calculated as
20x6 8x8
[ k ( )]
∑ ai 2
az = 10 log fw , (32)
where x1–x10 are the design variables, whose physical mean- 1
a0
ing and boundaries are presented in Table 10. w(x) denotes

13
Engineering with Computers

Fig. 11  FEA model and simulation results for the design optimization of the conical shell

Table 11  Optimal design points X1 X2 X3 X4 X5 X6 X7 X8 X9 X10

of different approaches
DSBA 240 16 100 14 100 6 40 6 20 8
EV 240 20 120 16 120 6 60 12 20 14
PoF 320 18 120 20 140 12 40 6 18 12
MAX-MSE 280 16 100 14 100 16 40 14 18 10
WAE 240 12 100 14 100 6 40 8 20 10
EAMGO 240 12 100 14 140 8 40 6 20 6
SCU-CI 300 16 100 18 120 12 40 12 20 10

where k is the number of the calculated frequencies, ai is Table 12  Corresponding objective functions, constraints, and the
the acceleration response under the ith order calculated fre- number of simulation calls for different approaches
quency, a0 = 10−6m/s2 is the basis acceleration, and fw is the Y G1 G2 NS
interval of sweep frequency. In this case, the constraints g1
and g2 are obtained from the expensive simulation models, DSBA 8249 − 4 − 0.55 5050
while the objective function and the rest constraints can be EV 10,959 − 10 1.58 255
PoF 11,195 − 6 0.22 191
calculated by arithmetic expression.
MAX-MSE 9566 − 12 0.02 215
The frequency of the first-order modal and the gross
WAE 8191 − 16 1.79 209
stage of the acceleration are assumed to be expensive to
EAMGO 7481 6 − 0.29 392
be obtained. In this simulation, the elastic modulus is
SCU-CI 10,871 − 6 − 0.94 200
E = 2.11 × 1011 Pa , the Poisson ratio is 𝜇 = 0.3 , and the den-
sity of the material is 𝜌=7850 kg/m3. A downward unit har-
monic force is applied at the small end of the conical shell.
Element Beam 188 and Shell 181 are used to simulate the As observed in Table 12, only the DSBA approach and
geometry model of this conical shell structure. The mapped proposed SCU-CI approach can obtain feasible solutions
grid model with 6180 elements is used for this problem. for this engineering case. The other approaches failed to
The FEA model and simulation results are shown in Fig. 11. obtain feasible solutions. In terms of the computational cost,
The proposed SCU-CI approach is compared with five the number of total sample points of DSBA is 5050, which
existing approaches. 60 initial sample points are gener- is actually a computation-prohibitive process for complex
ated for constructing the initial Kriging metamodel. The engineering problems. The total sample points of SCU-CI
optimal solutions of different approaches are summarized approach are about 20 times less than that of DSBA, indicat-
in Table 11. The corresponding objective functions, con- ing that the proposed SCU-CI approach can obtain feasible
straints, and the number of simulation calls are listed in optimal solution while significantly reduce the computa-
Table 12. tional cost.

13
 Engineering with Computers

5 Conclusion 5. Wang H, Chen L, Li E (2017) Time dependent sheet metal form-

ing optimization by using Gaussian process assisted firefly algo-
rithm. Int J Mater Form pp. 1–17
In this work, general sequential constraints updating 6. Song X, Sun G, Li G, Gao W, Li Q (2012) Crashworthiness opti-
approach based on the confidence intervals from the Kriging mization of foam-filled tapered thin-walled structure using mul-
surrogate model, SCU-CI, is proposed, in which an objec- tiple surrogate models. Struct Multidiscip Optim 47(2):221–231
7. Zhou Q, Jiang P, Shao X, Hu J, Cao L, Wan L (2017) A variable
tive switching and sequential updating strategy is introduced fidelity information fusion method based on radial basis function.
to address two issues in the updating of Kriging surrogate Adv Eng Inform 32:26–39
model. One issue is whether the actual simulation models 8. Bellary SAI, Samad A, Couckuyt I, Dhaene T (2015) A compara-
or the Kriging metamodel should be used to evaluate the tive study of kriging variants for the optimization of a turboma-
chinery system. Eng Comput 32(1):49–59
feasibility of the design alternatives. The other one is which 9. Jiang P, Zhang Y, Zhou Q, Shao X, Hu J, Shu L (2018) An adap-
design alternatives would be selected to improve the predic- tive sampling strategy for Kriging metamodel based on Delaunay
tion accuracy of the Kriging metamodel. The core idea is triangulation and TOPSIS. Appl Intell 48(6):1644–1656
based on whether the feasibility status of the design alterna- 10. Bouhlel MA, Martins JRRA (2018) Gradient-enhanced kriging
for high-dimensional problems. Eng Comput 35(1):157–173
tives would be changed because of the interpolation uncer- 11. Dong H, Song B, Dong Z, Wang P (2018) SCGOSR: surrogate-
tainty from the Kriging surrogate model or not. based constrained global optimization using space reduction. Appl
To demonstrate the applicability and efficiency of the pro- Soft Comput 65:462–477
posed SCU-CI approach, nine numerical examples and two 12. Clarke SM, Griebsch JH, Simpson TW (2005) Analysis of sup-
port vector regression for approximation of complex engineering
practical engineering cases are tested. The observations are analyses. J Mech Des 127(6):1077
summarized as follows, (a) the proposed SCU-CI approach 13. Zhou Q, Shao XY, Jiang P, Gao ZM, Zhou H, Shu LS (2016) An
can generally obtain the optimum that meets the require- active learning variable-fidelity metamodelling approach based
ments of the actual feasibility, while it is not the case in other on ensemble of metamodels and objective-oriented sequential
sampling. J Eng Des 27(4–6):205–231
five existing approaches due to the prediction uncertainty 14. Jiang C, Cai X, Qiu H, Gao L, Li P (2018) A two-stage support
from the Kriging surrogate model and (b) in terms of the vector regression assisted sequential sampling approach for global
computational efficiency, the proposed SCU-CI approach metamodeling. Struct Multidiscip Optim 58(4):1657–1672
shows obvious advantages compared to DSBA and is more 15. Chatterjee T, Chakraborty S, Chowdhury R (2017) A critical
review of surrogate assisted robust design optimization. Arch
efficient than EV, PoF, MAX-MSE, WAE, and EAMGO Comput Methods Eng 26:245–725
approaches when all of them can obtain the feasible solu- 16. Zhou Q, Shao XY, Jiang P, Zhou H, Cao LC, Zhang L (2015) A
tions. It should be pointed out that the proposed SCU-CI deterministic robust optimisation method under interval uncer-
approach is general in the sense that it can be integrated tainty based on the reverse model. J Eng Des 26(10–12):416–444
17. Assari P, Dehghan M (2017) The numerical solution of two-
with any efficient global optimization approaches. As part dimensional logarithmic integral equations on normal domains
of future work, combining the proposed approach with dif- using radial basis functions with polynomial precision. Eng Com-
ferent sequential objective functions updating strategies for put 33(4):853–870
complex engineering optimization design problems will be 18. Zhou Q, Wang Y, Choi S-K, Jiang P, Shao X, Hu J, Shu L (2018)
A robust optimization approach based on multi-fidelity meta-
investigated. model. Struct Multidisciplin Optimization 57(2):775–797
19. Kaintura A, Spina D, Couckuyt I, Knockaert L, Bogaerts W,
Acknowledgements  This work has been supported by the National Dhaene T (2017) A Kriging and stochastic collocation ensemble
Natural Science Foundation of China (NSFC) under Grant Nos. for uncertainty quantification in engineering applications. Eng
51805179, 51775203, and the Research Funds of the Maritime Defense Comput 33:935–949
Technologies Innovation, and the Research Funds of the defense tech- 20. Han Z, Zimmerman R, Görtz S (2012) Alternative cokrig-
nologies leadership. ing method for variable-fidelity surrogate modeling. AIAA J
50(5):1205–1210
21. Huang C, Radi B, El Hami A, Bai H (2018) CMA evolution strat-
egy assisted by kriging model and approximate ranking. Appl
References Intell 48:4288–4304
22. Shao W, Deng H, Ma Y, Wei Z (2011) Extended Gaussian Krig-
1. Hu Z, Mahadevan S (2017) A surrogate modeling approach for ing for computer experiments in engineering design. Eng Comput
reliability analysis of a multidisciplinary system with spatio–tem- 28(2):161–178
poral output. Struct Multidiscip Optim 56(3):553–569 23. Toal DJJ (2015) A study into the potential of GPUs for the effi-
2. Jiang C, Qiu H, Yang Z, Chen L, Gao L, Li P (2019) A general cient construction and evaluation of Kriging models. Eng Comput
failure-pursuing sampling framework for surrogate-based reliabil- 32(3):377–404
ity analysis. Reliab Eng Syst Saf 183:47–59 24. Cheng J, Jiang P, Zhou Q, Jiexiang H, Tao Y, Leshi S, Xinyu S
3. Han Z-H, Zhang Y, Song C-X, Zhang K-S (2017) Weighted gra- (2019) A lower confidence bounding approach based on the coef-
dient-enhanced Kriging for high-dimensional surrogate modeling ficient of variation for expensive global design optimization. Eng
and design optimization. AIAA J 55(12):4330–4346 Comput 1:2. https​://doi.org/10.1108/EC-08-2018-0390
4. Hu J, Zhou Q, Jiang P, Shao X, Xie T (2018) An adaptive 25. Zheng J, Li Z, Gao L, Jiang G, Owen D (2016) A parameterized
sampling method for variable-fidelity surrogate models using lower confidence bounding scheme for adaptive metamodel-based
improved hierarchical Kriging. Eng Optim 50(1):145–163 design optimization. Eng Comput 33(7):2165–2184

13
Engineering with Computers

26. Chen S, Jiang Z, Yang S, Chen W (2016) Multimodel fusion based 43. Bichon BJ, Eldred MS, Swiler LP, Mahadevan S, McFarland JM
sequential optimization. AIAA J 55(1):241–254 (2008) Efficient global reliability analysis for nonlinear implicit
27. Regis RG, Shoemaker CA (2007) A stochastic radial basis func- performance functions. AIAA J 46(10):2459–2468
tion method for the global optimization of expensive functions. 44. Li X, Qiu H, Chen Z, Gao L, Shao X (2016) A local Kriging
Informs J Comput 19(4):497–509 approximation method using MPP for reliability-based design
28. Hennig P, Schuler CJ (2012) Entropy search for information- optimization. Comput Struct 162:102–115
efficient global optimization. J Mach Learn Res 13:1809–1837 45. Shu L, Jiang P, Wan L, Zhou Q, Shao X, Zhang Y (2017) Met-
29. Krause A, Ong CS (2011) Contextual gaussian process bandit amodel-based design optimization employing a novel sequential
optimization. Adv Neural Inform Process Syst sampling strategy. Eng Comput 34(8):2547–2564
30. Viana FA, Haftka RT, Watson LT (2013) Efficient global optimi- 46. Liu H, Xu S, Chen X, Wang X, Ma Q (2016) Constrained global
zation algorithm assisted by multiple surrogate techniques. J Glob optimization via a DIRECT-type constraint-handling technique
Optim 56(2):669–689 and an adaptive metamodeling strategy. Struct Multidiscip Optim
31. Jones DR, Schonlau M, Welch WJ (1998) Efficient global 55:155–177
optimization of expensive black-box functions. J Glob Optim 47. Dong H, Song B, Dong Z, Wang P (2016) Multi-start space reduc-
13(4):455–492 tion (MSSR) surrogate-based global optimization method. Struct
32. Sóbester A, Leary SJ, Keane AJ (2005) On the design of optimi- Multidiscip Optim 54(4):907–926
zation strategies based on global response surface approximation 48. Shi R, Liu L, Long T, Wu Y, Tang Y (2019) Filter-based adaptive
models. J Glob Optim 33(1):31–59 Kriging method for black-box optimization problems with expen-
33. Wang H, Li E, Li GY (2009) The least square support vector sive objective and constraints. Comput Methods Appl Mech Eng
regression coupled with parallel sampling scheme metamodeling 347:782–805
technique and application in sheet forming optimization. Mater 49. Wu Y, Yin Q, Jie H, Wang B, Zhao J (2018) A RBF-based con-
Des 30(5):1468–1479 strained global optimization algorithm for problems with compu-
34. Zhan D, Qian J, Cheng Y (2017) Balancing global and local search tationally expensive objective and constraints. Struct Multidiscip
in parallel efficient global optimization algorithms. J Glob Optim Optim pp. 1–23
67(4):873–892 50. Sacks J, Welch WJ, Mitchell TJ, Wynn HP (1989) Design and
35. Dong H, Song B, Wang P, Dong Z (2018) Hybrid surrogate-based analysis of computer experiments. Stat Sci 4:409–423
optimization using space reduction (HSOSR) for expensive black- 51. Zhou Q, Wang Y, Choi S-K, Jiang P, Shao X, Hu J (2017) A
box functions. Appl Soft Comput 64:641–655 sequential multi-fidelity metamodeling approach for data regres-
36. Haftka RT, Villanueva D, Chaudhuri A (2016) Parallel surrogate- sion. Knowl-Based Syst 134:199–212
assisted global optimization with expensive functions—a survey. 52. Coello CAC (2000) Use of a self-adaptive penalty approach for
Struct Multidiscip Optim 54(1):3–13 engineering optimization problems. Comput Ind 41(2):113–127
37. Schonlau M (1997) Computer experiments and global
53. Zhu J, Wang Y-J, Collette M (2013) A multi-objective variable-
optimization fidelity optimization method for genetic algorithms. Eng Optim
38. Li Y, Wu Y, Zhao J, Chen L (2017) A Kriging-based constrained 46(4):521–542
global optimization algorithm for expensive black-box functions 54. Zhou H, Zhou Q, Liu C, Zhou T (2018) A kriging metamodel-
with infeasible initial points. J Glob Optim 67(1–2):343–366 assisted robust optimization method based on a reverse model.
39. Wang Z, Ierapetritou M (2018) Constrained optimization of black- Eng Optim 50(2):253–272
box stochastic systems using a novel feasibility enhanced Kriging- 55. Wang Z, Ierapetritou M (2018) Constrained optimization of black-
based method. Comput Chem Eng 118:210–230 box stochastic systems using a novel feasibility enhanced Kriging-
40. Zhang Y, Han ZH, Zhang KS (2018) Variable-fidelity expected based method. Comput Chem Eng 118:210–223
improvement method for efficient global optimization of expen- 56. Garcia S, Herrera F (2008) An extension on “statistical compari-
sive functions. Struct Multidiscip Optim 58:1431–1451 sons of classifiers over multiple data sets” for all pairwise com-
41. Parr JM, Keane AJ, Forrester AIJ, Holden CME (2012) Infill parisons. J Mach Learn Res 9:2677–2694
sampling criteria for surrogate-based optimization with constraint
handling. Eng Optim 44(10):1147–1166 Publisher’s Note Springer Nature remains neutral with regard to
42. Sasena MJ, Papalambros P, Goovaerts P (2002) Exploration of jurisdictional claims in published maps and institutional affiliations.
metamodeling sampling criteria for constrained global optimiza-
tion. Eng optim 34(3):263–278

13

View publication stats