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

net/publication/221927230

Surrogate-Based Optimization

Chapter · March 2012

DOI: 10.5772/36125 · Source: InTech

CITATIONS READS
122 6,875

2 authors:

Zhong-Hua Han ke-shi Zhang

Northwestern Polytechnical University Northwestern Polytechnical University
161 PUBLICATIONS 3,738 CITATIONS 34 PUBLICATIONS 983 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Zhong-Hua Han on 03 May 2016.

The user has requested enhancement of the downloaded file.

 17

Surrogate-Based Optimization
Zhong-Hua Han and Ke-Shi Zhang
School of Aeronautics, Northwestern Polytechnical University, Xi’an,
P.R. China

1. Introduction
Surrogate-based optimization (Queipo et al. 2005, Simpson et al. 2008) represents a class of
optimization methodologies that make use of surrogate modeling techniques to quickly find
the local or global optima. It provides us a novel optimization framework in which the
conventional optimization algorithms, e.g. gradient-based or evolutionary algorithms are
used for sub-optimization(s). Surrogate modeling techniques are of particular interest for
engineering design when high-fidelity, thus expensive analysis codes (e.g. Computation
Fluid Dynamics (CFD) or Computational Structural Dynamics (CSD)) are used. They can be
used to greatly improve the design efficiency and be very helpful in finding global optima,
filtering numerical noise, realizing parallel design optimization and integrating simulation
codes of different disciplines into a process chain. Here the term “surrogate model” has the
same meaning as “response surface model”, “metamodel”, “approximation model”,
“emulator” etc. This chapter aims to give an overview of existing surrogate modeling
techniques and issues about how to use them for optimization.

2. Overview of surrogate modeling techniques

For optimization problems, surrogate models can be regarded as approximation models for
the cost function (s) and state function (s), which are built from sampled data obtained by
randomly probing the design space (called sampling via Design of Experiment (DoE)). Once
the surrogate models are built, an optimization algorithm such as Genetic Algorithms (GA)
can be used to search the new design (based on the surrogate models) that is most likely to
be the optimum. Since the prediction with a surrogate model is generally much more
efficient than that with a numerical analysis code, the computational cost associated with
the search based on the surrogate models is generally negligible.
Surrogate modeling is referred to as a technique that makes use of the sampled data
(observed by running the computer code) to build surrogate models, which are sufficient to
predict the output of an expensive computer code at untried points in the design space.
Thus, how to choose sample points, how to build surrogate models, and how to evaluate the
accuracy of surrogate models are key issues for surrogate modeling.

2.1 Design of experiments

To build a surrogate model, DoE methods are usually used to determine the locations of
sample points in the design space. DoE is a procedure with the general goal of maximizing

www.intechopen.com
344 Real-World Applications of Genetic Algorithms

the amount of information gained form a limited number of sample points (Giunta et al.,
2001). Currently, there are different DoE methods which can be classified into two
categories: “classic” DoE methods and “modern” DoE methods. The classic DoE methods,
such as full-factorial design, central composite design (CCD), Box-Behnken and D-Optimal
Design (DOD), were developed for the arrangement of laboratory experiments, with the
consideration of reducing the effect of random error. In contrast, the modern DoE methods
such as Latin Hypercube Sampling (LHS), Orthogonal Array Design (OAD) and Uniform
Design (UD) (Fang et al., 2000) were developed for deterministic computer experiments
without the random error as arises in laboratory experiments. An overview of the classic
and modern DoE methods was presented by Giunta et al. (2001). A more detailed
description of existing DoE methods is beyond the scope of this chapter.
The schematics of 40 sample points selected by LHS and UD for a two-dimensional problem
are sketched in Figure 1.

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
V2

0.5
V2

0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
V1 V1

Fig. 1. Schematics of 40 sample points selected by Design of Experiments for a two-

dimensional problem (left: Latin Hypercube Sampling; right: Uniform Design).

2.2 Surrogate models

There are a number of surrogate models available in the literatures. Here we limit our
discussion to three popular techniques such as RSM (polynomial Response Surface Model),
Kriging, RBFs (Radial Basis Functions).

For an m-dimensional problem, suppose we are concerned with the prediction of the output
of a high-fidelity, thus expensive computer code, which is correspondent to an unknown
function y : m → . By running the computer code, y is observed at n sites (determined by
DoE)

S = [ x(1) ,..., x( n ) ]T ∈  n×m , x = { x1 ,.., xm } ∈  m (1)

with the corresponding responses

y S = [ y (1) ,..., y ( n ) ]T = [ y( x(1) ),..., y( x( n ) )]T ∈  n (2)

www.intechopen.com
Surrogate-Based Optimization 345

The pair ( S , y S ) denotes the sampled data sets in the vector space.

With the above descriptions and assumptions, our objective here is to build a surrogate
model for predicting the output of the computer code for any untried site x (that is, to
estimate y( x ) ) based on the sampled date sets ( S , y S ), in an attempt to achieve the desired
accuracy with the least possible number of sample points.

2.2.1 Quadratic response surface method

Here we use “RSM” to denote a polynomial approximation model in which the sampled
data is fitted by a least-square regression technique. In RSM-based optimization
applications, the “quadratic” polynomial model usually provides the best compromise
between the modeling accuracy and computational expense, when compared with the linear
or higher order polynomial models. An advantage of RSM is that it can smooth out the
various scales of numerical noise in the data while captures the global trend of the variation,
which makes it very robust and thus well suited for optimization problems in engineering
design.
The true quadratic RSM can be written in the following form:

y ( x ) = yˆ ( x ) + ε , x ∈  m , (3)

where ŷ ( x ) is the quadratic polynomial approximation and ε is the random error which is
assumed to be normally distributed with mean zero and variance of σ 2 . The error ε i at each
observation is supposed to be independent and identically distributed. The quadratic RSM
predictor ŷ ( x ) can be defined as:

yˆ ( x ) = β 0 +  β i xi +  β ii xi2 +  β ij xi x j ,
m m m m
(4)
i =1 i =1 i =1 j ≥i

where β 0 , β i , β ii and β ij are unknown coefficients to be determined. Since there are

totally p = (m + 1)(m + 2) / 2 unknown coefficients in Eq.(4), building a quadratic RSM with
m variables requires at least p sample points. Let β ∈  p be the column vector contains
these p unknown coefficients. The least square estimator of β is

β = ( U T U )−1 U T y S , (5)

where

 
 1 x1  x m 
(1) 2 (1) 2
(1) (1)
x1(1)x2(1)  (1) (1)
xm − 1 xm (x )  (x )
U =    ∈  n× p .
1 m

 
        (6)
 (n) 2 
1 x1 xm 

2
(n) (n)
 xm x1( n )x(2n )  (n) (n)
xm − 1 xm ( )
x2( n )  ( )
After the unknown coefficients in β are determined, the approximated response ŷ at any
untried x can be efficiently predicted by Eq. (4).

www.intechopen.com
346 Real-World Applications of Genetic Algorithms

2.2.2 Kriging model

Different from RSM, Kriging (Krige, 1951) is an interpolating method which features the
observed data at all sample points. Kriging provides a statistic prediction of an unknown
function by minimizing its Mean Squared Error (MSE). It can be equivalent to any order of
polynomials and is thus well suited for a highly-nonlinear function with multi extremes. For the
derivation of Kriging (Sacks et al., 1989), the output of a deterministic computer experiment is
treated as a realization of a random function (or stochastic process), which is defined as the sum
of a global trend function f T ( x)β and a Gaussian random function Z ( x) as following

y( x ) = f T ( x )β + Z( x ), x ∈  m , (7)

where f( x ) = [ f 0 ( x ),.., f p − 1 ( x )]T ∈  p is defined with a set of the regression basis functions
and β = [ β 0 ,.., β p − 1 ]T ∈  p denotes the vector of the corresponding coefficients. In general,
f T ( x)β is taken as either a constant or low-order polynomials. Practice suggests that the
constant trend function is sufficient for most of the problems. Thus, f T ( x)β is taken as a
constant β 0 in the text hereafter. In Eq.(7), Z (⋅) denotes a stationary random process with
zero mean, variance σ 2 and nonzero covariance of

Cov[ Z( x ), Z( x′)] = σ 2 R( x , x′) . (8)

Here R (x, x′) is the correlation function which is only dependent on the Euclidean distance
between any two sites x and x′ in the design space. In this study, a Gaussian exponential
correlation function is adopted, and it is of the form

R( x , x′) = exp[ −  θ k |x k − x′k |pk ] ,1 < pk ≤ 2 ,

m
(9)
k =1

where θ = [θ1 ,θ 2 ,...,θm ]T and p = [ p1 , p2 ,..., pm ]T denote the vectors of the unknown model
parameters (hyper parameters) to be tuned. The schematics of a Gaussian exponential
correlation function for one-dimensional problem is sketched in Figure 2.
1.2 1.2
p = 2.0 θ = 1.0
1 1
Correlation function

Correlation function

0.8 0.8
θ = 0.1

0.6 0.6

θ = 1.0
0.4 0.4

0.2 p = 0.5
0.2
p = 1.0
θ = 10.0 p = 2.0
0 0

-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4
xi - x xi - x

Fig. 2. Schematics of Gaussian exponential correlation function for different hyper

parameters (left: varying θ with a fixed p ; right: varying p with a fixed θ )

www.intechopen.com
Surrogate-Based Optimization 347

From the derivation by Sacks et al. (1989) the Kriging predictor ŷ ( x ) for any untried x can
be written as

yˆ ( x ) = β 0 + r T ( x )R −1 ( y S − β0 1) , (10)

where the generalized least square estimation of β 0 is

β 0 = ( 1T R −1 1)−1 1T R −1 y S , (11)

and 1 ∈  n is a vector filled with ones, and R , r are the correlation matrix and the
correlation vector, respectively. R and r are defined as

 R( x(1) , x(1) ) R( x(1) , x(2) )  R( x(1) , x( n ) )   R( x(1) , x ) 

   
 R( x(2) , x(1) ) R( x(2) , x(2) )  R( x(2) , x( n ) )  R( x(2) , x )
R= ∈ , r = 
n×n
∈ ,
n

   
(12)
    
 R( x( n ) , x(1) ) R( x( n ) , x(2) )  R( x( n ) , x( n ) )  R( x( n ) , x )
  

where R( x( i ) , x( j ) ) denotes the correlation between any two observed points x( i ) and x( j ) ;
R( x( i ) , x ) denotes the correlation between the i-th observed point x( i ) and the untried point
x.
A unique feature of Kriging model is that it provides an uncertainty estimation (or MSE) for
the prediction, which is very useful for sample-points refinement. It is of the form

sˆ 2 ( x ) = σ 2 [1.0 − r TR −1 r + ( r T R −1 1 − 1)2 / 1T R −1 1] . (13)

Assuming that the sampled data are distributed according to a Gaussian process, the
responses at sampling sites are considered to be correlated random functions with the
corresponding likelihood function given by

 1 ( y S − β 0 1)T R −1 ( y S − β 0 1) 
exp  −  .
1
L( β 0 ,σ 2 , θ , p) =
 2 
(14)
2π (σ 2 )n R σ2

The optimal estimation of β 0 and the process variance

β 0 (θ , p) = ( 1T R −1 1)−1 1T R −1 y S
1 (15)
σ 2 ( β 0 , θ , p) = ( y S − β 0 1)T R −1 ( y S − β 0 1)
n

are obtained analytically, yet depends on the remaining hyper-parameters θ = [θ 1 ,θ 2 ,...,θ m ]T

and p = [ p1 , p2 ,..., pm ]T . Substituting it into the associated Eq. (14) and taking the logarithm,
we are left with maximizing

. MLE(θ , p) = −n ln σ 2 ( θ) − ln R(θ) ., (16)

www.intechopen.com
348 Real-World Applications of Genetic Algorithms

which can be solved by a numerical optimization algorithm such as GA. The

hyperparamters tuning strategies are discussed by Toal et al. (2008). Note that the above
Kriging formulation can be extended by including gradient information obtained by Adjoint
method (Han et al. 2009, Laurenceau et al. 2008) or lower-fidelity data by lower-fidelity
analysis code (Han et al. 2010, Forrester et al. 2007).

2.2.3 Radial basis functions

In additional to Kriging, RBFs model (Hardy, 1971) is known as an alternative interpolation
method for surrogate modeling. For the RBFs approach by Powell (1987), the approximation
of the unknown function y( x ) at an untried x is formally defined as the linear combination
of the radial basis functions and a global trend function as

yˆ ( x ) =  ωiϕ ( x ) + P( x ) ,
n
(17)
i =1

where ωi are the i-th unknown weight coefficient, ϕ ( x ) = ϕ ( x( i ) − x ) are the basis functions
that depend on the Euclidean distance between the observed point x( i ) and the untried
point x (similar to the correlation function of kriging model); P( x ) is the global trend
function which is taken as a constant β 0 here. To ensure the function values at observed
points are reproduced by the RBFs predictor, the flowing constraints should be satisfied:

yˆ ( x( i ) ) = y ( i ) , i = 1,.., n
. (18)
Then the additional constraints for P (x) should be imposed as

 ωi = 0 .
n
(19)
i =0

Solving the linear equations formed by Eq. (18) and Eq. (19) for ωi and β 0 , and substituting
into Eq.(17) yields the RBFs predictor as

yˆ ( x ) = β 0 + φT ( x )Ψ −1 ( y S − β 0 1) . (20)

Where β 0 (θ) = ( 1T Ψ −1 1)−1 1T Ψ −1 y S and Ψ , φ( x ) are defined as

Ψ := [ϕ ( x( i ) − x( j ) )]ij ∈  n×n , φ( x ) := [ϕ ( x( i ) − x )]i ∈  n . (21)

When the above RBFs predictor is compared with the Kriging predictor (see Eq. (10)), one
can observe that they are essentially similar, only with the basis-function matrix Ψ (also
called Gram matrix) and the basis function vector φ(x) being different from the correlation
matrix R and the correlation vector r (x) of the Kriging predictor, respectively. In addition,
RBFs differs from Kriging at the following two aspects: 1) RBFs doesn’t provide the
uncertainty estimation of the prediction; 2) The model parameters can’t be tuned by MLE
like Kriging. Generally, Kriging can be regarded as a particular form of RBFs.

www.intechopen.com
Surrogate-Based Optimization 349

To build a RBFs model, one needs to prescribe the type of basis functions that only depends
on the Euclidean distance r = x − x′ between any two sites x and x′ . Compared to the
correlation function used for a Kriging model, more choices are available for a RBFs model,
which are partially listed in Table 1.

Basis functions Formulations

2
/ 2σ 2
Gaussian (GAUSS) φ (r ) = e − r 2
(e.g. σ = 1 )

Power function (POW) φ (r ) = r β ,1 ≤ β ≤ 3 (e.g. β = 1.8 )

2
Thin Plate Spline (TPS) φ (r ) = r ln(r )
Hardy’s Multiquadric (HMQ) φ (r ) = 1 + r 2
Hardy’s Inverse Multiquadric (HIMQ) φ (r ) = 1/ 1 + r 2
Table 1. Basis functions for RBFs surrogate model

All the basis functions listed in Table 1 can be classified into two categories: decaying
functions (such as GAUSS and HIMQ) and growing functions (POW, TPS and HMQ). The
decaying functions can yield positive definite matrix Ψ , which allows for the use of
Cholesky decomposition for its inversion; the growing functions generally result in a non-
positive definite matrix Ψ and thus LU decomposition is usually used alternatively. The
schematics of the basis functions for one-dimensional problem is sketched in Figure 3.
8
1

0.8 6
Basis functions
Basis functions

0.6
4
IHMQ
0.4

2
HMQ
0.2 GAUSS
TPS
POW

0
0
-4 -3 -2 -1 0 1 2 3 4
-4 -3 -2 -1 0 1 2 3 4
i xi - x
x -x

Fig. 3. Schematics of basis functions for Radial Basis Functions (left: decaying functions;
right: growing functions).

2.2.4 The big (ger) picture

In addition to what we mentioned above, there are also a few surrogate models available in
the literatures, such as Artificial Neutral Network (ANN) (Elanayar et al. 1994, Park et al.
1991), Multiple Adaptive Regression Splines (MARS) and Support Vector Regression (SVR)

www.intechopen.com
350 Real-World Applications of Genetic Algorithms

(Smola & Schoelkopf 2004). Although these methods are coming from different research
communities, the idea is similar when using them for function prediction in surrogate
modeling. They are not described in detail here due to the limited space. The readers are
referred to read the paper by Wang & Shan (2007) and the books written by Keane et al. (2005)
and by Forrester et al. (2008) for more description about surrogate modeling techniques.

2.3 Evaluation of approximation models

An important issue for the surrogate modeling is how to estimate the error of the
approximation models. Only when the surrogate model with sufficient accuracy is built can
the reliable optimum be obtained. Here we use two variables ( e and σ e ) to evaluate the
error of the approximation models at test points, which are also chosen by DoE method. The
average relative error is

 e( i ) ,
nt
1 yˆ t( i ) − yt( i )
e= e( i ) = , (22)
nt i =1 yt( i )

where nt is number of the test points; yt( i ) and yˆ t( i ) are the true value and predicted value
corresponding to the i-th test point, respectively. The root mean squared error is defined by

 (e
nt
(i ) 2
σe = ) nt . (23)
i =1

2.4 Framework of building surrogate models

A Generic framework of building a surrogate model is sketched in Figure 4. Note that the
initial surrogate model can be evaluated by Eqs. (22) and (23) and a branch for resampling is
denoted by black dashed line in Figure 4.

3. Use of surrogate models for optimization

Suppose we are concerned with solving a general optimization problem as

Objective minimize y(x )

s.t. g i (x ) ≤ 0, i = 1,...nc , (24)
xl ≤ x ≤ xu

where nc is the number of state functions which is in line with the number of inequality
constraints (assuming that all the equality constraints have been transformed into inequality
constraints.); x l and x u are the lower and upper bound of design variables, respectively;
the object function y(x ) and state functions g i (x ) are evaluated by an expensive analysis
code. Traditionally, the optimization problem is solved by either a gradient-based algorithm
or a gradient-free algorithm such as GA. It may become prohibitive due to the large
computational cost associated running the expensive analysis code. Alternatively, here we
are concerned with using surrogate modeling techniques to solve the optimization problem,
in an attempt to dramatically improve the efficiency.

www.intechopen.com
Surrogate-Based Optimization 351

Design space

DoE

Run analysis code

Sampled database

Build surrogate models

Evaluate surrogate models

Surrogate models

Fig. 4. Frameworks of building surrogate models

3.1 Framework of surrogate-based optimization

3.1.1 A simple framework
The basic idea of using surrogate models in optimization can be quite simple. First, the
surrogate models for the object function(s) and state function(s) with sufficient accuracy are
built (see Figure 2); second, the optimum is found by an optimizer, with the object
function(s) or state function(s) evaluated by surrogate models, rather than by the expensive
analysis code. Since prediction with the surrogate models is much more efficient than that
by the expensive analysis code, the optimization efficiency can be greatly improved. The
comparison of the conventional optimization and surrogate-based optimization is sketched
in Figure 5. In addition to improve optimization efficiency, surrogate models also serve as
an interface between the analysis code and the optimizer, which makes the establishment of
an optimization procedure much easier. One of the examples for such a surrogate-based
optimization framework can be found in paper by Han et al. (2010).

3.1.2 A bi-level framework

Although the framework of the surrogate-based optimization sketched in Figure 5. (b) is
very intuitive and simple, questions may arise: are the surrogate models accurate enough?
has the true optimum been reached? In fact, the optimum gained by the surrogate models is
only an approximation to the true optimum (see Figure 5. (a)). One has to refine the
surrogate models by adding new sample points, which is to be observed by running the
analysis code. The procedure of augmenting new sample point(s) to the current sampled
data sets is the so-called “sample-point refinement”. The rules of determining the new
sample sites towards the true optimum are called “infill criteria”, which will be discussed in
section 3.2. The flowchart of a surrogate-based optimization with additional process of

www.intechopen.com
352 Real-World Applications of Genetic Algorithms

sample-point refinement is sketched in Figure 6. It can be regarded as a bi-level optimization

framework, with the process of building and refining the surrogate models (which needs to
run the expensive analysis code) acting as the main optimization and the process of using
surrogate models to determine the new sample sites acting as the sub-optimization(s).

Baseline Analysis code Optimizer Converge? Optimum

New design(s)

(a) Conventional optimization

Baseline Surrogate models Optimizer Converge? Approxi.

Optimum

New design(s)

(b) Surrogate-based optimization

Fig. 5. Comparison of frameworks for conventional optimization and surrogate-based

optimization with a simple framework

design space

DoE

distributed
simulate sample
computing

sampled data set

surrogate models

Optimizer
Main new e.g. GA sub-
optimization samples optimizations
no
Converge ?
yes

no
Converge ?

yes

Optimum design

Fig. 6. Flowchart of the surrogate-based optimization with a bi-level framework (main

optimization: building and refining the surrogate models which needs to run the expensive
analysis code; sub-optimization(s): using surrogate models to determine new sample sites)

www.intechopen.com
Surrogate-Based Optimization 353

3.2 Infill criteria

The infill criterion is related to the determination of the new sample sites by solving sub-
optimization problem(s). Three infill criteria are discussed here: Searching Surrogate
Model(s) (SSM), Expected Improvement (EI) and Lower-Bounding Confidence (LCB).

3.2.1 Searching surrogate model(s)

Provided that the initial surrogate models have been built, an optimizer such as GA can be
used to find the optimum, which in turn can be employed to refine the surrogate models.
The mathematical model of the sub-optimization for determining the new sample site is of
the form

Objective ˆ x)
minimize y(
s.t. gˆ i (x ) ≤ 0, i = 1, nc , (25)
xl ≤ x ≤ xu

where ŷ(x ) and ĝ i (x ) are surrogate models of y(x ) and g i (x ) , respectively. With the
optimal design variables x̂ opt gained by the surrogate models in hand, one needs to run the
expensive analysis code to compute the corresponding true function value and compare it
with what predicted by the surrogate models. If the error between them is blow a threshold,
the optimization process can be terminated; if not, the new sample point is augmented to the
sampled data sets and the surrogate models are rebuilt; the process is repeated until the
optimum solution is approached.
This criterion applies for all the surrogate models and is very efficient for local exploitation
of the promising region in the design space.

3.2.2 Expected improvement

Surrogate model such as Kriging provides not only an function prediction but also an
estimation of the mean squared error (MSE). In fact, the prediction by a Kriging model,
yˆ ( x ) , at any point can be regarded as a Gaussian random variable with the mean given by
the Kriging predictor, and the variance given by the mean squared error, s 2 ( x ) (see section
2.2.2). Viewed in this way, a probability can be computed that the function value at any
untried x would fall below the minimum among the sample points observed so far. Then
Expected Improvement (EI) function (Jones et al 1998, Jeong et al. 2005) can be calculated to
account for the improvement of the object function we expect to achieve at any untried x .
The definition of EI is of the form

  y min − yˆ ( x )   ymin − yˆ ( x ) 
( ymin − yˆ ( x ))Φ   +sˆ( x )φ   if sˆ > 0
E[ I ( x )] =     
 0
sˆ( x ) sˆ( x ) (26)
 if sˆ = 0

where Φ() and φ ( ) are the cumulative distribution function and probability density
function of a standard normal distribution, respectively. y min = Min( y (1) , y (2) ,..., y ( n ) )

www.intechopen.com
354 Real-World Applications of Genetic Algorithms

denotes the minimum of the observed data so far. The greater the EI, the more improvement
we expect to achieve. The point with maximum EI is located by a global optimizer such as
GA then observed by running the analysis code. For this infill criterion, the constraints can
be accounted by introducing the probability that the constraints are satisfied. The
corresponding sub-optimization problem can be modeled as

nc
Objective maximize E[ I ( x )] ⋅ ∏ P[Gi ( x ) ≤ 0]
i =1
, (27)
s.t. xl ≤ x ≤ xu

where P[Gi ( x ) ≤ 0] denotes the probability that i-th constraint may be satisfied and Gi ( x ) is
a random function corresponding to i-th state function gi ( x ) . P[Gi ( x ) ≤ 0] → 1 when the
constraint is satisfied and P[Gi ( x ) ≤ 0] → 0 when the constraint is violated. P[Gi ( x ) ≤ 0] can
be calculated by

 − gˆ ( x ) 
−∞ e dGi ( x ) = Φ  i 
1 0 −[ Gi ( x ) − gˆ i ( x )]2 /2 sˆi2 ( x )

 sˆi ( x ) 
P[Gi ( x ) ≤ 0] = (28)
ˆsi ( x ) 2π

where sˆi ( x ) denotes the estimated standard error corresponding to the surrogate model
gˆ i ( x ) .
The optimum site x̂ opt obtained by solving Eq. (27) is observed by running analysis code
and the new sample point is added to the sampled date sets; the surrogate models are
rebuilt and the whole process is repeated until the global optimum is approached.

3.2.3 Lower-bounding confidence (LCB)

The LCB function is defined as the weighted sum of predicted function value yˆ (x ) and the
standard error of the prediction sˆ(x ) . For an optimization problem of finding the minimum
of the unknown function y( x ) , a simple expression for LCB function is of the form

LCB= yˆ (x ) − Asˆ(x ) , (29)

where A is a constant which balances the influence of the predicted function and the
corresponding uncertainty. Best practice suggests A = 1 works well for a number of realistic
problems. The corresponding sub-optimization problem can be modeled as

Objective minimize yˆ (x ) − Asˆ(x )

s.t. gˆ i (x ) ≤ 0, i = 1, nc , (30)
xl ≤ x ≤ xu

The above optimization problem can be solved via a global optimizer such as GA. Since the
point with smallest value of LCB indicates the possible minimum of the unknown function,
the optimum site x̂ opt is then observed and added to sampled data sets to refine the
surrogate models. This procedure is performed iteratively until the global optimum is
reached.

www.intechopen.com
Surrogate-Based Optimization 355

4. Examples for applications to aircraft design

4.1 Airfoil design
Using an in-house Navier-Stokes flow solver, the objective of the problem is to minimize the
o
drag of an RAE2822 airfoil at the flow condition of Ma = 0.73 , α =2.7 , Re = 6.5 × 10 6 , subject to
3 constraints:

Objective : Minimize Cd
st. : (1) Area ≥ 0.99 × Area0
, (31)
: (2) C l ≥ 0.99 × C l 0
: (3) C m ≤ C m 0

where Area0 ,Cl0 , Cm0 are the area, lift coefficient, and moment coefficient of the baseline
airfoil, respectively. The first constraint is in consideration of the structural design of the
wing to guarantee the volume of the wing; the second one is to enforce a constant lift of the
wing in order to balance the weight of the aircraft at cruise condition; the third one is to
control the pitching moment of the airfoil to avoid large drag penalty of the horizontal tail
paid for balancing the aircraft.
The initial number of samples for Kriging is set to 20, selected by the Latin Hypercube
Sampling (LHS). The airfoil is parameterized by 10 Hicks-Henne bump functions (Hicks &
Henne, 1978); and the maximum amplitude of each bump is Amax / c = 0.544% . Both of the
SSM and EI infill strategies are adopted in the surrogate refinement. Table 2 presents the
optimization results of the two optimization method. The optimized and initial airfoils and
the corresponding pressure coefficient distributions are compared in Figure 7. Note that the
aerodynamic coefficients of the initial airfoil RAE2822 are set to 100. Obviously, the Kriging-
based optimization method gives better result, and with higher efficiency, and is more likely
to find the global optimum.

initial initial
RSM RSM
Kriging Kriging
-1

0.05

-0.5
y/c
Cp

0 0

0.5
o
Ma = 0.73, Re=6.5e6, α=2.7
-0.05
1

0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 1

x/c x/c

Fig. 7. Aerodynamic shape optimization of a transonic airfoil (RAE 2822) via Kriging and
quadratic Response Surface Model (left: pressure distribution; right: airfoil shape); by using
Kriging model with Expected Improvement infill criteria, the drag is reduced by 33.6% with
only 56 calling of Navier-Stokes flow solver.

www.intechopen.com
356 Real-World Applications of Genetic Algorithms

Cd Cl Cm Area NS flow solver calls

baseline 100 100 100 100 -

RSM 73.0 101.1 99.3 99.9 102

(-27.0%) (+1.1%) (-0.7%) (-0.1%)
Kriging 70.7 101.1 96.5 100.4 57
(-29.3%) (+1.1%) (-3.5%) (+0.4%)

Table 2. Drag reduction of an RAE 2822 airfoil via Kriging and RSM-based optimizations

4.2 Wing design

Here we are concerned with the preliminary design for a high-subsonic transport-aircraft
wing of a wing/body combination, considering aerodynamic, structure and static
aeroelastic effect. The calculation of the external flow is carried out by numerical solutions of
the full potential equation in conservative form (Kovalev & Karas, 1991). The FEM-based
commercial software ANSYS is used for analyzing the structural performance of the wing
with double-beams sandwich structure. Weak coupling method is adopted for static
aeroelastic analysis.
The optimization objectives are to maximize the aircraft lift-to-drag ratio and minimize the
weight of wing for a fixed maximum take-off weight of 54 tons and cruise Mach number of
0.76 at 10,000 meters high. The wing is composed of inner and outer wing. The reference
area of wing is 105 square meter. The mathematical model for optimization is of the form

max L D
min Wwing
s.t. L ≥ 54 / 10 3 kg
(32)
100 ≤ Swing ≤ 110 / m 2
σ max ≤ σ b / 109 pa
δ max ≤ 1 /m

Eight supercritical airfoils are configured along the span. The optimization is subject to 4
constraints. The first constraint is to enforce a constant lift of the wing in order to balance
the weight of the aircraft at cruise condition; the second one is to guarantee a near constant
wing loading; the third and fourth constraints are to make sure that the strength and rigidity
requirements are satisfied. The definition for the limits of design variables is listed in Table
3. The first four design variables define the aerodynamic configuration of the wing. The four
remain are for structure design. The detail can be found in paper by Zhang et al. (2008).
The uniform design table, U100(108), is used to creates one hundred candidate wings for
building surrogate models. The other forty-five candidate wings are created by the uniform
design table, U45(108), for evaluating the approximation model. For each wing, the static
aeroelastic analysis is performed to obtain the responses of lift (L), lift-to-drag ratio (L/D),
wing area (Swing), maximum stress (σmax), maximum deformation (δmax) and wing weight

www.intechopen.com
Surrogate-Based Optimization 357

(Wwing). Then the average relative errors and the root mean squared errors are calculated to
evaluate the approximation models, as listed in Table 4. In this case Kriging and RSM have
comparative high accuracy.

Design variable Unit Lower limit Upper limit

Span m 26 34
Taper ratio 0.2 0.4
Linear twist angle degree -3 -1
Sweep angle on leading edge degree 25 35
Thickness of front-spar web mm 2 6
Thickness of back-spar web mm 2 6
Thickness of lower skin mm 3 7
Thickness of Upper skin mm 3 7
Table 3. Definition of design variables for preliminary design of a high-subsonic transport-
aircraft wing

Parameter Surrogate e σe Parameter Surrogate e σe

model model
L RSM 0.0360 0.0213 σmax RSM 0.0563 0.0535
Kriging 0.0362 0.0213 Kriging 0.0515 0.0522
L/D RSM 0.0122 0.0099 δmax RSM 0.0227 0.0241
Kriging 0.0123 0.0097 Kriging 0.0227 0.0241
Swing RSM 0.0071 0.0051 Wwing RSM 0.0140 0.0104
Kriging 0.0071 0.0051 Kriging 0.0142 0.0107
Table 4. Evaluation of modeling accuracy

Then the multi-objective optimization for the supercritical wing is performed based on RSM
due to its higher computational efficiency. Weighted sum method is used to transform the
multi-objective optimization into a single-objective optimization. Sequential quadratic
programming method is employed to solve the optimization. One of the candidate wings
with better performance, are selected as the initial point for optimization. The optimal
design is observed by running the analysis codes and the results are listed in Table 5. Where
X 0 and Y 0 is the initial wing scheme and its response, respectively; X * and Y * is the
optimal wing scheme and its actual response, respectively; Ŷ is the response at X *
calculated by the approximation models. For the optimal wing scheme, the largest relative
error of approximation models is no more than 3 percent. It again proves the high accuracy
of the approximation models.
Figure 8 shows the contour of the equivalent stress of the optimal wing. It shows that the
stress is larger in the location of intersection of inner wing and outer wing due to the
inflexion. Figure 9 shows pressure distribution of the optimal wing. It shows that the wing

www.intechopen.com
358 Real-World Applications of Genetic Algorithms

basically meets the design requirements of a supercritical wing. A little bit non-smoothness
of the pressure distribution may be caused by non-uniform deformation of the skin. Figure
10 shows the convergence history of aeroelastic deformation, which shows that fast
convergence of the aeroelastic deformation of the optimal wing.
The optimization , together with the aeroelastic analysis of all candidate wings, only takes
about two days on a personal computer of Pentium(R) 4 CPU 2.8GHz. If more computers
are used to concurrently calculate the performance of different candidate wings, the cost can
be further greatly reduced.

B/m λ θ/(o) Λ/(o) TFS/mm TBS/mm TLS/mm TUS/mm

X0 34.00 0.244 -1.667 29.444 3.333 3.778 6.556 3.889
X* 31.69 0.200 -1.563 28.233 2.232 2.000 4.396 4.057
Wwing
L/103kg L/D Swing/m2 σmax/109pa δmax /m
/103kg

Y0 51.50 27.81 111.94 0.311 1.191 3.850

Y* 53.14 27.40 107.25 0.275 0.991 3.063

Ŷ 54.00 27.45 106.27 0.267 1.000 3.003

Modeling error 1.61% 0.16% 0.91% 2.97% 0.87% 1.85%

Table 5. Optimization results when considering the aeroelastic effect

Fig. 8. Contour of equivalent stress of the optimal wing

www.intechopen.com
Surrogate-Based Optimization 359

-1

-0.5

Cp
-1 0

0.5
z/b = 0.25
-0.5

Cp
0 1 0 0.25 0.5 0.75 1

-1
x/c
0.5 z/b = 0.55
-0.5

1
0 0.25 0.5 0.75 1
x/c
Cp

0.5
z/b = 0.85

1
0 0.25 0.5 0.75 1
x/c

-1

-0.5
-1
Cp

0
-0.5
-1
0.5
z/b = 0.40
Cp

0
-0.5
1
z/b = 0.70 0 0.25 0.5 0.75 1
0.5 x/c
Cp

0.5
z/b = 0.95 1
0 0.25 0.5 0.75 1
x/c
1
0 0.25 0.5 0.75 1
x/c

Fig. 9. Pressure distribution of the optimal wing (Ma=0.76, Re=0.257E+08, α=0o)

1.5
-1.5
Wing-tip Deform in Y-direction /m

1.4
Wing-tip torsion angle /( )
o

-2.0
1.3
-2.5
1.2
-3.0
1.1
-3.5
1.0
-4.0
0.9
-4.5
0.8
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Iteration Iteration

Fig. 10. Convergence history of Y-direction deform and torsion deform on wing tip

www.intechopen.com
360 Real-World Applications of Genetic Algorithms

5. Conclusion
An overview of the existing surrogate models and the techniques about how to use them for
optimization is presented in this chapter. Among the surrogate models, the regression
model such as the quadratic response surface model (RSM) is well suited for a local
optimization problem with relatively simpler design space; interpolation models such as
Kriging or RBFs can be used for highly-nonlinear, multi-modal functions, and thus well
suited for a global problem with relatively more complicated design space. From an
application point of view, the simple framework of surrogate-based optimization is a good
choice for an engineer design, due to the fact that surrogate model can act as an interface
between the expensive analysis code and the optimizer and one doesn’t need to change the
analysis code itself. The drawback of this framework is that the accuracy of optimum only
depends on the approximation accuracy of surrogate model and we generally get an
approximation to the true optimum. In contrast, the bi-level framework with different infill
criteria provides an efficient way to quickly find true optimum without the need of building
globally accurate surrogate models. Multiple infill criteria seem to be a better way to
overcome the drawback of the single infill criterion.
Examples for airfoil and wing designs show that surrogate-based optimization is very
promising for aerodynamic problem with number of design variables being less than about
10. For higher-dimensional problem, the computational cost increases very quickly, which
can be prohibitive. Thus, use of surrogate model for high(er)-dimensional optimization
problems would become an important issue of future work.

6. Acknowledgements
This research was sponsored by the National Natural Science Foundation of China (NSFC)
under grant No. 10902088 and the Aeronautical Science Foundation of China under grant
No. 2011ZA53008

7. References
Elanayar, S. V. T., and Shin, Y. C., “Radial basis function neural network for approximation
and estimation of nonlinear stochastic dynamic systems,” IEE Transactions on
Neural Networks, Vol. 5, No. 4, 1994, pp. 594-603.
Fang, K. T., Lin, D., Winker, P., Zhang, Y., “Uniform design: Theory and application,”
Technometrics, Vol. 42, No. 3, 2000, pp. 237-248.
Forrester, A. I. J., Sóbester, A., and Keane, A. J., “Multi-Fidelity Optimization via Surrogate
Modeling,” Proceedings of the Royal Society A, Vol. 463, No. 2088, 2007, pp. 3251-
3269.
Forrester, A. I. J., Sóbester, A., and Keane, A., “Engineering Design via Surrogate Modeling:
A Practical Guide,” Progress in Astronautics and Aeronautics Series, 226, published
by John Wiley & Sons, 2008.
Giunta, A. A., Wojtkiewicz Jr, S. F., and Eldred, M. S. , “Overview of Modern Design of
Experiments Methods for Computational Simulations,” AIAA paper 2003-649,
2001.

www.intechopen.com
Surrogate-Based Optimization 361

Han, Z. -H., Zhang, K. -S., Song, W. -P., and Qiao, Z. -D., "Optimization of Active Flow
Control over an Airfoil Using a Surrogate-Management Framework," Journal of
Aircraft, 2010, Vol. 47, No. 2, pp. 603-612.
Han, Z.-H., Görtz, S., Zimmermann, R. “On Improving Efficiency and Accuracy of Variable-
Fidelity Surrogate Modeling in Aero-data for Loads Context,” Proceeding of CEAS
2009 European Air and Space Conference, Manchester, UK, Oct. 26-29, 2009.
Han, Z.-H., Zimmermann, R., and Görtz, S., “A New Cokriging Method for Variable-Fidelity
Surrogate Modeling of Aerodynamic Data,” AIAA Paper 2010-1225, 48th AIAA
Aerospace Sciences Meeting Including the New Horizons Forum and Aerospace
Exposition, Orlando, Florida, Jan. 4-7, 2010.
Hardy, R. L., “Multiquadric Equations of Topography and Other Irregular Surface,” Journal
of Geophysical Research, Vol. 76, March, 1971, pp. 1905-1915.
Hicks, R. M., and Henne, P. A., “Wing Design by Numerical Optimization,” Journal of
Aircraft, Vol. 15, No. 7, 1978, pp. 407–412.
Jeong, S., Murayama, M., and Yamamoto, K., "Efficient Optimization Design Method Using
Kriging Model," Journal of Aircraft, Vol. 42, No. 2, 2005, pp. 413- 420.
Jones, D., Schonlau, M., Welch W., “Efficient Global Optimization of Expensive Black-Box
Functions,” Journal of Global Optimization, 1998, Vol. 13, pp. 455-492.
Keane, A. J., Nair, P. B., “Computational Approaches for Aerospace Design: The Pursuit of
Excellence”, John Wiley & Sons, Ltd, Chichester, 2005.
Kovalev, V.E., Karas, O. V. “Computation of a Transonic Airfoil Flow Considering Viscous
Effects and Thin Separated Regions.” La Recherche Aerospatiale (English Edition)
(ISSN 0379-380X), No. 1, 1991, pp. 1-15.
Krige, D. G., “A Statistical Approach to Some Basic Mine Valuations Problems on the
Witwatersrand,” Journal of the Chemical, Metallurgical and Mining Engineering Society
of South Africa, Vol. 52, No. 6, 1951, pp. 119-139.
Laurenceau, J., and Sagaut, P. “Building Efficient Response Surfaces of Aerodynamic
Functions with Kriging and Cokriging,” AIAA Journal, Vol. 46, No. 2, 2008, pp. 498-
507.
Park. J., and Sandberg, I. W., “Universal Approximation Using Radial-Basis-Function
Networks,” Neural Computation, Vol. 3, No. 2, 1991, pp. 246-257.
Powell, M. J. D. “Radial Basis Functions for Multivariable Interpolation: A Review, ”
Algorithms for Approximation, edited by J. C. Mason and M. G. Cox, Oxford Univ.
Press, New York, 1987, Chap. 3, pp. 141-167.
Queipo, N. V., Haftka, R. T., Shyy W., Goel, T., Vaidyanathan, R., and Tucher, P. K.,
“Surrogate-based Analysis and Optimization,” Progress in Aerospace Sciences, Vol.
41, 2005, pp. 1-28.
Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. P., “Design and Analysis of Computer
Experiments,” Statistical Science, Vol. 4, 1989, pp. 409-423.
Simpson, T. W., Mauery, T. M., Korte, J. J., et al., “Kriging Models for Global Approximation
in Simulation-Based Multidisciplinary Design Optimization”, AIAA Journal, Vol. 39,
No. 12, 2001, pp. 2233-2241.
Simpson, T. W., Toropov, V., Balabanov, V., and Viana, F. A. C., “Design and Analysis of
Computer Experiments in Multidisciplinary Design Optimization: a Review of
How Far We Have Come – or Not,” AIAA Paper 2008-5802, 2008.

www.intechopen.com
362 Real-World Applications of Genetic Algorithms

Smola, A. J. and Schoelkopf, B., “A tutorial on support vector regression,” Statistics and
Computing, Vol. 14, 2004, pp. 199-222.
Toal, D. J. J., Bressloff, N. W., Kean, J., “Kriging Hyperparameter Tuning Strategies,” AIAA
Journal, Vol. 46, No .5, 2008, pp. 1240-1252.
Wang, G. G., Shan S., “Review of Metamodeling Techniques in Support of Engineering
Design Optimization,” Journal of Mechanical Design, Vol. 129, No. 4. 2007, pp. 370-
380.
Zhang, K. -S., Han, Z. -H., Li, W. -J., and Song, W. -P., “Coupled Aerodynamic/Structural
Optimization of a Subsonic Transport Wing Using a Surrogate Model,” Journal of
Aircraft, 2008, Vol. 45, No. 6, pp. 2167-2171.

www.intechopen.com
 Real-World Applications of Genetic Algorithms
Edited by Dr. Olympia Roeva

ISBN 978-953-51-0146-8
Hard cover, 376 pages
Publisher InTech
Published online 07, March, 2012
Published in print edition March, 2012

The book addresses some of the most recent issues, with the theoretical and methodological aspects, of
evolutionary multi-objective optimization problems and the various design challenges using different hybrid
intelligent approaches. Multi-objective optimization has been available for about two decades, and its
application in real-world problems is continuously increasing. Furthermore, many applications function more
effectively using a hybrid systems approach. The book presents hybrid techniques based on Artificial Neural
Network, Fuzzy Sets, Automata Theory, other metaheuristic or classical algorithms, etc. The book examines
various examples of algorithms in different real-world application domains as graph growing problem, speech
synthesis, traveling salesman problem, scheduling problems, antenna design, genes design, modeling of
chemical and biochemical processes etc.

How to reference
In order to correctly reference this scholarly work, feel free to copy and paste the following:

Zhong-Hua Han and Ke-Shi Zhang (2012). Surrogate-Based Optimization, Real-World Applications of Genetic
Algorithms, Dr. Olympia Roeva (Ed.), ISBN: 978-953-51-0146-8, InTech, Available from:
http://www.intechopen.com/books/real-world-applications-of-genetic-algorithms/surrogate-based-optimization

InTech Europe InTech China

University Campus STeP Ri Unit 405, Office Block, Hotel Equatorial Shanghai
Slavka Krautzeka 83/A No.65, Yan An Road (West), Shanghai, 200040, China
51000 Rijeka, Croatia
Phone: +385 (51) 770 447 Phone: +86-21-62489820
Fax: +385 (51) 686 166 Fax: +86-21-62489821
www.intechopen.com

View publication stats