AIAA-2000-4919

VORONOI DIAGRAM BASED CUMULATIVE APPROXIMATION

FOR ENGINEERING OPTIMIZATION

Kikuo Fujita, Noriyasu Hirokawa and Tomoya Tachikawa

Department of Computer-Controlled Mechanical Systems
Graduate School of Engineering
Osaka University
Suita, Osaka 565-0871, Japan

ABSTRACT based on sensitivity analysis, response surface with

This paper proposes a function approximation scheme
based on Voronoi diagram, a technique of computational
geometry, toward optimization cost saving in real
engineering applications. The scheme consists of a
blending function of local information over spatial
structure among sample points that is topologically
represented with Voronoi diagram, an estimation
method of function gradients through the method
of least squares along each Voronoi region, and
an associated overall optimization algorithm. The
scheme can locally establish and cumulatively update
an approximation function based on the nature of
Voronoi diagram, and this leads efficient and accurate
accumulation of samples toward superior global fidelity
in function approximation. Its validity and advantages
are demonstrated through its application to a two-
dimensional example problem and a welded beam
structure design problem.
1 INTRODUCTION
The surrounding situation of engineering design
optimization based on mathematical programming
techniques seems to be changing, after a wide variety
of algorithms and methods have been developed and
computer-aided engineering tools have become to meet
the needs in product development. That is, since
the effectiveness of optimization has been well proven
and engineering analysis has been well computerized
in conventional fields, how to meet optimization
computation cost and time scale in product development
becomes the most challenging topic in the research field
of engineering design optimization.
Function approximation methods are exploring cost
effectiveness in engineering optimization, for instance,
Copyright
c 2000 by The American Institute of Aeronautics and tion. Computational geometry (e.g., Edelsbrunner, 1987)
Astronautics Inc. All rights reserved.
polynomial expressions, neural networks, etc. (e.g.,
Bathelemy and Haftka, 1993). A major framework of
function approximation is that after an approximation
function is generated based on the information given
from analysis at sample points, an optimization
computation is executed over the approximated function
instead of direct analysis. This paper refers these
methods as batch approximation. While batch
approximation can save optimization cost, it is difficult to
adjust or improve solution fidelity, since approximation
and optimization are separated each other, that it
does not have any embedded feedback mechanism of
optimization results to function approximation.
To overcome such a limitation, Rasmussen (1998) pro-
posed the method called cumulative approximation. Its
concept is that cumulative addition of sample points and
consequent refinement of approximation functions can
gradually improve its fidelity. In his method, the ap-
proximation functions are formulated as a sum of locally
representative tendencies under attenuative weights, and
the sum is rearranged every time a new sample is added
through each optimization computation. This frame-
work tends to increase optimization fidelity simultane-
ously with cumulative addition of new samples. Shi et
al. (1999), Arakawa et al. (1999), etc. also proposed sim-
ilar approximation methods with neural networks, which
have cumulative learning ability. However, these meth-
ods implicitly assume isotropic effects of samplings, cu-
mulative addition of samples affects all structure of ap-
proximation, they cannot explicitly discuss the optimal-
ity of sampling placement, and so forth.
This paper proposes another method of cumulative
approximation based on Voronoi diagram, a technique
developed in computational geometry, in order to over-
come the above shortcomings of cumulative approxima-
is a set of theories and algorithms to topologically ma-

1
American Institute of Aeronautics and Astronautics
 Solution Solution Solution
Optimization Optimization Optimization
Approximation Approximation
Analysis Analysis Analysis
time time time
(i) Without (ii) Batch (iii) Cumulative
approximation approximation approximation
Figure 1 Frameworks of optimization without/with
approximation schemes
nipulate geometrical structure of information over multi-
dimensional space. It is earning much attention due to
its robustness of information processing in various fields.
A concrete application of Voronoi diagram is to find the
optimal point of a new additional facility such as bus
stop, post office under a given city map with existing
facilities so as that accessibility from all people within
the city to one of facilities is optimally increased. This
type of capabilities of Voronoi diagram must be suitable
for spatial structure among sample points toward func-
tion approximation and cumulatively optimal placement
of sample points in optimization computation. Under
these promises, this paper proposes a function approx-
imation method based on Voronoi diagram and its incre-
mental update mechanism toward cumulative approxi-
mation. Further, this paper demonstrates an optimization
computation to show their validity and effectiveness un-
der an overall optimization algorithm that is configured
so as to be integrated with the approximation method.
2 COMPROMISE BETWEEN SOLUTION
OPTIMALITY AND ANALYSIS COST
2.1 Optimization Frameworks
As mentioned in Introduction, as computational en-
gineering analysis has become common tools and de-
sign optimization has been recognized to be effective
in product development, the demand to apply optimiza-
tion methods to real engineering applications is realized.
However, computational analysis required in real engi-
neering applications is often so expensive, and conven-
tional optimization methods have become inadequate not
by their ability but by analysis cost required in their pro-
cess. This shortcoming leads the development of func-
tion approximation and then cumulative approximation.
By summarizing the trend in this direction, we can
define the following three categories of optimization
methods in regard of approximation schemes:
2
American Institute of Aeronautics and Astronautics
Optimization fidelity
(iii) Cumulative
approximation
(i) Optimization
(ii) Batch without
approxi- approximation
mation
Optimization cost
(mostly analysis cost)
Figure 2 Fidelity versus cost in optimization
(i) Optimization without approximation Analysis
is executed every time optimization requires any
information.
(ii) Optimization with batch approximation After
approximation functions are generated from a set
of analysis information, optimization is separately
executed with them.
(iii) Optimization with cumulative approximation As
optimization is proceeded, analysis information is
added step by step to improve approximation.
Since each method consists of analysis module,
optimization module and approximation module, the
relationships among modules can be illustrated as Fig. 1.
As shown in the figure, approximation acts as a cash
mechanism against real analysis. This effects smaller
number of analysis computation and leads to save total
optimization cost. Further, while batch approximation
cannot dynamically update cashed contents, cumulative
approximation can have such ability.
Under the above difference, the relationship between
optimization fidelity, which means how the obtained
solution is close to the true optimum, and optimization
cost, which includes all costs for analysis, approximation
and optimization, should be expected as shown in Fig. 2.
That is, since both (i) and (ii) methods execute only
one optimization run, the meaningful solution cannot
be obtained until the end of optimization process. On
the other hand, since (iii) methods iterate optimization
runs through accumulation in approximation, effective
solutions could be obtained at respective stages of
optimization process.
2.2 Cumulative Optimization
The above expecting advantage of (iii) methods with
cumulative approximation is under the compromise
between the two issues, optimization of a design problem
and optimization of the distribution of sample points for
analysis, at any stage of optimization process.
This compromise can be easily understood under
the concept of limited rationality rather than absolute P9
rationality. The traditional optimization paradigm tends P8
to guarantee to find the absolutely optimal solution,
and it is better to accomplish it with less cost. On P4
P3
the other hand, limited rationality asks us what is a
good optimization method if acceptable cost is anyhow P1 P7
less than necessary cost for the best solution. That is, P2
an appropriate optimization method must not find any
absolutely optimal but provide somehow an effective
solution within the limited cost. This kind of situations P6
P5
must be obvious in real engineering applications as
aforementioned due to expensive cost of analysis, which
can be easily dominated over any other cost. Figure 3 Voronoi diagram
Since such limited rationality depends on allowable
cost, fidelity of solutions must be gradually improved
as much cost is poured into optimization. This approximation scheme. At least, disadvantages of the
leads the situation of (iii) optimization methods above scheme are that the shape of (k) (x) in Eq. (2) is
with cumulative approximation brings the necessity spatially symmetric even if distribution of sample points
of compromising ability between solution optimality is not symmetrical and that the calculation of F(xk ) in
(fidelity) and sampling placement optimality toward less Eq. (3) is relatively expensive in comparison with the
iteration of analysis. calculation of F(xk ) since it usually requires numerical
differentiation in real engineering applications.

3 VORONOI DIAGRAM BASED 3.2 Voronoi Diagram

CUMULATIVE APPROXIMATION
As aforementioned, Voronoi diagram is a technique of
3.1 Superposable Function Approximation computational geometry (e.g., Edelsbrunner, 1987).
Figure 3 illustrates the essence of Voronoi diagram
The approximation of a function F(x) for design with a two-dimensional example, while Voronoi diagram
optimization is often based on local information L(k) (x) can handle any-dimensional space. After a set of points
at a set of sample points xk , (k = 1, 2, , K), where P1 , P2 , , P9 (although these points refer as generators
x = [x1 , x2 , , xJ ]T . in the terminology of computational geometry, this paper
Rasmussen (1998) used the approximation scheme calls them base points for convenience.) are given,
of the following equation with superposition of local perpendicular bisectors are drawn between particular
information in his cumulative approximation: pairs of points. This results in a set of polygons
K (polyhedra in the case of higher-dimensional space) as
(k) (x)L(k) (x) shown in the figure. These polygons are named as
P(K) (x) = k=1
(1) Voronoi regions. For convenience, the overall Voronoi
K diagram for points S = {P1 , P2 , } is denoted as D(S),
(k)
(x) the Voronoi region of a base point Pi is denoted as V (Pi ),
k=1
and V (Pi ) of D(S) is specifically denoted as V (Pi , D(S)),
Where, (k) (x) is the blending function, which is respectively in this paper. Each Voronoi region V (Pi )
denoted as follows with a characteristic distance d: means an area where its base point Pi is the closest base
" # point from any points within it, Pj V (Pi ). For instance,
(k) |x xk |2 this situation can be translated to that when a Voronoi
(x) = exp (2) diagram is drawn for the locations of post offices in a city,
d2
people living within each Voronoi region use the post
Where, || is Euclid norm. office at its base point because it is the closest. Under this
Further, when local information is linear approxima- nature, Voronoi diagram is applied to optimal placement
tion at each sample point, L(k) (x) is denoted as follows: problems of facilities and so forth.
An advanced mathematical nature of Voronoi diagram
L(k) (x) = F(xk ) + F(xk )T [ x xk ] (3) is that any operations can be executed based on
topological structure of graphs among base points rather
While the form of Eq. (1) is fairly general, the forms than analytical data, and it leads its robustness against
of (k) (x) and L(k) (x) can be refined toward another new numerical errors in computation. Another important

3
American Institute of Aeronautics and Astronautics
 P9 R3 Double Voronoi region
P8 of P 1 : W ( P 1 )
P4 Q3 Voronoi region
P4 P
4 P 3 of P 1 : V ( P 1 )
PN P3 P4
P1 Q2 R2
P7 PA
P2 P2
P1
R4 Q4 Q1
R1
P5 P6 P5

ctio g
d 1
Figure 4 Addition of a new point to Voronoi diagram

fun ndin
1

n
d 0

Ble
0

nature is that various operations such as addition of a

base point, deletion of a base point, small move of a base
point, can be executed locally within a small subset of
surrounding Voronoi regions.
For instance, Figure 4 shows a Voronoi diagram where
a new base point PN is added to the diagram D(S)
of Fig. 3 within V (P1 ). As the comparison between
both figures illustrates, it causes the modification of
Voronoi polygons, V (P1 ), V (P3 ), V (P4 ) and V (P9 ), but
the Voronoi polygons, V (P2 ), V (P5 ), V (P6 ), V (P7 ) and
V (P8 ), are not affected by it.
These two characteristics of Voronoi diagram is
expected to be novel for cumulative approximation,
where sample points are mapped to base points, in
addition to the characteristic that each Voronoi region
means its dominating area within a whole space. That
is, the first leads robust approximation scheme as well
as geometric computation. The second more importantly
leads that cumulative addition of a sample point is locally
handled with less cost even against a large number of
sample points.
3.3 Approximation Procedure
As shown in Fig. 3, each base point, i.e., sample point,
has a companying Voronoi region and such a region
corresponds to its dominating region against other base
points, i.e., sample points. When this nature is compared
with Eqs. (1) and (2), the blending function (k) (x) can
be defined so as to reflect the dominating nature of each
sample point through its associated Voronoi region over
the whole Voronoi diagram.
For this direction, we introduce double Voronoi region
as shown in Fig. 5. That is, the Voronoi region of
a base point P1 , V (P1 ), is the polygon Q1 Q2 Q3 Q4 in
the case of the figure. Against this, we introduce a
polygon with double size and the same center point, i.e.,
R1 R2 R3 R4 as its corresponding double Voronoi region.
For convenience, the double Voronoi region of a point
Pi is denoted as W (Pi ) in this paper. The adjoining base
Figure 5 Voronoi diagram based function approxima-
tion
points of Voronoi region of P1 , i.e., P2 , P3 , P4 and P5
are on the border edges of its double Voronoi region,
R1 R2 , R2 R3 , R3 R4 and R4 R1 , respectively. Since each
base point, i.e., sample point has its own approximation
information L(k) (x), approximation at a base point must
exactly follow to its own local information exclusive of
any influences from other points. This indicates that the
blending function of a point P1 , (1) (x), must be 1 at P1
and that it must be 0 at P2 , P3 , P4 and P5 . Furthermore, it
can be required that it is 0 outside of W (P1 ), i.e., on the
border edges, R1 R2 , R2 R3 , R3 R4 and R4 R1 and outside
of it. Under this, it may be also required that it is 0.5
on the border of V (P1 ), i.e., on the edges, Q1 Q2 , Q2 Q3 ,
Q3 Q4 and Q4 Q1 .
To realize such a blending function and to smoothly
interpolate the above characteristic values at specific
points, we introduce the following procedure:
(i) After the point to calculate an approximation is
given as PA , the coordinate vector representation of
which is xA , the double Voronoi regions where PA
belongs are listed. In the case of Fig. 5, PA belongs
to W (P1 ) as well as W (P4 ), W (P5 ) and else.
(ii) For each double Voronoi region, the following
procedure is executed, respectively. It is explained
on W (P1 ).
(a) Find the triangle where PA belongs from the
sub-triangles of the double Voronoi region,
i.e., 4P1 R1 R2 , 4P1 R2 R3 , 4P1 R3 R4 and
4P1 R4 R1 . In this case, it belongs to
4P1 R3 R4 .
(b) By using the distance between an edge R3 R4
and a point PA , which is denoted as d1 , and
the distance between an edge R3 R4 and a

4
American Institute of Aeronautics and Astronautics
 point P1 , which is denoted as d0 , the blending
weight (xA ) is calculated with the following
equation:
 
d1
1 + cos 1
d0
(xA ) = (4)
2 updating estimation of gradients, F(xk ), can be saved,
(iii) By blending all local approximation information
with all blending weights, (xA ), the approximated
value of the function at PA is determined with
Eq. (1).
In the step (iii), the sum of all local information in Eq. (1)
can be replaced with the sum only across the related
double Voronoi regions, since (xA ) is 0 outside such
double Voronoi regions. This means that topologically
local structure of Voronoi diagram is also useful for
saving computational cost of approximation.
The above operation can be extended from two-
dimensional space to any-dimensional one in the similar
way that operations of Voronoi diagram are extended.
3.4 Estimation of Gradient at Each Sample
Point
The linear approximation shown in Eq. (3) is most
simple but rather expensive as aforementioned. The
approximated gradient of F(x) is also estimated based on
Voronoi diagram and the method of least squares. While
least square estimation of gradient requires equal to or
more than n points in n-dimensional space, a Voronoi
region at a certain base point is directly connected
with more than n + 1 associated base points due to the
nature of polyhedron (Voronoi diagram is composed of
polyhedra in general, while it is composed of polygons in
two-dimensional space) rather than the nature of Voronoi
diagram. For instance, the Voronoi region of P1 , V (P1 ), is
associated with P2 , P3 , P4 and P5 in Fig. 5. Besides, when
a Voronoi region is faced on the border of overall target
region, it may not be surrounded with enough number
of base points. In such a case, the closest points must
be additionally supplied from its neighboring Voronoi
regions.
When denoted the associated base points of a base
point Pk as Pk,1 , Pk,2 , , Pk,M (where, M > n), the
estimation of F(xk ) is determined so as to minimize the
following square value E, that is, by solving the equation
E
F(x ) = 0:
k
n o 2
M F(xk,m ) F(xk ) + F(xk )T [xk,m xk ]
as to normalize the influences of respective base points in
accordance with their distances.
This procedure is effective to save the number of
analysis, since it does not require any additional analysis
beyond ones that are used in forming Voronoi diagram
for cumulative approximation. Further, the cost for
since its operation can be localized as well as update of
Voronoi diagram itself.
4 OPTIMIZATION ALGORITHM UNDER
CUMULATIVE APPROXIMATION
4.1 Form of Optimization Problems
Since Voronoi diagram requires a predefined target
region, we consider optimization problems with upper
and lower limits on all design variables. The form of
problems is generally standardized for convenience as
follows in this paper:
Minimize f (x)
Subject to g (x) 0
ig (i = 1, 2, , I )
g g
hih (x) = 0 (ih = 1, 2, , Ih )
x j L x j x jU for all j
(6)
4.2 Overall Algorithm for Optimization
The overall algorithm for optimization with cumula-
tive approximation consists of the following steps:
(i) Randomly provide a certain number of initial
sample points, and execute system analysis at those
points.
(ii) Establish or update the Voronoi diagram based
approximation of functions such as f (x), g1 (x),
g2 (x), , h1 (x), h2 (x), .
(iii) Find a tentative solution by solving an approxi-
mated problem.
(iv) If a solution has enough fidelity or if poured cost
exceeds the limit, terminate the overall procedure.
(v) Determine some new sample points, and execute
system analysis there.
(vi) Go to Step (ii).
To realize this procedure, how to determine new sample
points and which optimization algorithm is appropriate
E =  
xk,m xk 

must be discussed.
m=1

(5) 4.3 Addition of New Sample Information

Where, || is Euclid norm. The distance between base The criteria for selecting new samples in Step
points is added as a denominator in the above equation so (v) are both to select the points that enhance the

5
American Institute of Aeronautics and Astronautics
fidelity of the current tentative optimum and to select P4
the points that improve global approximation fidelity.
Furthermore, it is necessary to take care for maintaining P3
the quasi-isotopicness of each Voronoi region, since PN
irregularly distorted shape of Voronoi regions leads P5 P1
inferior approximation such as locally bumpy shapes P2
along their boundaries. To compromise these issues,
first Voronoi regions where new samples are added P6
are selected toward the former purposes, and then new P7
sampling points are arranged within the specific Voronoi
(a) Straightforward addition of a new sample point
regions, respectively, for the last purpose.
P4
(1) Selection of candidate Voronoi regions
Two criteria for region selection are considered toward P3
both local fidelity of the optimum and global fidelity of
PN
function approximation. That is, one is the region where P5 P 1= P n PN
the tentative optimal solution is gotten in Step (iii) exists. P2
The other is selected as a region where existing sample
points are depopulated.
P6
The latter depopulated region is determined as the P7
Voronoi region where the optimal solution of the
following minimization function exists: (b) Adjusted addition of a new sample point
" #
K
|x xk |2 Figure 6 Local adjustment of a new sample point
(x) = exp
(K)
(7)
k=1 dk 2
solution or the minimal point of (K) (x), under a set
Where, || is Euclid norm, and dk is a characteristic of sample points S and the associated Voronoi diagram
distance, the value of which is determined so as that D(S). Figure 6 (a) shows a situation where Voronoi
each term of the above sum becomes 0.2 at the average diagram is updated immediately after the addition of
distance from xk to all associated base points of Voronoi PN . In such a case, the original V (P1 , D(S)) is
region of xk in the later applications. Equation (7) almost split into the two subparts of V (P1 , D(S {PN }))
represents the density of sample points, since each term and V (PN , D(S {PN })), since PN is very close to
of the sum is attenuated as x is away from an existing P1 . If further new points would be added along the
sample xk . That is, the value of (K) (x) is expected to segment P1 PN , the associated Voronoi regions must
be small in the depopulated zones of samples and to be become to have very thin shapes, which causes inferior
larger in the crowded zones of samples. approximation through the blending mechanism shown
Besides, how frequently new samples are added in Fig. 5. This situation must occur at the neighborhood
under respective viewpoints after every optimization of the optimal solution.
computation of Step (iii) must be controlled, while a new To avoid such a scenario, the location of a new sample
sample is added for each viewpoint in the applications point must be adjusted from PN to P0 N within the Voronoi
demonstrated in the later sections. For instance, if the region as shown in Fig. 6 (b) by ideally solving the
latter viewpoint is occasionally eliminated, the tentative following optimization problem:
solution may be quickly converged to the true optimum,

but it would be possible to lose the global optimum. find P0 N



2

(2) Local adjustment of new sample points that minimize w1 P N PN 0

Z

The local arrangement of new sample points is under
x Pi dx
V (Pi , D(S {P0 N })) (8)
the same concept for adjusting the location of a new + w2 0 Z



facility in a specific region over a map with a set Pi S {P N } dx

of existing facilities so as to maximize the average V (Pi , D(S {P0 N }))

0


accessibility to any facility from all points. subject to P N V (Pn , D(S))
Figure 6 illustrates the necessity of local adjustment
with a situation where PN is determined as a candidate Where, w1 and w2 are weighting factors, and Pn is given
for a new sample point, which is either tentative optimal by PN V (Pn , D(S)), which means the base point of

6
American Institute of Aeronautics and Astronautics
the Voronoi region where PN exists. is Euclid norm
between two points. Further, w1 = 0 for the adjustment
of depopulated regions.
In the actual implementation, since the integral
over Voronoi regions in Eq. (8) is tremendously time-
consuming, and since the strictly optimal solution
of Eq. (8) is not necessary toward efficient function
approximation, another similar but simple formula with
less computation is introduced.
4.4 Optimization Method
The overall optimization algorithm must solve the
three types of optimization problems, the approximated
optimization problem, the problem for searching a
depopulated region, and the problem for locally adjusting
new sample points. These problems are expected to
be constrained nonlinear continuous-space problems,
and it is clear that the second is multi-modal. For
this reason, any algorithm that has a capability to
find a global optimum under multi-modality such as
genetic algorithms must be appropriate as a common
optimization algorithm. However, this paper uses
a method that first the SQP (successive quadratic
programming) method is run several times from different
initial points, which are randomly generated, and then
the best solution is selected from the gotten candidate
solutions. While this may take relatively much cost, it
is enough for the validation purpose of Voronoi diagram
based cumulative approximation and its associated
overall optimization algorithm.
in the figure, this problem has a global optimum
f = 2.210 at (0.129, 1.579) and two local optima
f = 1.247 at (1.490, 1.445) and f = 1.720 at
(1.574, 1.381).
5.2 Optimization Result
Figure 8 shows how approximation of functions, i.e.,
f (x1 , x2 ), g1 (x1 , x2 ) and g2 (x1 , x2 ) are updated through
accumulation of sample points and how the tentative
optimal solution is getting close to the true optimum.
In Fig. 8 (a), a solution is gotten as f = 1.726 at
(0.100, 1.248) under the initial approximation with six
sample points. The shapes of approximated functions
are rather different from the true ones shown in Fig. 7.
In Fig. 8 (b), the solution is updated as f = 2.218
at (0.118, 1.576) under the approximation with 16
sample points, i.e., after five iteration of optimization
runs. The shapes of approximated functions are still
different from the true ones, especially in the low-
middle zone, but the solution is fairly close to the true
one. In Fig. 8 (c), the solution is updated as f =
2.228 at (0.121, 1.575) under the approximation with
26 sample points, i.e., after 10 iteration of optimization
runs. The shapes of approximated functions are almost
similar to the true ones. In Fig. 8 (d), the solution
is finally converged to f = 2.211 at (0.129, 1.579)
under the approximation with 52 sample points, i.e.,
after 23 iteration of optimization runs. The shapes of
approximated functions are slightly improved from ones
shown in Fig. 8 (c), while it leads a solution that is
perfectly close to the true one.

5 TWO-DIMENSIONAL NUMERICAL 5.3 Comparison with Direct Optimization

EXPERIMENT In order to ascertain the advantage of Voronoi diagram
5.1 Example Problem based cumulative approximation, the optimization his-
tory is compared against the optimization where func-
This section demonstrates an optimization computa- tions are directly evaluated in optimization computation.
tion of a two-dimensional and multimodal problem to Figure 9 shows the history of objective function and
validate the effectiveness of our cumulative approxima- design variables in the optimization with cumulative
tion scheme. The example problem used here is as fol- approximation over the number of function evaluation
lows: calls. Since it depends on the random number generation

for initial sample points, the figure shows ten trials
Minimize


from different seeds of random number generator.
x1 2

f (x1 , x2 ) = x1 x2 sin x1 + + x1 1.5 x2

While the problem shown in Eq. (9) has two local
10


optima in addition to a global optimum, every trial
Subject to
can find the global optimum. Further, the solution is
2
g1 (x1 , x2 ) = 9 (x1 1) 2 + x2 0
5 (9) smoothly converged to the true solution as the times of



function evaluation is increased, similarly in all trials.
g2 (x1 , x2 ) = 14 x1 2 2 x2 0


This indicates that the overall optimization algorithm


2 x1 2
accomplishes the concept of optimization under limited


rationality shown in Fig. 2. In this case, it seems
2 x2 2
that about 25 times of function evaluation can reach
Figure 7 shows the contour plot of functions and an effective solution independent from initial samples,
perspective view of the objective function. As shown while the example is a very small problem.

7
American Institute of Aeronautics and Astronautics
 2.0 f = 3.5
3.0
1.5 2.5
global 2.0
optimum 1.0 1.5 4
1.0 3
0.5 2
0.5 1
0.0
f 0
feasible region -0.5
0.0 x2 -1.0
-1
-2
-1.5 -3
-0.5 -2.0 -4 2.0
1.5
local -2.5 1.0
optimum -1.0 -3.0 0.5
local -3.5 -2.0
-1.5 0.0
optimum -4.0 -1.0
-0.5 -0.5 x 2
-1.5 0.0
0.5 -1.0
x1 1.0 -1.5
g1 = 0 1.5 2.0 -2.0
-2.0
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 g2 = 0
x1
(a) Contour plot of functions (b) Perspective view of the objective function
Figure 7 An optimization problem

2.0 2.0

1.5 1.5

1.0 1.0

0.5 0.5

0.0 x2 0.0 x2
-0.5 -0.5

-1.0 -1.0

-1.5 -1.5

-2.0 -2.0
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
x1 x1
(b) with additional 10 points f = 3.5
(a) with initial six points
(16 points in total) 3.0
2.5
2.0
2.0 2.0 1.5
1.0
1.5 1.5 0.5
0.0
1.0 1.0 -0.5
-1.0
0.5 0.5 -1.5
-2.0
-2.5
0.0 x2 0.0 x2 -3.0
-3.5
-0.5 -0.5 -4.0
g1 = 0
-1.0 -1.0
g2 = 0
-1.5 -1.5 sample
point
-2.0 -2.0
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 tentative
x1 x1 optimum

(c) with additional 20 points (d) with additional 46 points

(26 points in total) (52 points in total)
Figure 8 Optimization through cumulative approximation

8
American Institute of Aeronautics and Astronautics
Optimal objective 1.0 1.0

Optimal objective
0.0 0.0

-1.0 -1.0 Local optimum

Global optimum
-2.0 Global optimum -2.0

-3.0 -3.0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Function evaluation Function evaluation
(a) Objective function (a) Objective function

2.0 2.0
x 2 (global optimum)
x2
Design variables

Design variables
1.0 1.0

x 1 (global optimum)
0.0 0.0
x1

-1.0 -1.0
x 2 (local optimum)
x 1 (local optimum)
-2.0 -2.0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Function evaluation Function evaluation
(b) Design variables (b) Design variables

Figure 9 Optimization history of ten trials with Figure 10 Optimization history of ten trials with direct
cumulative approximation function calls

Figure 10 shows the optimization history of SQP 6 APPLICATION TO DESIGN PROBLEM

with direct function evaluation. Since SQP cannot find
the global optimum of multi-modal problems due to
its characteristic and its result depends on the initial
point, the figure shows ten trials from different initial
points that are randomly generated as well as Fig. 9.
In these cases, five trials can find the global optimum
at (0.129, 1.579), while other five trials fall into a local
optimum at (1.490, 1.445). Further, the convergence
speed to the final solution against the iteration of
function evaluations is obviously slower than ones with
cumulative approximation shown in Fig. 9, while this
is mainly caused by the numerical differentiation of
functions in SQP process, which is shown as flat steps
in Fig. 10.
This comparison sufficiently shows the potential
advantages of the proposed approximation scheme.
OF WELDED BEAM STRUCTURE
6.1 Problem Formulation
This section demonstrates another application of the
proposed optimization scheme to the welded beam
structure design problem (Ragsdell and Philips, 1976;
Rao, 1996), where the welded beam shown in Fig. 11
is designed for minimum cost subject to constraints
on shear stress (), bending stress in the beam (),
buckling load on the bar (Pc ), end deflection on the beam
(), and side constraints. This optimization problem
is formulated under the design variable vector x =
[x1 , x2 , x3 , x4 ]T = [h, l, t, b]T as follows:

9
American Institute of Aeronautics and Astronautics
 6.0

Optimal objective
5.0
h h
A P
4.0
l t
L b
3.0

Figure 11 Design problem of welded beam structure

2.0
0 30 60 90 120 150 180
Function evaluation
(a) Objective function
Minimize
f (x) = 1.10471 x1 2 x2 10.0
+ 0.04811 x3 x4 (L + x2 )
x3
Subject to 8.0

Design variables
(x) max 0
x2
(x) max 0 6.0

x1 x4 0
4.0
0.10471 x1 2 + 0.04811 x3 x4 (L + x2 )
5.0 0 2.0
0.125 x1 0
x1 x4
(x) max 0 0.0
0 30 60 90 120 150 180
P Pc (x) 0 Function evaluation
0.1 x1 2.0, 0.1 x4 2.0
(b) Design variables
0.1 x2 10.0, 0.1 x3 10.0
Where, Figure 12 Optimization history of five trials for
r
x2 beam design problem with cumulative
(x) = (0 )2 + 2 0 00 + (00 )2
2R approximation
P
0 =
2 x1 x2
00 MR max = 0.25 in.
= The optimal solution of this optimization problem is
J
 x2  f = 2.386 at x = [0.2455, 6.1960, 8.2730, 0.2455]T
M = P L+
2
v ! 
u  6.2 Optimization Result
u x2 2
x1 + x3 2
R = t +
4 2 Figure 12 shows the five-trials of optimization history
"   # with Voronoi diagram based cumulative approximation.
x2 2 x1 + x3 2
J = 2 2x1 x2 + In each trial, after 20 sample points are randomly
12 2
given from different seeds of random number generator,
6PL optimization computation is executed. As shown in the
(x) =
x4 x3 2 figure, the procedure can find the optimal solution with
4 P L3 about 120 function evaluations, while the history are
(x) =
Ex3 3 x4 rather jumpy as few tentative solutions occasionally fall
q ! into inappropriate ones.
r
4.013 E x3 2 x4 6 /36 x E Figure 13 shows the five-trials of optimization history
Pc (x) = 1 3
L2 2L 4G by SQP with direct function evaluation from different
initial variables in order to verify the advantage of
Further, P = 6000 lb, L = 14 in, E = 30 106 psi, G = cumulative approximation. While the SQP based
12 106 psi, max = 13, 600 psi, max = 30, 000 psi, and optimization is very stable and smooth under a

10
American Institute of Aeronautics and Astronautics
 6.0 7 SUMMARY
This paper proposed a Voronoi diagram based function
Optimal objective

5.0 approximation scheme as a technique for engineering

design optimization, and demonstrated its validity and
effectiveness through a tentatively configured overall
4.0
optimization algorithm. The proposed approximation
scheme shows some advantages and promises in
3.0 efficiency, fidelity and robustness in approximation,
convergence smoothness through sample accumulation
and so forth under the nature of computational
2.0
0 30 60 90 240 270 300 geometry paradigm, while the overall algorithm and
Function evaluation associated optimization method must be refined and the
advantages must be further demonstrated in the real-scale
(a) Objective function
engineering applications as future works.
10.0

x3 ACKNOWLEDGMENTS
8.0
Design variables

This research is partially supported by Research

x2 for the Future Program of the Japan Society for
6.0
the Promotion of Science (JSPS) under Methodology
4.0 for Emergent Synthesis Project (project number
96P00702).
2.0

x1 x4
0.0
0 30 60 90 240 270
Function evaluation
(b) Design variables
Figure 13 Optimization history of five trials for beam
design problem with direct function calls
macro point of view, the required number of function
approximations reaches to over 250.
As demonstrated through the comparison between
Figs. 12 and 13, Voronoi diagram based cumulative
approximation has advantages in reduction of the
number of function evaluation that can lead computation
cost saving and in early convergence to semi-optimal
solutions. However, the optimization history of welded
beam structure that has four design variables shows some
unstable tendency, which does not appear in the two-
dimensional example problem of the previous section.
This implies that the proposed approximation scheme
must have some weakness against higher-dimension
optimization problems, while this may be a common
issue across any approximation schemes. Further, any
executable replacement of Eq. (8) for local adjustment of
new samples must become more difficult and sensitive in
the cased of higher-dimension problems. The study on
how to overcome such difficulties must be important and
essential future works.
REFERENCES
300
Arakawa, M., Nakayama, H., Ishikawa, H., 1999,
Optimum design using radial basis function network
and adaptive range genetic algorithm, Proceedings
of the 1999 ASME Design Engineering Technical
Conferences, Paper Number DETC99/DAC-8637.
Bathelemy, J. -F. M, Haftka, R. T., 1993, Approximation
concepts for optimum structural design a review,
Structural Optimization, Vol. 5, pp. 129-144.
Edelsbrunner, H., 1987, Algorithms in Computational
Geometry, Springer-Verlag.
Rao, S. S., Engineering Optimization, Theory and
Practice, Third Edition, (1996), pp. 534-535.
Rasmussen, J., 1998, Nonlinear programming by
cumulative approximation refinement, Structural
Optimization, Vol. 15, pp. 1-7.
Shi, Q., Hagiwara, I., Takashima, F., 1999, The
most probable optimal design method for global
optimization, Proceedings of the 1999 ASME Design
Engineering Technical Conferences, Paper Number
DETC99/DAC-8635.
Ragsdell, K. M, and Philips, D. T., Optimal Design
of a Class of Welded Structures Using Geometric
Programming, Transactions of the ASME, Journal
of Engineering for Industry, Vol. 98, No. 3, (1976),
pp. 1021-1025.

11
American Institute of Aeronautics and Astronautics