Professional Documents
Culture Documents
net/publication/228934887
CITATIONS READS
77 232
3 authors:
Kwok-Leung Tsui
City University of Hong Kong
284 PUBLICATIONS 8,538 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
General Research Fund: Reliability and Degradation Modeling for Rechargeable Battery View project
All content following this page was uploaded by Ying Xiong on 20 May 2014.
DETC2007-35782
The other important issue associated with a surrogate model The proposed variable-fidelity optimization framework in this
based design optimization is the sequential updating of the work is depicted in Figure 1. In contrast to the classical variable
surrogate model with additional data, or called sequential fidelity optimization strategy, none of the LF and HF models is
sampling. One category of existing sequential sampling directly invoked during optimization. Instead, a model fusion
techniques aims at global accuracy in a space-filling fashion, technique is applied to combine information from both LF
e.g., based on criteria such as the Integrated Mean-Squared model yl(x) and HF simulations yh(xi) (i=1,…,N) to yield a
Error, MaxiMin/MiniMax, and Maximum Entropy. Recent surrogate model ys(x), over which optimization is performed.
studies [11-16] have proposed adaptive strategies based on the The proposed model fusion approach follows a Bayesian
knowledge obtained from the preceding surrogate model. modeling framework, in which the surrogate model is a
However, it is generally unaffordable to conduct enough combination of an augmented LF model with linear scaling plus
simulation runs to cover the entire input space when a bias function that characterizes the remaining difference with
simulations are expensive. A more efficient sequential the HF model. With the Bayesian approach, the uncertainty of
sampling strategy is the so-called objective-oriented approach ys(x) in predicting HF models can be quantified. As previously
that brings the design objective into account. The initial stated, such type of uncertainty is called the ‘interpolation
utilizations of the objective-oriented approach and the concept uncertainty’ due to the lack of sufficient HF simulations. When
of interpolation uncertainty could be seen in the Efficient a LF model is cheap to run, it will be directly used for model
Global Optimization (EGO) algorithm developed by Jones et al. fusion without fitting a surrogate model to replace it. When a
[17] using the idea of expected improvement (EI) originally LF simulation is expensive, the surrogate of LF model (dashed
proposed by Mockus et al. [18]. Although EI was demonstrated line box in Figure 1) could be used to replace the original LF
to be well suited for the global deterministic optimization, model. After the surrogate model ys(x) is obtained based on
alternative sequential sampling criteria [19] [13] were model fusion, design optimization is performed using the
investigated and shown to have various merits in making the predictor (or posterior mean) of the Bayesian surrogate model
trade-off between optimizing a design objective and reducing ys(x). The design confidence (DC) of the newly identified
the interpolation uncertainty. All of these existing objective- optimum x* will then be assessed considering the interpolation
{ }.
y
(12) requires the search of the most competing design w.r.t. the
optimal design x*, called the most competing (MC) point xmc,
by minimizing the probability P and treating the -1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
indifferentiable region X0 as the constraint. The obtained xmc is x
then used to calculate the design confidence of x* through Figure 5. The HF model and LF model, with 3 initial HF
DC (x*) = P[ Z x* (x mc ) > 0] . (17) samples (Example 1)
For sequential sampling in global optimization, the commonly
used stopping criteria are based on the convergence behavior in Different from using a single fidelity model, in variable fidelity
either design space of x, or performance space of y [3] [27]. optimization, very few data from HF simulations are available,
Although generally applicable, none of them provide thus LF model may be used to capture the global trend of a HF
probabilistic measure regarding the validity of an optimal model. We note in Figure 5 that the scaled LF model is fairly
design considering model uncertainty. In this work, we view close to the HF model. The trend information provided by the
the sequential sampling as a process of reducing the scaled LF model is integrated into the Bayesian surrogate
interpolation uncertainty of surrogate models, as well as model to enhance the accuracy of prediction.
improving the confidence in accepting a design solution. We
propose to use the stopping criterion based on the design From Stages 1 to 5, additional 5 sampling points are
confidence values of any two consecutive stages. If any two sequentially collected following the PSC strategy we proposed
consecutive design confidences DC(x*) meet (is higher than) a in Section 2.3. Using the Bayesian Gaussian process modeling
desired confidence level (e.g., 90%, 95%) prescribed by approach described in Section 2.2, the posterior mean and the
designers, the sequential sampling process can then be interpolation uncertainty (95% prediction interval (PI)) of
terminated. Often times, once resources (time and cost, etc) are y s (x) are shown in Figure 6, with all 5 sequential points
exhausted, designers would have to accept the current best annotated with the stage number besides. It is noted that
design. With the information of design confidence, designers although two local minimums exist in the HF model, only one
of the 5 samples is used to explore the secondary local
y
100
In Figure 6, points in the local region (0.215~0.253) 0 1
surrounding x* (=0.2330) is identified as indifferentiable to x*
0 0.5
with certain design tolerance H (=0.023) and confidence 0.5
1 0
C X 0 (=99%). The design point x=0.746, shown with large x1
x2
interpolation uncertainty, is identified as the most competing Figure 7. The 3-D plots of HF and LF models
point xmc (marked with triangle). The xmc point is considered as
1 15
the most competing point to x* among all design points outside 0 0
15 3 0
70
10
90 170gloabl optimum
30
HF 1
the indifferentiable region X0 because its lower bound of the
-1 0
0.9 13 0
0 HF local optimum 11
90
prediction interval is very close to the performance at x*. 11
10
0.8 50 0
Stage (5) 70 90
-1 0
30
0.7
Ys hat 30 70 50
95% PI 0.6
30
10
1 50
initial HF samples
0.5
50
x2
sequential HF samples
10
x star from Ys
11 0
7090
0.4
10
10
30
most competing point
0.5
30
10
0.3
Ys hat
0.2
15 01711900
50
0
23 0 25
0.1
2
mc
0
x
13
30
70 0
11
10
10
0
0
3
9
1 0
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
2 x1
-0.5
5 4
Figure 8. The contour plot of HF model marked with x*HF
=[0.0912 0.9325], yh(x*HF)=-19.142.
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Stage (0)
x 1
Initial HF samples
Branin Function)
0.6
In this example, the effectiveness of the proposed sequential
0.5
x2
has three global minima with exactly equal performance values. 0.3
The problem has been studied in literature for various purposes. 0.2
In this work we modified the original Branin function by
0.1
adding an additional small ‘tip’ term so that it has only one
global minimum, while the other two become local minima but 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
stay competing to the global one. The modified Branin function x1
0.4 2 both Figures 10 and 11. It is noted that xmc is always located
0.3 54
outside of X0 and represents the most competing design w.r.t
8 10 x*. Based on Eqs. (17) and (18), the design confidence DC(x*)
0.2 6
9 3 achieved using the EI criterion and our proposed PSC strategy
0.1
7
is 87.12% and 99.99%, respectively.
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 15
x1 PSC strategy
10 EI criteria
s
Figure 10. The plot of yˆ (x) at Stage 10, with x*= [0.5276, One-shot sampling
ymin from HF
5
0.2132]: sequential sampling points generated by EI criterion
Stage (10) 0
ymin(k) from Ys
1
2 1
0.9 10 -5
4 Initial HF samples
0.8 5 Sequential HF samples
-10
x star from HF
0.7 x star from Ys
3 -15
Indifferentiable region
0.6
most competing point
-20
0.5
x2
8
0.4 -25
0 1 2 3 4 5 6 7 8 9 10
Stage k
0.3
9 7
0.2 Figure 12. The comparison of the history plots of y*
0.1 1
6 PSC strategy
0 0.95 EI criteria
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
One-shot sampling
x1
0.9
Figure 11. The plot of yˆ s (x) at Stage 10, with x*= [0.1011, 0.85
0.75
In Figure 11, where the PSC is applied, x* (marked by square)
0.7
is identified at [0.1011, 0.9156], fairly close to x*HF. From the
locations of the sequential points (marked by solid circles), it is 0.65
found that most of the sampling points are placed in the local 0.6
region around the global minimum, while the rest of them are 0.55
placed elsewhere to reduce the interpolation uncertainty of the
surrogate model. It is found that at the early stages when the 0.5
0 1 2 3 4 5 6 7 8 9 10
uncertainty of the surrogate model is large, the sampling Stage k
procedure explores the model space more rather than focusing Figure 13. The comparison of the history plots of DC(x*)
on any local promising region. After sufficient samples have (H = 12 (3% of yrange) and C X 0 = 0.95)
been accumulated and the uncertainty of a surrogate model is
reduced, more samples are used for local refinement of the
global surrogate model in the region that is in favor of the Figures 12 and 13 show the history plots of the response value
design objective. y* and design confidence value DC(x*) from the surrogate
models at different stages. The results from using the EI
The most competing point (xmc) w.r.t. the optimal x* obtained criterion and the PSC strategy are compared in both figures. It
from the surrogate model, is marked as a triangle in both plots is observed that the DC level consistently increases with more
in Figures 10 and 11, where the indifferentiable region X0 is sample points when using the PSC strategy which yields a high
1
0.4
0.95
0.3
0.9 Setting A: H=0.24 Cx0=0.95
0.2
Setting B: H=0.08 Cx0=0.99
0.85
0.1
0.8
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
DC
x1 0.75
0.65
(One-shot sampling of 5+10=15 points)
0.6