Volume 8 Number 0408 ISSN 1979-3898

Journal of
Theoretical and Computational
Studies

Latin Hypercube Sampling for uncertainty analysis

N. A. Wahanani, A. Purwaningsih, T. Setiadipura
J. Theor. Comput. Stud. 8 (2009) 0408
Received: July 7th , 2009; Accepted for publication: August 30th , 2009

Published by
Indonesian Theoretical Physicist Group Indonesian Computational Society
http://situs.opi.lipi.go.id/gfti/ http://situs.opi.lipi.go.id/mki/
 Journal of Theoretical and Computational Studies
Journal devoted to theoretical study, computational science and its cross-disciplinary studies
URL : http://situs.jurnal.lipi.go.id/jtcs/

Editors
A. Purwanto (ITS) J.M. Tuwankotta (ITB)
A. S. Nugroho (BPPT) L.T. Handoko (LIPI)
A. Sopaheluwakan (LabMath) M. Nurhuda (UNIBRAW)
A. Sulaksono (UI) M. Sadly (BPPT)
B. E. Gunara (ITB) M. Satriawan (UGM)
B. Tambunan (BPPT) P. Nurwantoro (UGM)
F.P. Zen (ITB) P. W. Premadi (ITB)
H. Alatas (IPB) R.K. Lestari (ITB)
I.A. Dharmawan (UNPAD) T. Mart (UI)
I. Fachrudin (UI) Y. Susilowati (LIPI)
Z. Su’ud (ITB)
Honorary Editors
B.S. Brotosiswojo (ITB) M.O. Tjia (ITB)
M. Barmawi (ITB) P. Anggraita (BATAN)
M.S. Ardisasmita (BATAN) T.H. Liong (ITB)
Guest Editors
H. Zainuddin (UPM) K. Yamamoto (Hiroshima)
T. Morozumi (Hiroshima)

Coverage area
1. Theoretical study : employing mathematical models and abstractions of a particular field in an attempt to explain known or
predicted phenomenon. E.g. : theoretical physics, mathematical physics, biomatter modeling, etc.
2. Computational science : constructing mathematical models, numerical solution techniques and using computers to analyze
and solve natural science, social science and engineering problems. E.g. : numerical simulations, model fitting and data
analysis, optimization, etc.
3. Cross-disciplinary studies : inter-disciplinary studies between theoretical study and computational science, including the
development of computing tools and apparatus. E.g. : macro development of Matlab, paralel processing, grid infrastructure,
etc.
Types of paper
1. Regular : an article contains an original work.
2. Comment : an article responding to another one which has been published before.
3. Review : an article which is a compilation of recent knowledges in a particular topic. This type of article is only by invitation.
4. Proceedings : an article which has been presented in a scientific meeting.
Paper Submission
The submitted paper should be written in English using the LATEX template provided in the web. All communication thereafter
should be done only through the online submission page of each paper.
Referees
All submitted papers are subject to a refereeing process by an appropriate referee. The editor has an absolute right to make the
final decision on the paper.

Reprints
Electronic reprints including covers for each article and content pages of the volume are available from the journal site for free.

Indonesian Theoretical Physicist Group Indonesian Computational Society

Secretariat Office : c/o Group for Theoretical and Computational Physics, Research Center for Physics -
LIPI, Kompleks Puspiptek Serpong, Tangerang 15310, Indonesia
http://situs.opi.lipi.go.id/gfti/ http://situs.opi.lipi.go.id/mki/

c 2009 GFTI & MKI

° ISSN 1979-3898
J. Theor. Comput. Stud. Volume 8 (2009) 0408

Latin Hypercube Sampling for uncertainty analysis

N. A. Wahanani, A. Purwaningsih and T. Setiadipura
Pusat Pengembangan Informatika Nuklir BATAN, Kawasan Puspiptek Serpong Gedung 71 (lt. 1), Tangerang
15314, Indonesia

Abstract : The assessment and presentation of the effects of uncertainty are now widely recognized as important
parts of anlysis for complex system. Thus, uncertainty analysis is an important tools for a wide area of application, from
the nuclear reactor waste analysis to economic calculation. Sampling-based method is one of the powerful method for
the uncertainty analysis. There are several steps needed to run an uncertainty analysis,(i) construction of distribution to
characterize subjective uncertainty, (ii) sampling procedures, (iii) propagation of uncertainty through model, (iv) display
of uncertainty in model prediction. In this paper, development of the Latin Hypercube Sampling (LHS) module at PPIN-
BATAN as the sampling procedure in the uncertainty analysis is reported. The LHS module is able to demonstrate the
stability of the LHS method in estimating the cumulative density function. This development is a platform for further
advance sampling analysis method, and a step to build an in-house complete tool of uncertainty analysis. Review of the
LHS method and its superior to standard random sampling and importance sampling is also discussed.
Keywords : sampling-based uncertainty analysis, Latin Hypercube Sampling
E-mail : sintaadi@batan.go.id

Received: July 7th , 2009; Accepted for publication: August 30th , 2009

1 INTRODUCTION analysis inputs x = [x1, x2, ., xnX]. In turn, uncer-

The assessment and presentation of the effects of
uncertainty are now widely recognized as important
parts of anlysis for complex system1. Thus, uncer-
tainty analysis is an important tools for a wide area
of application, from the nuclear reactor waste analysis
to economic calculation.
Specifically, uncertainty analysis refers to the de-
termination of the uncertainty in analysis results that
derives from uncertainty in analysis inputs, slightliy
different from sensitivity analysis which is refers to
the determination of the contributions of individual
uncertain inputs to the uncertainty in analysis results.
The uncertainty under consideration here is often ref-
fered to as epistemic uncertainty which derives from a
lack of knowledge about the appropriate value to use
for a quantity that is assumed to hace a fixed value in
the context of a particular analysis.
This uncertainty, in the conceptual and computa-
tional organization of an analysis, is generally consid-
ered to be distinct from aleatory uncertainty which
arises from an inherent randomness in the behaviour
of the system under study. [2]
The underlying idea is that analysis results y(x) =
[y1(x), y2(x), ., ynY(x)] are functions of uncertain
tainty in x results in a corresponding uncertainty in
y(x). This leads to two questions: (i) What is the
uncertainty in y(x) given the uncertainty in x, and
(ii) How important are the individual elements of x
with respect to the uncertainty in y(x) The goal of
uncertainty analysis is to answer the first question,
and the goal of sensitivity analysis is to answer the
second questions.
In practice, the implementation of an uncertainty
analysis and the implementation of a sensitivity anal-
ysis are very closely connected on both a conceptual
and a computational level. [3] The illustration of the
above problem is depicted in Fig. 1.
There are several steps for the sampling-based un-
certainty analysis, as will be mentioned in the next
section, this paper focus on the sampling phase. The
objective of the research is to develop a sampling mod-
ule using latin hypercube method as part of the un-
certainty analysis tools, and to demonstrate the sta-
bility of the latin hypercube sampling estimate com-
pare to standard random sampling. This development
is a platform for further advance sampling analysis
method, and a step to build an in-house complete tool
of uncertainty analysis. Review of the LHS method
and its superior to standard random sampling and im-

c 2009 GFTI & MKI

° 0408-1
2 Latin Hypercube Sampling for uncertainty...
Figure 1: Illustration of uncertainty analysis of the out-
put which arises from uncertainty of the input4 [4].
portance sampling is also discussed.
2 SAMPLING-BASED UNCERTAINTY
ANALYSIS
Sampling-based methods are based on the use of a
probabilistic procedure to select model input and re-
sult in a mapping between analysis inputs and analy-
sis outcomes that is then used to produce uncertainty
and sensitivity analysis results. Desirable features of
Monte Carlo analysis include (i) extensive sampling
from the ranges of the uncertain variables, (ii) un-
certainty results that are obtained without the use
of surrogate models (e.g., Taylor series in differential
analysis and response surfaces in RSM), (iii) extensive
modifications of, or manipulations with, the original
model are not required (Le., as is the case for the other
techniques), (iv) the extensive sampling from the indi-
vidual variables facilitates the identification of nonlin-
earities, thresholds and discontinuities, (v) a variety
of sensitivity analysis procedures are available, and
(vi) the approach is conceptually simple, widely used,
and easy to explain. The major drawback is computa-
tional cost. This is especially the case if long-running
models are under consideration or probabilities very
Figure 2: Comparation of SRS (left) and LHS (right)
where x1(100, 3, 0.1) and x2(100, 5, 0.2) are normal distri-
bution.
0408-2
Figure 3: Figure3.Comparation of SRS (left) and LHS
(right) where x1 and x2 are triangular distribution.
close to zero or one must be estimated [1].
There are five basic components that underlie the
implementation of a sampling-based uncertainty and
sensitivity analysis: (i) Definition of distributions
D1, D2, ., DnX that characterize the uncertainty in
the components x1, x2, ., xnX of x, (ii) Generation of
a sample x1, x2, ., xnS from the x.s in consistency with
the distributions D1, D2, ., DnX, (iii) Propagation of
the sample through the analysis to produce a mapping
[xi, y(xi)], i = 1, 2, ., nS, from analysis inputs to anal-
ysis results, (iv) Presentation of uncertainty analysis
results (i.e.,approximations to the distributions of the
elements of y constructed from the corresponding ele-
ments of y(xi), i = 1, 2, ., nS), and (v) Determination
of sensitivity analysis results (i.e., exploration of the
mapping [xi, y(xi)], i = 1, 2, ., nS) [3].
This paper is limited to the second phase, gener-
ation of sample, particularly using latin hypercube
sampling. Some type of sampling procedure are used
to generate the sample in Monte Carlo analysis, such
as Simple Random sampling , Stratified sampling and
Latin Hypercube sampling. With random sampling,
there is no assurance that a sample element will be
generated from any particular subset of the sample
space. In particular, important subsets of sample
space with low probability but high consequences are
likely to be missed. Stratified sampling has the ad-
vantage of forcing the inclusion of specified subsets of
sample space while maintaining the probabilistic char-
acter of random sampling but there is a problem with
stratified sampling, the necessity of defining the strata
and calculating their probabilities. when the dimen-
sionality of sample space is high, the determination of
strata and strata probabilities become a major under-
taking.
The determinations are further complicated when
many analysis outcomes are under consideration, in
particular, strata definitions that are appropriate for
one analysis outcome may be inappropriate for other
analysis outcomes. Latin hypercube sampling is based
on a combination of simple random sampling and
stratified sampling techniques that leads to statisti-
cally significant results with substantially fewer real-
3 Latin Hypercube Sampling for uncertainty...
Figure 4: Comparation of SRS (left) and LHS (right)
where x1 and x2 are uniform distribution.
izations [1]. Latin hypercube sampling (LHS) displays
properties between simple random sampling, which in-
volves no stratification, and stratified sampling, which
stratifies on sample space.
The basis of LHS is a full stratification of the sam-
pled distribution with a random selection inside each
stratum, sample values are randomly shuffled among
different variables. To generate probability distribu-
tions, the LHS was performed using the following
steps: (i) Assign an inverse cumulative distribution
function (cdf) for each input variable, (ii) choose the
number of simulations (N) to be performed, (iii) di-
vide the cdf for each variable into N equi-probable in-
tervals, (iv) for each interval, choose a random sample,
x, from the inverse cdf and develop a data set for each
parameter, (v) randomly select from each parameter
input data set to create N model input sets, (vi) use
an analytical or numerical model to determine a real-
ization for each model input set. The current devel-
opment accomodate uniform, triangular, and normal
distribution which are commonly used in practice [5].
The particular different of the LHS and standard ran-
dom sampling is in the third step of the above steps.
3 DEVELOPMENT RESULTS AND ANAL-
YSIS
A software of the LHS and standard random sam-
pling based on the above mentioned procedure is de-
veloped as a stand alone software using JAVA pro-
gramming language. Result of the sampling is tested
to produce an estimate of a monoton simple function
which also demonstrate the powerfull of LHS method
to SRS method. For the normal distribution the in-
puts needed is the sample number, mean, and stan-
dard deviation.
The triangular distribution using four parameter
which are the sample number, minimum value, modus,
and maximum value. The uniform distribution need
three parameter the sample number,minimum, dan
maximum value. For the purpose of comparing
LHS and SRS sampling, simple monotonic function
y(x1, x2) = x1 + x2, where x1 and x2 assumed to be
0408-3
Figure 5: Comparation of SRS (left) and LHS (right)
where x1, x2, and x3 are Triangular (100, 2.5, 6), Uniform
(100, 2, 6), and Normal (100, 7, 0.1) distribusion, respec-
tively.
uncorrelated. The estimate for the cumulative density
function of the y(x1, x2) from both sampling method
are compared.
To check the combination of different input
distribution another simple monotonic function
y(x1, x2, x3) = x1+x2+x3 to accommodate the three
distribution involved. The result of the cdf of the y is
depicted in Fig. 5.
ACKNOWLEDGMENTS
Authors are very gratefull to Dr.Syahril at BATAN
for trigerring the uncertainty problem at computa-
tional division, particularly for the techno-economic
problem. Also to Mrs. Khairina, chair of the Compu-
tational Division for its support in the development
procceses.
JTCS
REFERENCES
[1] J.C. Helton and F. J. Davis, Sandia Report
SAND2001-0417 (2002).
[2] J.C. Helton, J.D. Johnson, C.J. Sallaberry and
C.B. Storlie, Sandia Report SAND2006-2901
(2006).
[3] J.C. Helton, Sensitivity Analysis of Model Out-
put, M.H. Kenneth and F.M. Hernez (eds), Los
Alamos National Laboratory (2005).
[4] S. Tarantula, Global Sensitivity Analysis JRC-
EC (2008).
[5] L.P. Swiller and G.D. Wyss, Sandia Report
SAND2004-2439 (2004).