This paper applies inverse transform sampling to sample training points for surrogate models. Inverse transform sampling uniformly generates a sequence of real numbers ranging from 0 to 1 as the probabilities at sample points. The coordinates of the sample points are evaluated using the inverse functions of Cumulative Distribution Functions (CDF). The inputs to surrogate models are assumed to be independent random variables. The sample points obtained by inverse transform sampling can effectively represent the frequency of occurrence of the inputs. The distributions of inputs to the surrogate models are fitted to their observed data. These distributions are used for inverse transform sampling. The sample points have larger densities in the regions where the Probability Density Functions (PDF) are higher. This sampling approach ensures that the regions with higher densities of sample points are more prevalent in the observations of the random variables. Inverse transform sampling is applied to the development of surrogate models for window performance evaluation. The distributions of the following three climatic conditions are fitted: (i) the outside temperature, (ii) the wind speed, and (iii) the solar radiation. The sample climatic conditions obtained by the inverse transform sampling are used as training points to evaluate the heat transfer through a generic triple pane window. Using the simulation results at the sample points, surrogate models are developed to represent the heat transfer through the window as a function of the climatic conditions. It is observed that surrogate models developed using the inverse transform sampling can provide higher accuracy than that developed using the Sobol sequence directly for the window performance evaluation.
1. Improving the Accuracy of Surrogate Models
Using Inverse Transform Sampling
Junqiang Zhang*, Achille Messac#, Jie Zhang*, and Souma Chowdhury*
* Rensselaer Polytechnic Institute, Department of Mechanical, Aerospace, and Nuclear Engineering
# Syracuse University, Department of Mechanical and Aerospace Engineering
53rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics
and Materials Conference
8th AIAA Multidisciplinary Design Optimization Specialist Conference
April 23 - 26, 2012
Honolulu, Hawaii
2. Introduction
• Sampling is an important component of optimization, numerical
simulations, design of experiments and uncertainty analysis.
• Surrogate modeling is concerned with the construction of
approximation models to estimate the system performance, and to
develop relationships between specific system inputs and outputs.
• It is expected that an intelligent selection of sample points can
increase the accuracy of surrogate models.
2
Surrogate
3. 3
Sampling Based on Probability Distribution
• Observations of inputs often follow a distribution.
• A set of sample points representative of the naturally
occurring distribution of inputs is often desirable.
Distribution of a population Sample points
-20
0
20
-20
0
6
4
2
0
20
x 10-3
x2 x1
PDF
x1
x2
20
10
0
-10
-20
-20 -10 0 10 20
4. Presentation Outline
4
Research Objectives and Motivation
Probability-based sampling methods overview
Inverse transform sampling
Surrogate model development
• Surrogate model performance comparison
• Performance in increasing sample space
Concluding remarks
5. Certain inputs occur more frequently, comprising regions of
high interests in the condition space.
It is desirable to have higher accuracy in the system response
(surrogate) in the regions of higher interest.
5
Motivation and Research Objectives
Motivation:
Objectives:
Develop a sampling strategy for surrogate model
development, which promotes higher accuracy in regions of
high interest (of the observed input).
7. 7
Inverse Transform Sampling: Key Features
Inverse transform can
• Sample more points in the regions where random variables
have higher probability densities; and
• Sample fewer points in the regions where random variables
have low probability densities.
The probability of random variables is used as the metric of
distance instead of the Euclidean distance.
Sample points are uniform in terms of the probability
differences.
8. 8
Procedure: Step 1
Random Variable Observations
Distribution Function Fitting
Generating the Sequence of CDFs
Coordinates Evaluation
Step 1
Step 2
Step 3
Step 4
The occurrence of sampling
variables should be sufficiently
observed.
10
5
0
-5
-10
-15
-20
-5 0 5 10 15
x1
x2
9. 9
Procedure : Step 2
Approaches
• The least squares method
• The least absolute deviations method
• The generalized method of moments
• The Maximum Likelihood Estimation
Random Variable Observations
Distribution Function Fitting
Generating the Sequence of CDFs
Coordinates Evaluation
Step 1
Step 2
Step 3
Step 4
-20
0
20
-20
0
0.015
0.01
0.005
0
20
x2 x1
PDF
10. 10
Procedure : Step 3
• CDF increases from 0 to 1.
• Low-discrepancy sampling methods
generate uniformly distributed
sequences between 0 and 1 in all
dimensions of a sample space.
• Van der Corput sequence
• Halton/Hammersley sequence
• Sobol sequence
• Faure sequence
Random Variable Observations
Distribution Function Fitting
Generate the Sequence of CDFs
Coordinates Evaluation
Step 1
Step 2
Step 3
Step 4
1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
CDF(x1)
CDF(x2)
11. 11
Procedure : Step 4
• Coordinates are evaluated using the
inverse function of CDF.
• Analytical expressions
• Numerical approaches
• The Newton’s method
• The Levenberg-Marquardt algorithm
• The trust region methods
Random Variable Observations
Distribution Function Fitting
Generating the Sequence of CDFs
Coordinates Evaluation
Step 1
Step 2
Step 3
Step 4
x1
x2
20
10
0
-10
-20
-20 -10 0 10 20
13. 13
Window Performance Evaluation
• The heat transfer rate through a triple pane window is
evaluated under varying climatic conditions.
• A CFD model of the triple pane window is created.
• Sample climatic conditions are boundary conditions of the
window CFD model.
Cross Section
14. 14
Window Performance Evaluation
Step 1 Random Variable Observations
Three climatic conditions:
• Air temperature
• Wind speed
• Solar radiation
.
Michigan, ND.
3720 hourly observations for either
January or August from 2006 to 2010
15. 15
Window Performance Evaluation
Step 2 Distribution Function Fitting
Three climatic conditions:
• Air temperature: Gaussian
• Wind speed: Weibull
• Solar radiation: Gamma
Parameters are fitted using the Maximum Likelihood Estimation.
Michigan, ND.
3720 hourly observations for either
January or August from 2006 to 2010
16. Three climatic conditions:
• Air temperature: Gaussian
• Wind speed: Weibull
• Solar radiation: Gamma
Parameters are fitted using the Maximum Likelihood Estimation.
Michigan, ND.
3720 hourly observations for either
January or August from 2006 to 2010
16
Window Performance Evaluation
Step 3 Generate the Sequence of CDFs
1
0.8
0.6
0.4
0.2
0
270 280 290 300 310
Temperature
CDF
1
0.8
0.6
0.4
0.2
0
0 5 10 15
Wind Speed
CDF
1
0.8
0.6
0.4
0.2
0
0 500 1000 1500
Solar radiation
CDF
Sobol sequence
17. Three climatic conditions:
• Air temperature: Gaussian
• Wind speed: Weibull
• Solar radiation: Gamma
Parameters are fitted using the Maximum Likelihood Estimation.
Michigan, ND.
3720 hourly observations for either
January or August from 2006 to 2010
17
Window Performance Evaluation
Step 4 Coordinates Evaluation
1
0.8
0.6
0.4
0.2
0
270 280 290 300 310
Temperature
CDF
1
0.8
0.6
0.4
0.2
0
0 5 10 15
Wind Speed
CDF
1
0.8
0.6
0.4
0.2
0
0 500 1000 1500
Solar radiation
CDF
18. 18
Distribution of Sample Points
• Sample climatic conditions for January
• Sample climatic conditions for August
Sample points crowd in the region where PDF is high.
19. 19
Surrogate Model Development
• The heat transfer rate through the window is evaluated using
31 sample climatic conditions for either January or August,
respectively.
• Two surrogate models are developed for January and August
using Kriging, respectively.
Outdoor temperature
Wind speed
Solar radiation
Heat flux
Kriging
Inputs
Output
In this paper, we use a Matlab Kriging
toolbox DACE (Design and Analysis
of Computer Experiments), developed
by Dr. Nielsen.
20. 20
Surrogate Model Performance Criteria
For January and August, 3720 climatic conditions are used to
evaluate errors of each surrogate.
The performance of the surrogate can be evaluated using:
• Root Mean Squared Error (RMSE)
• Root Mean Squared Percentage Error (RMSPE)
• Maximum Absolute Error (MAE)
• Maximum Percentage Error (MPE)
21. 21
Surrogate Model Performance Comparison
Month Method RMSE MAE RMSPE MPE
January Inverse 0.047 0.49 0.64% 7.2%
Sobol 0.054 0.30 0.68% 9.3%
August Inverse 0.079 0.54 11% 318%
Sobol 0.094 0.32 85% 4373%
• RMSE, RMSPE, and MPE: Inverse transform sampling
performs better than the Sobol sequence.
• MAE: Inverse transform sampling has a larger MAE values.
22. 22
Performance in Increasing Sample Space
• All the hourly climatic conditions are classified into regions
with increasing PDF values in the sample space.
• The performance of the surrogate models is evaluated in
increasing sample space.
280
285
290
295
300
305
2
4
6
8
800
600
400
200
0
10
Wind speed (m/s) Temperature (K)
Solar radiation (W/m2)
100%
…
…
0.8%
0.1%
3720 climatic conditions
23. Root Mean Squared Percentage Error
23
Performance in Increasing Sample Space
The surrogate model for January
Root Mean Squared Error
Increasing percentage of sample space
Increasing percentage of sample space
24. The surrogate model for January
Maximum Percentage Error
24
Performance in Increasing Sample Space
Maximum Absolute Error
Increasing percentage of sample space
Increasing percentage of sample space
25. Root Mean Squared Percentage Error
25
Performance in Increasing Sample Space
The surrogate model for August
Root Mean Squared Error
Increasing percentage of sample space
Increasing percentage of sample space
26. The surrogate model for August
Maximum Percentage Error
26
Performance in Increasing Sample Space
Maximum Absolute Error
Increasing percentage of sample space
Increasing percentage of sample space
27. 27
Conclusions
• Inverse transform sampling is uniquely helpful for surrogate
development where the system inputs follow a certain distribution.
• The CDF of the inputs are made to follow a pseudorandom
sequence (such as Sobol).
• For window performance evaluation, the surrogate models
developed using inverse transform sampling have lower root mean
squared error than those developed using the Sobol sequence.
• For window performance evaluation, the surrogate models
developed using inverse transform sampling have higher maximum
absolute error than those developed using the Sobol sequence.
28. 28
Future Work
• Extend the applicability of inverse transform sampling to
correlated multi-variate/multi-input systems.
29. Acknowledgement
• I would like to acknowledge my research adviser
Prof. Achille Messac, for his immense help and
support in this research.
• Support from the NSF Awards is also
acknowledged.
29
30. 30
Selected References
• Husslage, B. G., Rennen, G., van Dam, E. R., and den Hertog, D., “Space-filling Latin Hypercube Designs for Computer
Experiments,” Optimization and Engineering, Vol. 12, 2011, pp. 611–632.
• Clarkson, K. L. and Shor, P. W., “Applications of Random Sampling in Computational Geometry, II,” Discrete and Computational
Geometry, Vol. 4, 1989, pp. 387–421.
• Goldreich, O., Computational Complexity: A Conceptual Perspective, Cambridge University Press, 1st ed., 2008.
• LaValle, S. M., Planning Algorithms, Cambridge University Press, 2006.
• Niederreiter, H., “Point Sets and Sequences with Small Discrepancy,” Monatshefte fr Mathematik, Vol. 104, December 1987, pp.
273–337.
• van der Corput, J. G., “Verteilungsfunktionen,” Nederl. Akad. Wetensch. Proc., Vol. 38, 1935, pp. 813–821.
• Diaconis, P., “The Distribution of Leading Digits and Uniform Distribution Mod 1,” The Annals of Probability, Vol. 5, No. 1, Feb
1977, pp. 72–81.
• Sobol, I. M., “Uniformly Distributed Sequences with an Additional Uniform Property,” USSR Computational Mathematics and
Mathematical Physics, Vol. 16, 1976, pp. 236–242.
• Faure, H., “Discrpances de suites associes un systme de numration en dimension s,” Acta Arithmetica, Vol. 41, 1982, pp. 337–351.
• Miller, F., Vandome, A., and John, M., Inverse Transform Sampling, VDM Verlag Dr. Mueller e.K., 2010.
• von Neumann, J., “Various Techniques Used in Connection with Random Digits,” Nat. Bureau Stand. Appl. Math. Ser., Vol. 12,
1951, pp. 3638.
• Marshall, A. W., “The Use of Multi-stage Sampling Schemes in Monte Carlo Computations,” H. A. Meyer (ed.), Symposium on
Monte Carlo Methods, edited by N. Y. John Wiley & Sons, Inc., 1956, p. 123140.
• Gilks, W., Gilks, W., Richardson, S., and Spiegelhalter, D., Markov Chain Monte Carlo in Practice, Interdisciplinary Statistics,
Chapman & Hall, 1996.
31. 31
Performance in Increasing Sample Space
• All the hourly climatic conditions are classified into regions
with increasing PDF values in the sample space.
• For each variable, the probability is the integral of the fitted
PDF in the shortest interval.
• The performance of the surrogate models is evaluated in
increasing sample space.
32. 32
Review
• Sampling sequences
• Latin hypercube
• Random
• Pseudorandom
• Low-dispersion
• Low-discrepancy
• Generating sample points from a probability distribution
• Inverse transform sampling
• rejection sampling
• importance sampling
• Markov Chain Monte Carlo
• Metropolis-Hastings Sampling
• Gibbs Sampling
33. Comparisons and Analyses
33
Sobol sequence Inverse transform
• A Voronoi diagram is a special kind of decomposition of a metric space
determined by distances to a specified discrete set of points in the space.
• Each point has a cell that includes the region closer to the point than to
any others.
• The lines are equidistant to the two nearest points.
Hinweis der Redaktion
Inverse transform sampling was applied to a unimodal distribution to sample 31 points in previous slides to show how the sampling approach works.
It can also be used to sample more points. The figure on the left shows that, when the number of sample points is increased from 31 to 127, the sample points obtained from this approach still crowd in the region where the probability density is high. Although the number of points in the regions with low probability density also increase, it is increasing at a lower rate than that of points in the regions with high probability density.
Inverse transform sampling can also be used to sample multi-modal distributions. The two figures show the sample points for a bimodal distribution and a quad-modal distribution. The sample points crowd in the regions where the probability density is high.
Higher MAE: It is because in the region with very low distribution density, the sample points are far from each other. The absolute errors at some points in this region are higher for the inverse than the sobol.
August high MPE at 4373%: At some points, their values are very close to zero. These values are used as denominators.
For the sobol sequence, Root Mean Squared Error and Root Mean Squared Percentage Error are similar in different percentages of sample space.
For the inverse transform sampling, Root Mean Squared Error and Root Mean Squared Percentage Error are increasing as the percentage increases. Their overall performance is still better than the sobol sequence.
As the percentage of sample space increases, maximum absolute error and maximum percentage error are both increasing.
When the whole space is reached, the density of sample points obtained by inverse transform sampling in some space becomes lower than those for sobol. The maximum absolute error for inverse becomes higher than that of sobol. However, the maximum percentage error for inverse is still lower than that for sobol
The spike for the root mean squared percentage error is because in the 12.5% region but not in the 10.0% region, the actual heat flux through the window is very close to zero. It is used as the denominator for percentage error evaluation.
The spike for the maximum percentage error is because in the 12.5% region but not in the 10.0% region, the actual heat flux through the window is very close to zero. It is used as the denominator for percentage error evaluation.