10.06.2016 Views

eldo_user

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Monte Carlo Analysis<br />

Latin Hypercube Sampling<br />

The quasi-Monte Carlo efficiency relies on the empirical convergence rate of commonly used<br />

sequences. This convergence rate is governed by the discrepancy that is asymptotically given<br />

by:<br />

Asymptotically this is smaller than the standard error of a pseudorandom estimate, the error<br />

converges in O(n −1 ) instead of O(n −1/2 ) for traditional Monte Carlo. We then have a method<br />

that can potentially compete with MC to run the simulations. Unfortunately the convergence<br />

rate depends on the dimension d. For example, with d = 10, we have:<br />

for n ≤ 10 34<br />

We conclude that the current implementation of the QMC techniques should not be used as an<br />

option to accelerate the Monte Carlo analysis when only a moderate number of simulations is<br />

allowed and when the number of parameters is greater than 5.<br />

The estimators of the mean, yield, variance, and so on, are the same as in the Monte Carlo<br />

method. However, a disadvantage of this approach is that the use of a deterministic source<br />

means that the unbiased and the usual error estimation methods are lost. Therefore it is difficult<br />

to assess the final value of the integral (the mean value).<br />

For a detailed overview of QMC methods, refer to: P. L’Ecuyer and C. Lemieux, Recent<br />

Advances in Randomized Quasi-Monte Carlo Methods in Modeling Uncertainty: An<br />

Examination of Stochastic Theory, Methods and Applications, Kluwer, 2001.<br />

Related Topics<br />

Sampling Plan Methods<br />

Latin Hypercube Sampling<br />

Latin hypercube sampling (LHS) may be considered as a particular case of stratified sampling.<br />

The purpose of stratified sampling is to achieve a better coverage of the sample space of the<br />

input factors. In the latin hypercube, sampling is undertaken as follows:<br />

• The range of each input factor x j is divided into n intervals of equal marginal probability<br />

1/n.<br />

• One observation of each input factor is made in each interval using random sampling<br />

within that interval.<br />

Eldo® User's Manual, 15.3 457

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!