A comparison of global sensitivity techniques and sampling method

Inspired by Tarantola et al. (2012), we extend their analysis to include the Latin hypercube and Random sampling methods. In their paper, they compared Sobol’ quasi-Monte Carlo and Latin supercube sampling methods by using a V-function and variance-based sensitivity analysis. In our case we compare the convergence rate and average error between Sobol’, Latin hypercube, and Random sampling methods from the Chaospy library, keeping everything else the same as in their paper. We added the Random sampling method to test if the other two sampling methods are indeed superior. The results from our code confirm the results of their paper, where Sobol’ has better performance than Latin hypercube sampling in most cases, whilst they both have higher efficiency than is achieved with Random sampling. In addition we compared the explicit forms of ‘Jansen 1999’ total effects estimator used in Tarantola et al. (2012) with the ‘Sobol’ 2007’ estimator, again keeping sample sizes and the test function the same. Results confirm that the ‘Jansen 1999’ estimator is more efficient than ‘Sobol’ 2007’. The presentation will also include the Morris sampling method and other test functions to further test efficiency among all the sampling methods on different cases.