Morris method normal distribution sample

[MatthiVH]

Hi,

I was wondering how to create a sample with a normal distribution using the Morris sampler.
Since 'dists' is not yet supported in the morris sampler, I tried a workaround (tested with the saltelli-sampler that has 'dists' and it worked), but I receive lots of infinite-values in the sample with the morris method when converting uniform to normal distributions. Is there a solution for this?

Tried code:

from SALib.sample import morris

import pandas as pd

import scipy.stats.distributions as ssd

import lhsmdu



# problem definition ('num_vars', 'names', 'bounds', 'dists')

prob_dists_code = {'num_vars': 3,
                      
                      'names': ['P1', 'P2', 'P3'],

                     'groups': None,

                     'bounds': [[0.0, 1.0], [0.0, 1.0], [0.0, 1.0]]}

np_array = morris.sample(prob_dists_code, N=1000, num_levels=4, optimal_trajectories=None)

p1 = ssd.uniform(loc=0.9, scale=0.2)
p2 = ssd.norm(loc=45, scale=10)
p3 = ssd.norm(loc=0, scale=1)

new_samples = np_array

new_samples[:, 0] = lhsmdu.inverseTransformSample(p1, np_array[:, 0])

new_samples[:, 1] = lhsmdu.inverseTransformSample(p2, np_array[:, 1])

new_samples[:, 2] = lhsmdu.inverseTransformSample(p3, np_array[:, 2])


df = pd.DataFrame(new_samples)

df.columns = ['P1', 'P2', 'P3']

print(df)

Kind regards,
Matthias

[ConnectedSystems]

Hi @MatthiVH

Could you try the most recent development release and let me know if this is still an issue?

Recommend doing so in a fresh conda or virtual environment

Install with:

pip install --pre salib

[MatthiVH]

Hi,
Yes, I'll report back by the beginning of next week.
Kind regards,
Matthias

[MatthiVH]

Hi @ConnectedSystems,

Yes it is still an issue with the latest release. The 'dists'-statement is not recognized in the 'problem' and if I want to post-process the uniform samples towards the distribution I want, lots of infinity values show up (which does happen when using sobol, fast, ...

Kind regards,
Matthias

[ConnectedSystems]

Hi again,

Looking into this deeper, the Morris method does not, as originally conceived, support alternate distributions. The trajectory sampling employed assumes an uniform distribution of the input space, and scaling the sampled points to other distributions don't always make sense (as you have encountered).

There are improvements to Morris to allow its use with alternate input distributions (see for example, [1]), but we currently have no plan in the short-term to implement this.

[1] Feng, K., Lu, Z., Yang, C., 2019. Enhanced Morris method for global sensitivity analysis: good proxy of Sobol’ index. Struct Multidisc Optim 59, 373–387. https://doi.org/10.1007/s00158-018-2071-7