Closed tadhg-moore closed 4 years ago
But are there any parameters that really have a uniform distribution, as I would reckon we don't know the real PDF of any parameter. Wouldn't then uniformly sampling be slow by assuming that every point in the range has the same probability?
I noticed it when I used LHS on a logarithmic parameter and it didn't sample the parameter space thoroughly esepcially where the parameter was had a better fitness. See the plot where I did LHC on the full space but didn't account for the log but using a uniform distribution (red square) I was able to sample the logarithmic space thoroughly. And the idea of this is to provide information on the parameter space before a more targeted calibration tool would be used.
This has not been implemented yet, but because it is more of an enhancement than an issue, we decided to move this to the "Enhancements" project in the Project Tracker, and close the issue.
An alternative to Latin hypercube sampling would be to uniformly sample within the ranges. This would be particularly useful when there are logarithmic parameters e.g. k_min to ensure that the logarithmic distribution is sampled effectively.
Useful function for this here: https://rdrr.io/cran/KScorrect/man/dlunif.html