dinver for arbitrary problems

[mao]

One question which might be of more borad intrest to others, too.
When using dinver with one's own forward problem, then it is possible to
use the plugins as you describe in the documentation.

What, if we would just like to use the importance sampling starting
from an already computed set of models (parameters+misfits) - how can one
fake a run report in order to make it joint to the importance sampling?

bye
Matthias

[admin]

Hi Matthias,

I'll try to run an example to answer... give me some more time. It will be also a good opportunity to thoroughly debug the importance sampling

[freitag]

Hi Marc and Mathias,

I couldn't find any documentation or forum entries on the importance sampling apart from this one. Does it the same thing that Malcolm Sambridge does with his NA_bayes code (evaluating bayesian integrals using the Neighbourhood algorithm)? If so, how are the results stored in the report files and how could I access them?

Cheers,
Yannik

[admin]

Dear Yannik,

Yes it is the same as NA_bayes from Malcom Sambridge. The only difference is its ability to generate samples in a conditioned parameter space as described in Wathelet, 2008.

Models stored in .report files can be accessed with gpdcreport

To get parameter values and misfit:

Code: Select all

gpdcreport run.report -pm

To get 1D ground models:

Code: Select all

gpdcreport run.report -gm

Models produced this way can be piped through gpprofile or any other program to analyze or extract statistics.

Note that the importance sampling code is still experimental. For instance the definition of PPD function is based on misfit values with a fixed log expression. If you want to change it you must modify the code and re-compile.

[freitag]

Thanks Marc