#71 New function to compare posterior density to observations

[STROH]

I propose a function to compute the Continous Ranked Probability Score (CRPS), which compares a predictive density to an observation, in the case where the predictive density is Gaussian.

This function can be combined with stk_predict_leaveoneout. However, I do not know where store the file in STK.

[Julien Bect]

We already have several functions related to the Gaussian distribution: they are all named stk_distrib_normal_XXX.m and located in misc/distrib.

So, I suggest renaming this function to stk_distrib_normal_crps.m and putting it in misc/distrib as well.

Concerning author names : you don't have to add Emmanuel and I to each file that you publish, unless you really copied a significant portion of some other file that Emmanuel and/or I wrote.

Please try to follow more consistently the formatting convention of the STK toolbox. In particular, use one space between a function name and the opening parenthesis that follows it. Also, in the help text, the header should not exceed one line; function arguments are in capitals; etc.

You can prepare a patch for default with these indications. Don't forget to fill the ChangeLog, and also to update NEWS and INDEX.

@++

[STROH]

The patch that takes into account the suggestions above. Modifies ChangeLog, News and INDEX.
Also improves the functionalities of the function (can consider cases with noise).

Concerning author names: in this particular case, the script of the function is based on the script of stk_distrib_normal_ei.m, so I added the authors of the EI function to the list of authors.

[Julien Bect]

I have pushed your patch and made a few changes afterwards.

In particular, I have removed the fourth (optional) argument, which didn't make much sense for an stk_distrib_normal_* function.

I think we should compute a bunch of reference values, using for instance the 'verification' package for R, and turn them into unit tests.

Also, the is a line that I don't understand in the unit tests:

%! c = 2 / (sqrt (pi)) * (sqrt ((sigma .^ 2 + noistd .^ 2) / 2) - (sigma + noistd) / 2);

Is this supposed to be a test?

[Julien Bect]

I have removed the example from the help text. Read the ChangeLog to know why:
https://sourceforge.net/p/kriging/hg/ci/f4bbf8c2ce44ee78fc62a2088204b9cb1eaa2192/