Randomvars

[jpn--]

This is a start of what I am thinking about to address #46, allowing arbitrary distributions. I have not written any unit tests for it yet, nor solved anything about unbounded distributions discussed in #46, however it does pass all existing unit tests and does not break anything as far as I can see.

[coveralls]

Coverage decreased (-0.1%) to 73.735% when pulling 69ea09f on jpn--:randomvars into e83946d on quaquel:master.

[quaquel]

I have started in a new 2.1 branch on integrating this idea. However, after thinking about it at length, I have decided to use a factory method from_stats to support alternative distributions. The reason for this is twofold: (1) in exploratory modelling using a uniform distribution is the default option, I want to keep it this way; (2) alternative distributions can be supported but at the api level, I want it to look clearly different, using from_stats achieves this.

At present, the parameters.py file is finished. I have implemented the functionality in a slightly different way that you did (using various python tricks to keep the code clean etc.)

Next up is editing the samplers. Here I want to do a few things:

• add an LHS sampler that uses upper and lower bounds and defaults to uniform (as you have)
• add a sampler that generates an LHS over the deep uncertainties, as well as an LHS over the well characterised uncertainties, and then creates a full factorial over both sets of samples. In my personal view, this is the most defendable approach for handling deep and well characterised uncertainties simultaneously.
• update various utility functions (like loading parameters from a csv, saving to a csv, the repr magic method)

Do you have any further thoughts or requests in this space?

[jpn--]

This sounds reasonable. You may want to add a flag on uncertainties to mark if they are deep or well-characterized, as that won't be identifiable from the distribution alone (well characterized uncertainties can have uniform distributions).

[quaquel]

I have implemented all this drawing from your code in the 2.1 branch