Create parameter_names and sampled_parameter_values attributes for LCA about brightway2-calc HOT 5 CLOSED

Original comment by Chris Mutel (Bitbucket: cmutel, GitHub: cmutel).

Can you think about when we would want common indexing, and when we wouldn't? I am not sure.

By common indexing, I mean that we create one RNG that creates an index, and we select that column from every presample package array. I could imagine that we would want every presample package to have the same index, but I can equally well imagine that we wouldn't want this behaviour if presample packages should be considered independent. We can stack arrays in a single package relatively easily, so we could go with option two.

Another alternative would be to use a consistent sampler only if we were inside a campaign.

from brightway2-calc.

Original comment by Pascal Lesage (Bitbucket: MPa, ).

Common indexing occurs when the same  seed is passed when creating the presamples package. If common indexing is required, the same seed value is passed. If different indexing is needed, different seed values is passed. If it doesn't matter, then the default (random) seed value is used.

from brightway2-calc.

Original comment by Chris Mutel (Bitbucket: cmutel, GitHub: cmutel).

OK, now you have opened another door of confusion :)

We know that:

1. IrregularPresamplesArray is a subclass of RandomState, and only draws from the RNG to get an index.
2. MatrixPresamples creates a seeded IrregularPresamplesArray for each presamples package.
3. LCA (and its subclasses) create one instance of MatrixPresamples if needed on instantiation.

BTW, we might want to change the name of IrregularPresamplesArray, as we now ensure that presamples packages are regular.

So, I think my concern was unfounded; we get consistent sampling within presamples packages automatically with IrregularPresamplesArray, regardless of whether the RNG was seeded; we get consistent sampling across presample packages if a seed value is passed. Obviously, this should be documented and thoroughly tested.

"Common indexing occurs when the same seed is passed when creating the presamples package." I am not sure this is true, but probably we have to be more precise about what we mean here; for example, as the values from the first package be used to generate values from the second? If not, I don't think the seed matters; for example, using the same seed for two different versions of ecoinvent wouldn't lead to any correlation.

from brightway2-calc.

Original comment by Pascal Lesage (Bitbucket: MPa, ).

I think the simplest and quite logical solution is this: Seed values from the LCA object should not be passed to matrix_presamples in the LCA (here). Presamples seeds should be specific to the presamples package, and should not be overwritten during use of presamples.

As to "Common indexing occurs when the same seed is passed when creating the presamples package." This should be true for presamples with the same number of columns. This is ensured for parameter_presamples and matrix_presamples jointly created.

from brightway2-calc.

Original comment by Chris Mutel (Bitbucket: cmutel, GitHub: cmutel).

Fixed in upstream presamples code. Can access presamples.parameters to get all requested data.

from brightway2-calc.