BAT.jl bat_marginalize should come with an algorithm type and defaults

bat_marginalize should always return a NamedTuple (result = ...,), to leave room for additional, algorithm-specific output values. This is something like a standard for all bat_....() functions now. Sorry, forgot to mention that on the PR that added bat_marginalize.

CC @Cornelius-G, @VasylHafych

Jun 09 '20 10:06 oschulz

I guess this issue can also be closed as it was fixed with PR #133

Jun 23 '20 11:06 Cornelius-G

Oh, yes. But we have the more rigorous API for bat_functions now, so I'll change the title of this issue. :-)

Jun 23 '20 13:06 oschulz

So we should add the concept of marginalization to algotypes and algodefaults.

Jun 23 '20 13:06 oschulz

Ah, ok. I will take a look at this. We also wanted to think about the general structure of the marginalizations again. So currently they have the original valueshape and the indices of the included parameters. But I guess we wanted to have them only the valueshape of the contained params, right?

Jun 23 '20 14:06 Cornelius-G

Hm, from a purist point of view, they would not include the indices of the original parameters. Do we need them?

Actually, I've been thinking about adding a close relative to ValueShapes.jl that allows assigning a varshape to any generic multivariate Distribution. With that, we wouldn't even need a special data structure for marginalizations in BAT itself, it would just be a combinations of ValueShapes and EmpiricalDistributions.

Jun 23 '20 16:06 oschulz

Ah, what should we now do about this issue?

Hm, from a purist point of view, they would not include the indices of the original parameters. Do we need them?

I will check again if we need the original shapes for a correct matching of names to indices in plotting, but I think we actually would not need them for this. If not needed, should we then make MarginalDist contain only the distribution and a subset of the original shape?

Or will you take a look into this proposed

close relative to ValueShapes.jl that allows assigning a varshape to any generic multivariate Distribution.

?

Jul 29 '20 22:07 Cornelius-G

think we actually would not need them for this.

Then let's drop them, it'll be much cleaner without.

If not needed, should we then make MarginalDist contain only the distribution and a subset of the original shape

Yes - I'm thinking about adding a generic "add NamedTupleShape to any dist" wrapper to ValueShapes - that would then make it possible to replace MarginalDist by that wrapper combined EmpiricalDistributions.

Jul 30 '20 07:07 oschulz