Update variance estimation methods for RF learners

[ja-thomas]

• Extratrees is now an available splitting method for classification and regression

• Add se estimation method for regr.ranger. (This is mostly copy paste code from regr.randomForest) and only one of the three alternatives there.

[berndbischl]

isnt it a bit inconsitent to add only 1 se method from rf...?

yes, it is a clear improvement. can you easily convince me that we should not copy the other 2 methods?

[ja-thomas]

yup you're right, all three methods would make sense.

But I just needed one for now to test some stuff for MBO. But in general it would make sense if the se estimations for random forest would work for all supported RFs, but thats more work.

So the short answer is I was too lazy to copy all three methods... And this is mostly a PR so that the code/branch does not get lost

[jakob-r]

We should really have extra methods for se calculation for bagged learners...

[SteveBronder]

What else needs done here? May have time next week to help if you need any

[jakob-r]

@berndbischl We should merge and be okay with this double code because calculating the SE for bagged learners in general would be a seperate issue. This should not hold back such important progress for the usage of ranger.
In other words: The benefit of being able to calculate SE for ranger outweighs the disadvantage of the little code clutter in the inside. Especially as the latter can be easily fixed later.

[berndbischl]

i really dont want the package to become inconsistent here.
if people want to have this so urgently, please help making this better.

it is really not that hard here.

[jakob-r]

Okay. I added generalized functions for sd and jackknife. We might want to put them in a separate file. For a better diff i will do that later (name suggestions are welcome).

[jakob-r]

@berndbischl your opinion would be much appreciated 😉

[larskotthoff]

I've tried to sort out the merge conflicts. Could someone please have a look whether this still makes sense? @jakob-r

[jakob-r]

I solved the conflict. If travis passes this can be merged.

[larskotthoff]

Ok, I've fixed some obvious stuff, but Travis is still failing and have no idea why. Can't reproduce locally.

[pat-s]

@ja-thomas just FYI, I merged the master branch. Let's see if the build passes. Do you still think that this is ready to merge then?

(+ an entry to NEWS.md)

[pat-s]

@jakob-r
For some reason the par.vals list in

mlr/tests/testthat/test_regr_ranger.R

Line 31 in c89a3c2

learner = makeLearner("regr.ranger", predict.type = "se", par.vals = par.vals)

is not present during the train() call in

mlr/R/RLearner_regr_ranger.R

Line 41 in c89a3c2

tn = getTaskTargetNames(.task)

That is what causes Travis to fail.

Do you know if this is a general problem of par.vals not being forwarded?

[jakob-r]

ranger now uses only its own se methods (jacknife and infjacknife (default)).
Normal rf should work again.
all necessary parameters are set automatically if predict.type = "se" is demanded.

From my side this should be mergable if no errors occur (already checked locally)

[pat-s]

Cool 👍
The doc still needs to be transformed to markdown and we a NEWS entry, then we're good to go here.