deprecate Learner2D

Once LearnerND becomes good enough we should remove Learner2D, as it will no longer be needed.

If good enough == better than.

good enough == as good as?

mentioned in issue #84

I think good enough == better than the Learner2d.

• Because right now the learner2d takes all the points into account when computing the loss for each triangle. Although this is something we want to lose, this has some benefits, as you can for example, take the second derivative of the function into account (or as the default loss is currently implemented, taking the deviation from the linear interpolation + gradient)
• Also I am not sure if people have written custom loss functions, which are going to be different for the LearnerND, so it is not a drop-in replacement.

Point #1 (closed) is fair.

We don't care about backwards compatibility at this point, as we are < v1.0

I take it you didn't want to refer to issue 1?

In the future I would like to implement a way to maintain locality and still be able to take the second derivative into account (like taking the loss over a simplex and it's neighbours or something like that). But to me this sounds like a non-trivial task. However, once implemented, it could really improve the quality of the LearnerND

added Learner2D label

mentioned in merge request !87 (closed)

mentioned in merge request !132