2D: use the same loss for choosing points and loss_improvement

This seems to work really badly:

Screen_Shot_2017-10-04_at_12.46.36

Merge request reports

Loading