Piecewise Nonlinear => PWL nudge

[michaelbynum]

Summary

@emma58 - This line has caused me some problems. If I remember correctly, this was preventing me from evaluating the PWL function at the bounds. I will go back and refresh my memory to make sure that was the only problem, but I wanted to see if you remember why this is needed. I've had better luck using 1e-10 instead of 1e-4. Tolerances like this are tricky, and I'm always hesitant to change them. Maybe we should expose this as an option? If so, what do you think of min_distance_to_bounds for a name?

[sadavis1]

This was my code... I'm pretty sure I did this because going all the way to the edges triggered a problem with the linear tree package. If it's not working around some other flaw, I don't see why this code should exist at all. I propose we should make it a flag to _get_uniform_point_grid, so it's used when we're going to use the linear tree but we just do things correctly otherwise. If that sounds good I can PR that change.

[michaelbynum]

Thanks, @sadavis1! That is helpful. I have not been using the linear tree functionality, which is probably why I have not had problems with smaller values. I think you plan is a good one.