Kodi Arfer

> Implementing the slope for pseudo-Huber What does this mean? By the slope, do you mean the gradient? It looks like the gradient is already defined properly for this objective.

By this way, this kind of investigation may be easier with the somewhat simpler log-cosh objective function, with gradient `tanh(x)` and Hessian `1/cosh(x)^2`.

Oh, I see. Yeah, I was thinking that the δ value should be documented in the description of `reg:pseudohubererror`, and perhaps customizable.

I don't think that fixes the problem. Changing the slope just makes pseudo-Huber less like absolute error.

Oh, very cool, I'd hadn't heard. I'll have to look into it.

@trivialfis I'm getting poor fits with the new `reg:absoluteerror` and XGBoost 18a38f7ca0ee7a60c03632a560419c4f507432ae in simple examples like the ones I've tried before with psuedo-Huber and log-cosh; see below for a complete...

Thanks, that seems to work well in this case. Do you recommend `hist` over `approx` in most cases? > Automate the `base_score=median(y)` in your example Intelligent selection of the default...

I've also noticed that my example problem has trouble with DART, even with `tree_method = "hist"`, at least if the RNG is re-seeded in the inner loop: ```R rmse =...

@trivialfis I was excited to use `reg:absoluteerror` in some of my group's projects, but I again have performance problems: no longer catastrophically bad, but still worse than can be achieved...

There may be considerable value in just checking whether a column of doubles has only a few unique values before encoding it with the generic doubles compressor. I find that...