Is it possible to have an aggregate function instead of the active binding?

[schalkdaniel]

Wouldn't it be nice to have a function

bmr$aggregate(params = TRUE)

which returns not just the aggregated benchmark object but also the parameter of a single learner? I have written a wrapper in tuning, but it feels a bit misplaced there.

[mllg]

Looking at your code is on my TODO (e.g. https://github.com/mlr-org/mlr3tuning/pull/29/files#diff-8d41897158e70fbb081fd674a26bffd0R80).

I'll get back to you tomorrow.

[mllg]

This is now a function, as requested. The interface might change again though, depending on #155.

[mllg]

Actually, it might be better to provide the parameters as list column and let the user unnest them. @schalkdaniel What do you think?

[schalkdaniel]

I think the data.table structure with R6 class columns is complicated enough. Unnesting the parameters manually requires to know mlr3misc and from there the unnest function. Providing some help here is ok (in my opinion).

On the other hand, tuning over a large amount of parameter (as it is for pipelines?) can bloat the aggregated data table.

So, maybe returning the parameters as list column by default, but giving the possibility to unnest them automatically?

[mllg]

We have a no sugar agreement. You get the list column for now. 😉

I've added an example to demonstrate how to access the params.