should a learner have a model as a state member variable?

[mlr-org]

should a learner have a model as a state member variable?

[berndbischl]

and sklearn

[larskotthoff]

Sounds to me like having a workflow or similar object to combine learner and model would be better. It's not clear to me how a learner with state would fit into the current framework. For example, what happens when a learner with a model is wrapped? Model removed or automatically retrained? Same with changing hyperparameters for a learner.

[berndbischl]

we already decided this for the new framework. it makes the new API better.
you can see this when we document this a bit better.

short answer:

Sounds to me like having a workflow or similar object to combine learner and model would be better.

we do this. the object "Learner" now combines calls regarding training, predictions and model. this makes everything simpler and more natural.