Closed heidmic closed 4 days ago
Yes! More flat parameters would definitely help with things like parameter tuning. E.g. I would like to be able to do
SupRB(rd_delay=10, sc_selection_type="Tournament", sc_selection_params={"k" : 2})
or even
SupRB(rd_delay=10, sc_selection_type="Tournament", sc_selection_param_k=2)
I guess, however, that both forms would require some minimal form of parameter validation, though, especially since some parameters are conditional.
This should probably (in my opinion) be optional, though; if someone wants to create objects and supply them as arguments, they should be able to do so as well.
Note that there is something faintly similar to this parameter name chaining (using sklearn.utils.Bunch
etc.) in suprb-experimentation already which could maybe be adjusted and incorporated into suprb itself.
Currently, you need to generate objects from deep within suprb to set some parameters. That is fine when interacting with suprb-experimentation but should be different for a standalone solution