One nice thing to do would be to have initialization heuristics, i.e. the ability to initialize the probabilities of using independent variables according to a lasso regression or using something like that.
Should go into a canonical initialization stage of the algorithm that uses all possible data transformations. At the end of the stage, eliminate variables with negligible probabilities of being involved.
One nice thing to do would be to have initialization heuristics, i.e. the ability to initialize the probabilities of using independent variables according to a lasso regression or using something like that.