ja-thomas

Open Questions

1

- Definition of the paramater space (ranges) - How large to set the validation set - How large should early_stopping_limit be

Fix Autoxgboost reader issue

3

Now merging in the correct branch. This might fix the CI problems.

``` task = convertOMLTaskToMlr(getOMLTask(2073)) Error in instantiateResampleInstance.CVDesc(desc, length(ci), task) : Cannot use more folds (10) than size (5)! ``` The mlr task: ``` Browse[2]> mlr.task Supervised task: yeast Type: classif...

(Crosspost from https://github.com/mlr-org/mlrCPO/issues/49) CPOs don't work with runTaskMlr (see link for example) Not sure if this has to be solved in mlrCPO or in Openml

In mboostLSS_fit() all offset values of the distribution parameters (mu, sigma, ...) are written in each distribution parameter model. e.g. The fit object of the mu parameter has a variable...

``` > cb3 = cb$clone(deep = TRUE) Error in envRefInferField(x, what, getClass(class(x)), selfEnv) : '.__enclos_env__' is not a valid field or method name for reference class "Rcpp_BlearnerFactoryList" ```

should be either caught or (preferably) handled correctly as a pure intercept model

I think another reason to use a data abstraction layer. But we want to have the ability to do random subsampling in every iteration.

BaseLearnerFactory will always copy data

1

In some cases e.g. splines we need to actually create new data but for simple base learner e.g. linear, a pointer to the data column would be enough

There is the package [DiceDesign](https://cran.r-project.org/web/packages/DiceDesign/index.html) which contains a large number of possible ways to create initial designs. May be worth a look