Create pre-defined parameter grids for common models

Usage

tl_default_param_grid(method, size = "medium", is_classification = TRUE)

Arguments

method: Model method ("tree", "forest", "boost", "svm", "xgboost", etc.)
size: Grid size: "small", "medium", "large"
is_classification: Whether the grid is for a classification task (the default) or a regression one. For regression, an "svm" grid also tunes epsilon, the width of the band within which e1071's regression SVM ignores errors, and the large "forest" grid centres nodesize on randomForest's regression default of 5 rather than its classification default of 1. The other grids are the same for both tasks.

Value

A named list of parameter values suitable for passing to tl_tune_grid or tl_tune_random. Each element is a numeric or character vector of candidate values for that hyperparameter, or for "deep"'s hidden_layers a list of layer-size vectors. The grid is built without the data, so a "forest" mtry can exceed the number of predictors; the tuners cap it. "polynomial" tunes degree. The "xgboost" grids draw on the values tl_tune_xgboost searches by default, and add nrounds, which that function chooses by early stopping and tl_tune_grid has to tune. "linear" and "logistic" have no tuneable hyperparameter and return an empty list with a warning, as does an unknown method.

Examples

# \donttest{
grid <- tl_default_param_grid("tree", size = "small")
grid <- tl_default_param_grid("forest", size = "medium")
grid <- tl_default_param_grid("svm", is_classification = FALSE)
# }