# 6 Extending

This chapter gives instructions on how to extend mlr3 and its extension packages with custom objects.

The approach is always the same:

1. determine the base class you want to inherit from,
2. extend the class with your custom functionality,
4. (optionally) add new object to the respective Dictionary.

The chapter Create a new learner illustrates the steps needed to create a custom learner in mlr3.

Here, we show how to create a custom mlr3learner step-by-step using mlr3extralearners::create_learner.

It is strongly recommended that you first open a learner request issue to discuss the learner you want to implement if you plan on creating a pull request to the mlr-org. This allows us to discuss the purpose and necessity of the learner before you start to put the real work in!

This section gives insights on how a mlr3learner is constructed and how to troubleshoot issues. See the Learner FAQ subsection for help.

Summary of steps for adding a new learner

1. Check the learner does not already exist here.
2. Fork, clone and load mlr3extralearners.
3. Run mlr3extralearners::create_learner.
4. Add the learner param_set.
5. Manually add .train and .predict private methods to the learner.
6. If applicable add importance and oob_error public methods to the learner.
7. If applicable add references to the learner.
8. Check unit tests and paramtests pass (these are automatically created).
9. Run cleaning functions
10. Open a pull request with the new learner template.

(Do not copy/paste the code shown in this section. Use the create_learner to start.)

### 6.1.1 Setting-up mlr3extralearners

In order to use the mlr3extralearners::create_learner function you must have a local copy of the mlr3extralearners repository and must specify the correct path to the package. To do so, follow these steps:

1. Fork the repository
2. Clone a local copy of your forked repository.

Then do one of:

We recommend the last option. It is also important that you are familiar with the three devtools commands:

### 6.1.2 Calling create_learner

The learner classif.rpart will be used as a running example throughout this section.

library("mlr3extralearners")
create_learner(
pkg = ".",
classname = "Rpart",
algorithm = "decision tree",
type = "classif",
key = "rpart",
package = "rpart",
caller = "rpart",
feature_types = c("logical", "integer", "numeric", "factor", "ordered"),
predict_types = c("response", "prob"),
properties = c("importance", "missings", "multiclass", "selected_features", "twoclass", "weights"),
references = TRUE,
gh_name = "RaphaelS1"
)

The full documentation for the function arguments is in mlr3extralearners::create_learner, in this example we are doing the following:

1. pkg = "." - Set the package root to the current directory (assumes mlr3extralearners already set as the working directory)
2. classname = "Rpart" - Set the R6 class name to LearnerClassifRpart (classif is below)
3. algorithm = "decision tree" - Create the title as “Classification Decision Tree Learner”, where “Classification” is determined automatically from type and “Learner” is added for all learners.
4. type = "classif" - Setting the learner as a classification learner, automatically filling the title, class name, id (“classif.rpart”) and task type.
5. key = "rpart" - Used with type to create the unique ID of the learner, classif.rpart.
6. package = "rpart" - Setting the package from which the learner is implemented, this fills in things like the training function (along with caller) and the man field.
7. caller = "rpart" - This tells the .train function, and the description which function is called to run the algorithm, with package this automatically fills rpart::rpart.
8. feature_types = c("logical", "integer", "numeric", "factor", "ordered") - Sets the type of features that can be handled by the learner. See meta information.
9. predict_types = c("response", "prob"), - Sets the possible prediction types as response (deterministic) and prob (probabilistic). See meta information.
10. properties = c("importance", "missings", "multiclass", "selected_features", "twoclass", "weights") - Sets the properties that are handled by the learner, by including "importance" a public method called importance will be created that must be manually filled. See meta information.
11. references = TRUE - Tells the template to add a “references” tag that must be filled manually.
12. gh_name = "RaphaelS1" - Fills the “author” tag with my GitHub handle, this is required as it identifies the maintainer of the learner.

The sections below demonstrate what happens after the function has been run and the files that are created.

### 6.1.3 learner_package_type_key.R

The first script to complete after running create_learner is the file with the form learner_package_type_key.R, in our case this will actually be learner_rpart_classif_rpart.key. This name must not be changed as triggering automated tests rely on a strict naming scheme. For our example, the resulting script looks like this:

#' @title Classification Decision Tree Learner
#' @author RaphaelS1
#' @name mlr_learners_classif.rpart
#'
#' @template class_learner
#' @templateVar id classif.rpart
#' @templateVar caller rpart
#'
#' @references
#' <FIXME - DELETE THIS AND LINE ABOVE IF OMITTED>
#'
#' @template seealso_learner
#' @template example
#' @export
LearnerClassifRpart = R6Class("LearnerClassifRpart",
inherit = LearnerClassif,

public = list(
#' @description
#' Creates a new instance of this [R6][R6::R6Class] class.
initialize = function() {
# FIXME - MANUALLY ADD PARAM_SET BELOW AND THEN DELETE THIS LINE
ps = <param_set>

# FIXME - MANUALLY UPDATE PARAM VALUES BELOW IF APPLICABLE THEN DELETE THIS LINE.
# OTHERWISE DELETE THIS AND LINE BELOW.
ps$values = list(<param_vals>) super$initialize(
id = "classif.rpart",
packages = "rpart",
feature_types = c("logical", "integer", "numeric", "factor", "ordered"),
predict_types = c("response", "prob"),
param_set = ps,
properties = c("importance", "missings", "multiclass", "selected_features", "twoclass", "weights"),
man = "mlr3extralearners::mlr_learners_classif.rpart"
)
},

# FIXME - ADD IMPORTANCE METHOD HERE AND DELETE THIS LINE.
# <See LearnerRegrRandomForest for an example>
#' @description
#' The importance scores are extracted from the slot <FIXME>.
#' @return Named numeric().
importance = function() { }

),

private = list(

pars = self$param_set$get_values(tags = "train")

# set column names to ensure consistency in fit and predict
self$state$feature_names = task$feature_names # FIXME - <Create objects for the train call # <At least "data" and "formula" are required> formula = task$formula()
data = task$data() # FIXME - <here is space for some custom adjustments before proceeding to the # train call. Check other learners for what can be done here> # use the mlr3misc::invoke function (it's similar to do.call()) mlr3misc::invoke(rpart::rpart, formula = formula, data = data, .args = pars) }, .predict = function(task) { # get parameters with tag "predict" pars = self$param_set$get_values(tags = "predict") # get newdata newdata = task$data(cols = task$feature_names) pred = mlr3misc::invoke(predict, self$model, newdata = newdata,
type = type, .args = pars)

# FIXME - ADD PREDICTIONS TO LIST BELOW
list(...)
}
)
)

.extralrns_dict$add("classif.rpart", LearnerClassifRpart) Now we have to do the following (from top to bottom): 1. Fill in the references under “references” and delete the tag that starts “FIXME” 2. Replace <param_set> with a parameter set 3. Optionally change default values for parameters in <param_vals> 4. As we included “importance” in properties we have to add a function to the public method importance 5. Fill in the private .train method, which takes a (filtered) Task and returns a model. 6. Fill in the private .predict method, which operates on the model in self$model (stored during $train()) and a (differently subsetted) Task to return a named list of predictions. ### 6.1.4 Meta-information In the constructor (initialize()) the constructor of the super class (e.g. LearnerClassif) is called with meta information about the learner which should be constructed. This includes: • id: The ID of the new learner. Usually consists of <type>.<algorithm>, for example: "classif.rpart". • packages: The upstream package name of the implemented learner. • param_set: A set of hyperparameters and their descriptions provided as a paradox::ParamSet. For each hyperparameter the appropriate class needs to be chosen. When using the paradox::ps shortcut, a short constructor of the form p_*** can be used: • predict_types: Set of predict types the learner is able to handle. These differ depending on the type of the learner. See mlr_reflections$learner_predict_types for the full list of feature types supported by mlr3.
• LearnerClassif
• response: Only predicts a class label for each observation in the test set.
• prob: Also predicts the posterior probability for each class for each observation in the test set.
• LearnerRegr
• response: Only predicts a numeric response for each observation in the test set.
• se: Also predicts the standard error for each value of response for each observation in the test set.
• feature_types: Set of feature types the learner is able to handle. See mlr_reflections$task_feature_types for feature types supported by mlr3. • properties: Set of properties of the learner. See mlr_reflections$learner_properties for the full list of feature types supported by mlr3. Possible properties include:
• "twoclass": The learner works on binary classification problems.
• "multiclass": The learner works on multi-class classification problems.
• "missings": The learner can natively handle missing values.
• "weights": The learner can work on tasks which have observation weights / case weights.
• "parallel": The learner supports internal parallelization in some way. Currently not used, this is an experimental property.
• "importance": The learner supports extracting importance values for features. If this property is set, you must also implement a public method importance() to retrieve the importance values from the model.
• "selected_features": The learner supports extracting the features which where used. If this property is set, you must also implement a public method selected_features() to retrieve the set of used features from the model.
• man: The roxygen identifier of the learner. This is used within the $help() method of the super class to open the help page of the learner. ### 6.1.5 ParamSet The param_set is the set of hyperparameters used in model training and predicting, this is given as a paradox::ParamSet. The set consists of a list of hyperparameters, each has a specific class for the hyperparameter type (see above). For classif.rpart the following replace <param_set> above: ps = ParamSet$new(list(
ParamInt$new(id = "minsplit", default = 20L, lower = 1L, tags = "train"), ParamInt$new(id = "minbucket", lower = 1L, tags = "train"),
ParamDbl$new(id = "cp", default = 0.01, lower = 0, upper = 1, tags = "train"), ParamInt$new(id = "maxcompete", default = 4L, lower = 0L, tags = "train"),
ParamInt$new(id = "maxsurrogate", default = 5L, lower = 0L, tags = "train"), ParamInt$new(id = "maxdepth", default = 30L, lower = 1L, upper = 30L, tags = "train"),
ParamInt$new(id = "usesurrogate", default = 2L, lower = 0L, upper = 2L, tags = "train"), ParamInt$new(id = "surrogatestyle", default = 0L, lower = 0L, upper = 1L, tags = "train"),
ParamInt$new(id = "xval", default = 0L, lower = 0L, tags = "train"), ParamLgl$new(id = "keep_model", default = FALSE, tags = "train")
))
data = task$data() mlr3misc::invoke(rpart::rpart, formula = formula, data = data, .args = pars) } ### 6.1.7 Predict function The internal predict method .predict() also operates on a Task as well as on the fitted model that has been created by the train() call previously and has been stored in self$model.

The return value is a Prediction object. We proceed analogously to what we did in the previous section. We start with a version without any mlr3 objects and continue to replace objects until we have reached the desired interface:

# inputs:
self = list(model = rpart::rpart(task$formula(), data = task$data()))

data = iris
response = predict(self$model, newdata = data, type = "class") prob = predict(self$model, newdata = data, type = "prob")

The rpart::predict.rpart() function predicts class labels if argument type is set to to "class", and class probabilities if set to "prob".

Next, we transition from data to a task again and construct a list with the return type requested by the user, this is stored in the $predict_type slot of a learner class. Note that the task is automatically passed to the prediction object, so all you need to do is return the predictions! Make sure the list names are identical to the task predict types. The final .predict() method is below, we could omit the pars line as there are no parameters with the "predict" tag but we keep it here to be consistent: .predict = function(task) { pars = self$param_set$get_values(tags = "predict") # get newdata and ensure same ordering in train and predict newdata = task$data(cols = self$state$feature_names)
if (self$predict_type == "response") { response = mlr3misc::invoke(predict, self$model,
newdata = newdata,
type = "class",
.args = pars)

return(list(response = response))
} else {
prob = mlr3misc::invoke(predict,
self$model, newdata = newdata, type = "prob", .args = pars) return(list(prob = prob)) } } Note that you cannot rely on the column order of the data returned by task$data() as the order of columns may be different from the order of the columns during $.train. The newdata line ensures the ordering is the same by calling the saved order set in $.train, don’t delete either of these lines!

### 6.1.8 Control objects/functions of learners

Some learners rely on a “control” object/function such as glmnet::glmnet.control(). Accounting for such depends on how the underlying package works:

• If the package forwards the control parameters via ... and makes it possible to just pass control parameters as additional parameters directly to the train call, there is no need to distinguish both "train" and "control" parameters. Both can be tagged with “train” in the ParamSet and just be handed over as shown previously.
• If the control parameters need to be passed via a separate argument, the parameters should also be tagged accordingly in the ParamSet. Afterwards they can be queried via their tag and passed separately to mlr3misc::invoke(). See example below.
control_pars = mlr3misc::(<package>::<function>,
self$param_set$get_values(tags = "control"))

train_pars = self$param_set$get_values(tags = "train"))

mlr3misc::invoke([...], .args = train_pars, control = control_pars)

### 6.1.9 Testing the learner

Once your learner is created, you are ready to start testing if it works, there are three types of tests: manual, unit and parameter.

#### 6.1.9.1 Train and Predict

For a bare-bone check you can just try to run a simple train() call locally.

task = tsk("iris") # assuming a Classif learner
lrn = lrn("classif.rpart")
lrn$train(task) p = lrn$predict(task)
p$confusion If it runs without erroring, that’s a very good start! #### 6.1.9.2 Autotest To ensure that your learner is able to handle all kinds of different properties and feature types, we have written an “autotest” that checks the learner for different combinations of such. The “autotest” setup is generated automatically by create_learner and will open after running the function, it will have a name with the form test_package_type_key.R, in our case this will actually be test_rpart_classif_rpart.key. This name must not be changed as triggering automated tests rely on a strict naming scheme. In our example this will create the following script, for which no changes are required to pass (assuming the learner was correctly created): install_learners("classif.rpart") test_that("autotest", { learner = LearnerClassifRpart$new()
expect_learner(learner)
result = run_autotest(learner)
expect_true(result, info = result$error) }) For some learners that have required parameters, it is needed to set some values for required parameters after construction so that the learner can be run in the first place. You can also exclude some specific test arrangements within the “autotest” via the argument exclude in the run_autotest() function. Currently the run_autotest() function lives in inst/testthat of the mlr_plkg("mlr3") and still lacks documentation. This should change in the near future. To finally run the test suite, call devtools::test() or hit CTRL + Shift + T if you are using RStudio. #### 6.1.9.3 Checking Parameters Some learners have a high number of parameters and it is easy to miss out on some during the creation of a new learner. In addition, if the maintainer of the upstream package changes something with respect to the arguments of the algorithm, the learner is in danger to break. Also, new arguments could be added upstream and manually checking for new additions all the time is tedious. Therefore we have written a “Parameter Check” that runs for every learner asynchronously to the R CMD Check of the package itself. This “Parameter Check” compares the parameters of the mlr3 ParamSet against all arguments available in the upstream function that is called during $train() and $predict(). Again the file is automatically created and opened by create_learner, this will be named like test_paramtest_package_type_key.R, so in our example test_paramtest_rpart_classif_rpart.R. The test comes with an exclude argument that should be used to exclude and explain why certain arguments of the upstream function are not within the ParamSet of the mlr3learner. This will likely be required for all learners as common arguments like x, target or data are handled by the mlr3 interface and are therefore not included within the ParamSet. However, there might be more parameters that need to be excluded, for example: • Type dependent parameters, i.e. parameters that only apply for classification or regression learners. • Parameters that are actually deprecated by the upstream package and which were therefore not included in the mlr3 ParamSet. All excluded parameters should have a comment justifying their exclusion. In our example, the final paramtest script looks like: library("mlr3extralearners") install_learners("classif.rpart") test_that("classif.rpart train", { learner = lrn("classif.rpart") fun = rpart::rpart exclude = c( "formula",# handled internally "model", # handled internally "data", # handled internally "weights", # handled by task "subset", # handled by task "na.action", # handled internally "method", # handled internally "x", # handled internally "y", # handled internally "parms", # handled internally "control", # handled internally "cost" # handled internally ) ParamTest = run_paramtest(learner, fun, exclude) expect_true(ParamTest, info = paste0( "Missing parameters:", paste0("- '", ParamTest$missing, "'", collapse = "
")))
})

test_that("classif.rpart predict", {
learner = lrn("classif.rpart")
fun = rpart:::predict.rpart
exclude = c(
"object", # handled internally
"newdata", # handled internally
"type", # handled internally
"na.action" # handled internally
)

ParamTest = run_paramtest(learner, fun, exclude)
expect_true(ParamTest, info = paste0(
"Missing parameters:",
paste0("- '", ParamTest$missing, "'", collapse = " "))) }) ### 6.1.10 Package Cleaning Once all tests are passing, run the following functions to ensure that the package remains clean and tidy 1. devtools::document(roclets = c('rd', 'collate', 'namespace')) 2. If you haven’t done this before run: remotes::install_github('pat-s/styler@mlr-style') 3. styler::style_pkg(style = styler::mlr_style) 4. usethis::use_tidy_description() 5. lintr::lint_package() Please fix any errors indicated by lintr before creating a pull request. Finally ensure that all FIXME are resolved and deleted in the generated files. You are now ready to add your learner to the mlr3 ecosystem! Simply open a pull request to with the new learner template and complete the checklist in there. Once the pull request is approved and merged, your learner will automatically appear on the package website. ### 6.1.11 Thanks and Maintenance Thank you for contributing to the mlr3 ecosystem! When you created the learner you would have given your GitHub handle, meaning that you are now listed as the learner author and maintainer. This means that if the learner breaks it is your responsibility to fix the learner - you can view the status of your learner here. ### 6.1.12 Learner FAQ Question 1 How to deal with Parameters which have no default? Answer If the learner does not work without providing a value, set a reasonable default in param_set$values, add tag "required" to the parameter and document your default properly.

Question 2

Where to add the package of the upstream package in the DESCRIPTION file?

Add it to the “Imports” section. This will install the upstream package during the installation of the mlr3learner if it has not yet been installed by the user.

Question 3

How to handle arguments from external “control” functions such as glmnet::glmnet_control()?

Question 4

How to document if my learner uses a custom default value that differs to the default of the upstream package?

If you set a custom default for the mlr3learner that does not cope with the one of the upstream package (think twice if this is really needed!), add this information to the help page of the respective learner.

You can use the following skeleton for this:

#' @section Custom mlr3 defaults:
#' - <parameter>:
#'   - Actual default: <value>
#'   - Reason for change: <text>

Question 5

When should the "required" tag be used when defining Params and what is its purpose?

The "required" tag should be used when the following conditions are met:

• The upstream function cannot be run without setting this parameter, i.e. it would throw an error.
• The parameter has no default in the upstream function.

In mlr3 we follow the principle that every learner should be constructable without setting custom parameters. Therefore, if a parameter has no default in the upstream function, a custom value is usually set for this parameter in the mlr3learner (remember to document such changes in the help page of the learner).

Even though this practice ensures that no parameter is unset in an mlr3learner and partially removes the usefulness of the "required" tag, the tag is still useful in the following scenario:

If a user sets custom parameters after construction of the learner

lrn = lrn("<id>")
lrn$param_set$values = list("<param>" = <value>)

Here, all parameters besides the ones set in the list would be unset. See paradox::ParamSet for more information. If a parameter is tagged as "required" in the ParamSet, the call above would error and prompt the user that required parameters are missing.

Question 6

What is this error when I run devtools::load_all()

> devtools::load_all(".")
Warning message:
call: vapply(hooks, function(x) environment(x)$pkgname, NA_character_) error: values must be length 1, but FUN(X[[1]]) result is length 0 Answer This is not an error but a warning and you can safely ignore it! ## 6.2 Adding new Measures In this section we showcase how to implement a custom performance measure. A good starting point is writing down the loss function independently of mlr3 (we also did this in the mlr3measures package). Here, we illustrate writing measure by implementing the root of the mean squared error for regression problems: root_mse = function(truth, response) { mse = mean((truth - response)^2) sqrt(mse) } root_mse(c(0, 0.5, 1), c(0.5, 0.5, 0.5)) ## [1] 0.4082 In the next step, we embed the root_mse() function into a new R6 class inheriting from base classes MeasureRegr/Measure. For classification measures, use MeasureClassif. We keep it simple here and only explain the most important parts of the Measure class: MeasureRootMSE = R6::R6Class("MeasureRootMSE", inherit = mlr3::MeasureRegr, public = list( initialize = function() { super$initialize(
# custom id for the measure
id = "root_mse",

# additional packages required to calculate this measure
packages = character(),

# properties, see below
properties = character(),

# required predict type of the learner
predict_type = "response",

# feasible range of values
range = c(0, Inf),

# minimize during tuning?
minimize = TRUE
)
}
),

private = list(
# custom scoring function operating on the prediction object
.score = function(prediction, ...) {
root_mse = function(truth, response) {
mse = mean((truth - response)^2)
sqrt(mse)
}

root_mse(prediction$truth, prediction$response)
}
)
)

This class can be used as template for most performance measures. If something is missing, you might want to consider having a deeper dive into the following arguments:

• properties: If you tag you measure with the property "requires_task", the Task is automatically passed to your .score() function (don’t forget to add the argument task in the signature). The same is possible with "requires_learner" if you need to operate on the Learner and "requires_train_set" if you want to access the set of training indices in the score function.
• aggregator: This function (defaulting to mean()) controls how multiple performance scores, i.e. from different resampling iterations, are aggregated into a single numeric value if average is set to micro averaging. This is ignored for macro averaging.
• predict_sets: Prediction sets (subset of ("train", "test")) to operate on. Defaults to the “test” set.

Finally, if you want to use your custom measure just like any other measure shipped with mlr3 and access it via the mlr_measures dictionary, you can easily add it:

mlr3::mlr_measures$add("root_mse", MeasureRootMSE) Typically it is a good idea to put the measure together with the call to mlr_measures$add() in a new R file and just source it in your project.

## source("measure_root_mse.R")
msr("root_mse")
## <MeasureRootMSE:root_mse>
## * Packages: -
## * Range: [0, Inf]
## * Minimize: TRUE
## * Parameters: list()
## * Properties: -
## * Predict type: response

This section showcases how the mlr3pipelines package can be extended to include custom PipeOps. To run the following examples, we will need a Task; we are using the well-known “Iris” task:

library("mlr3")
task$data() ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: setosa 1.4 0.2 5.1 3.5 ## 2: setosa 1.4 0.2 4.9 3.0 ## 3: setosa 1.3 0.2 4.7 3.2 ## 4: setosa 1.5 0.2 4.6 3.1 ## 5: setosa 1.4 0.2 5.0 3.6 ## --- ## 146: virginica 5.2 2.3 6.7 3.0 ## 147: virginica 5.0 1.9 6.3 2.5 ## 148: virginica 5.2 2.0 6.5 3.0 ## 149: virginica 5.4 2.3 6.2 3.4 ## 150: virginica 5.1 1.8 5.9 3.0 mlr3pipelines is fundamentally built around R6. When planning to create custom PipeOp objects, it can only help to familiarize yourself with it. In principle, all a PipeOp must do is inherit from the PipeOp R6 class and implement the .train() and .predict() functions. There are, however, several auxiliary subclasses that can make the creation of certain operations much easier. ### 6.3.1 General Case Example: PipeOpCopy A very simple yet useful PipeOp is PipeOpCopy, which takes a single input and creates a variable number of output channels, all of which receive a copy of the input data. It is a simple example that showcases the important steps in defining a custom PipeOp. We will show a simplified version here, PipeOpCopyTwo, that creates exactly two copies of its input data. The following figure visualizes how our PipeOp is situated in the Pipeline and the significant in- and outputs. #### 6.3.1.1 First Steps: Inheriting from PipeOp The first part of creating a custom PipeOp is inheriting from PipeOp. We make a mental note that we need to implement a .train() and a .predict() function, and that we probably want to have an initialize() as well: PipeOpCopyTwo = R6::R6Class("PipeOpCopyTwo", inherit = mlr3pipelines::PipeOp, public = list( initialize = function(id = "copy.two") { .... }, ), private == list( .train = function(inputs) { .... }, .predict = function(inputs) { .... } ) ) Note, that private methods, e.g. .train and .predict etc are prefixed with a .. #### 6.3.1.2 Channel Definitions We need to tell the PipeOp the layout of its channels: How many there are, what their names are going to be, and what types are acceptable. This is done on initialization of the PipeOp (using a super$initialize call) by giving the input and output data.table objects. These must have three columns: a "name" column giving the names of input and output channels, and a "train" and "predict" column naming the class of objects we expect during training and prediction as input / output. A special value for these classes is "*", which indicates that any class will be accepted; our simple copy operator accepts any kind of input, so this will be useful. We have only one input, but two output channels.

By convention, we name a single channel "input" or "output", and a group of channels ["input1", "input2", …], unless there is a reason to give specific different names. Therefore, our input data.table will have a single row <"input", "*", "*">, and our output table will have two rows, <"output1", "*", "*"> and <"output2", "*", "*">.

All of this is given to the PipeOp creator. Our initialize() will thus look as follows:

    initialize = function(id = "copy.two") {
input = data.table::data.table(name = "input", train = "*", predict = "*")
# the following will create two rows and automatically fill the train
# and predict cols with "*"
output = data.table::data.table(
name = c("output1", "output2"),
train = "*", predict = "*"
)
super$initialize(id, input = input, output = output ) } #### 6.3.1.3 Train and Predict Both .train() and .predict() will receive a list as input and must give a list in return. According to our input and output definitions, we will always get a list with a single element as input, and will need to return a list with two elements. Because all we want to do is create two copies, we will just create the copies using c(inputs, inputs). Two things to consider: • The .train() function must always modify the self$state variable to something that is not NULL or NO_OP. This is because the $state slot is used as a signal that PipeOp has been trained on data, even if the state itself is not important to the PipeOp (as in our case). Therefore, our .train() will set self$state = list().

• It is not necessary to “clone” our input or make deep copies, because we don’t modify the data. However, if we were changing a reference-passed object, for example by changing data in a Task, we would have to make a deep copy first. This is because a PipeOp may never modify its input object by reference.

Our .train() and .predict() functions are now:

.train = function(inputs) {
self$state = list() c(inputs, inputs) } .predict = function(inputs) { c(inputs, inputs) } #### 6.3.1.4 Putting it Together The whole definition thus becomes PipeOpCopyTwo = R6::R6Class("PipeOpCopyTwo", inherit = mlr3pipelines::PipeOp, public = list( initialize = function(id = "copy.two") { super$initialize(id,
input = data.table::data.table(name = "input", train = "*", predict = "*"),
output = data.table::data.table(name = c("output1", "output2"),
train = "*", predict = "*")
)
}
),
private = list(
.train = function(inputs) {
self$state = list() c(inputs, inputs) }, .predict = function(inputs) { c(inputs, inputs) } ) ) We can create an instance of our PipeOp, put it in a graph, and see what happens when we train it on something: library("mlr3pipelines") poct = PipeOpCopyTwo$new()
gr = Graph$new() gr$add_pipeop(poct)

print(gr)
## Graph with 1 PipeOps:
##        ID         State sccssors prdcssors
##  copy.two <<UNTRAINED>>
result = gr$train(task) str(result) ## List of 2 ##$ copy.two.output1:Classes 'TaskClassif', 'TaskSupervised', 'Task', 'R6' <TaskClassif:iris>
##  $copy.two.output2:Classes 'TaskClassif', 'TaskSupervised', 'Task', 'R6' <TaskClassif:iris> ### 6.3.2 Special Case: Preprocessing Many PipeOps perform an operation on exactly one Task, and return exactly one Task. They may even not care about the “Target” / “Outcome” variable of that task, and only do some modification of some input data. However, it is usually important to them that the Task on which they perform prediction has the same data columns as the Task on which they train. For these cases, the auxiliary base class PipeOpTaskPreproc exists. It inherits from PipeOp itself, and other PipeOps should use it if they fall in the kind of use-case named above. When inheriting from PipeOpTaskPreproc, one must either implement the private methods .train_task() and .predict_task(), or the methods .train_dt(), .predict_dt(), depending on whether wants to operate on a Task object or on its data as data.tables. In the second case, one can optionally also overload the .select_cols() method, which chooses which of the incoming Task’s features are given to the .train_dt() / .predict_dt() functions. The following will show two examples: PipeOpDropNA, which removes a Task’s rows with missing values during training (and implements .train_task() and .predict_task()), and PipeOpScale, which scales a Task’s numeric columns (and implements .train_dt(), .predict_dt(), and .select_cols()). #### 6.3.2.1 Example: PipeOpDropNA Dropping rows with missing values may be important when training a model that can not handle them. Because mlr3 Tasks only contain a view to the underlying data, it is not necessary to modify data to remove rows with missing values. Instead, the rows can be removed using the Task’s $filter method, which modifies the Task in-place. This is done in the private method .train_task(). We take care that we also set the $state slot to signal that the PipeOp was trained. The private method .predict_task() does not need to do anything; removing missing values during prediction is not as useful, since learners that cannot handle them will just ignore the respective rows. Furthermore, mlr3 expects a Learner to always return just as many predictions as it was given input rows, so a PipeOp that removes Task rows during training can not be used inside a GraphLearner. When we inherit from PipeOpTaskPreproc, it sets the input and output data.tables for us to only accept a single Task. The only thing we do during initialize() is therefore to set an id (which can optionally be changed by the user). The complete PipeOpDropNA can therefore be written as follows. Note that it inherits from PipeOpTaskPreproc, unlike the PipeOpCopyTwo example from above: PipeOpDropNA = R6::R6Class("PipeOpDropNA", inherit = mlr3pipelines::PipeOpTaskPreproc, public = list( initialize = function(id = "drop.na") { super$initialize(id)
}
),

private = list(
self$state = list() featuredata = task$data(cols = task$feature_names) exclude = apply(is.na(featuredata), 1, any) task$filter(task$row_ids[!exclude]) }, .predict_task = function(task) { # nothing to be done task } ) ) To test this PipeOp, we create a small task with missing values: smalliris = iris[(1:5) * 30, ] smalliris[1, 1] = NA smalliris[2, 2] = NA sitask = as_task_classif(smalliris, target = "Species") print(sitask$data())
##       Species Petal.Length Petal.Width Sepal.Length Sepal.Width
## 1:     setosa          1.6         0.2           NA         3.2
## 2: versicolor          3.9         1.4          5.2          NA
## 3: versicolor          4.0         1.3          5.5         2.5
## 4:  virginica          5.0         1.5          6.0         2.2
## 5:  virginica          5.1         1.8          5.9         3.0

We test this by feeding it to a new Graph that uses PipeOpDropNA.

gr = Graph$new() gr$add_pipeop(PipeOpDropNA$new()) filtered_task = gr$train(sitask)[[1]]
print(filtered_task$data()) ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: versicolor 4.0 1.3 5.5 2.5 ## 2: virginica 5.0 1.5 6.0 2.2 ## 3: virginica 5.1 1.8 5.9 3.0 #### 6.3.2.2 Example: PipeOpScaleAlways An often-applied preprocessing step is to simply center and/or scale the data to mean $$0$$ and standard deviation $$1$$. This fits the PipeOpTaskPreproc pattern quite well. Because it always replaces all columns that it operates on, and does not require any information about the task’s target, it only needs to overload the .train_dt() and .predict_dt() functions. This saves some boilerplate-code from getting the correct feature columns out of the task, and replacing them after modification. Because scaling only makes sense on numeric features, we want to instruct PipeOpTaskPreproc to give us only these numeric columns. We do this by overloading the .select_cols() function: It is called by the class to determine which columns to pass to .train_dt() and .predict_dt(). Its input is the Task that is being transformed, and it should return a character vector of all features to work with. When it is not overloaded, it uses all columns; instead, we will set it to only give us numeric columns. Because the levels() of the data table given to .train_dt() and .predict_dt() may be different from the task’s levels, these functions must also take a levels argument that is a named list of column names indicating their levels. When working with numeric data, this argument can be ignored, but it should be used instead of levels(dt[[column]]) for factorial or character columns. This is the first PipeOp where we will be using the $state slot for something useful: We save the centering offset and scaling coefficient and use it in $.predict()! For simplicity, we are not using hyperparameters and will always scale and center all data. Compare this PipeOpScaleAlways operator to the one defined inside the mlr3pipelines package, PipeOpScale. PipeOpScaleAlways = R6::R6Class("PipeOpScaleAlways", inherit = mlr3pipelines::PipeOpTaskPreproc, public = list( initialize = function(id = "scale.always") { super$initialize(id = id)
}
),

private = list(
task$feature_types[type == "numeric", id] }, .train_dt = function(dt, levels, target) { sc = scale(as.matrix(dt)) self$state = list(
center = attr(sc, "scaled:center"),
scale = attr(sc, "scaled:scale")
)
sc
},

.predict_dt = function(dt, levels) {
t((t(dt) - self$state$center) / self$state$scale)
}
)
)

(Note for the observant: If you check PipeOpScale.R from the mlr3pipelines package, you will notice that is uses “get("type")” and “get("id")” instead of “type” and “id”, because the static code checker on CRAN would otherwise complain about references to undefined variables. This is a “problem” with data.table and not exclusive to mlr3pipelines.)

We can, again, create a new Graph that uses this PipeOp to test it. Compare the resulting data to the original “iris” Task data printed at the beginning:

gr = Graph$new() gr$add_pipeop(PipeOpScaleAlways$new()) result = gr$train(task)

result[[1]]$data() ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: setosa -1.3358 -1.3111 -0.89767 1.01560 ## 2: setosa -1.3358 -1.3111 -1.13920 -0.13154 ## 3: setosa -1.3924 -1.3111 -1.38073 0.32732 ## 4: setosa -1.2791 -1.3111 -1.50149 0.09789 ## 5: setosa -1.3358 -1.3111 -1.01844 1.24503 ## --- ## 146: virginica 0.8169 1.4440 1.03454 -0.13154 ## 147: virginica 0.7036 0.9192 0.55149 -1.27868 ## 148: virginica 0.8169 1.0504 0.79301 -0.13154 ## 149: virginica 0.9302 1.4440 0.43072 0.78617 ## 150: virginica 0.7602 0.7880 0.06843 -0.13154 ### 6.3.3 Special Case: Preprocessing with Simple Train It is possible to make even further simplifications for many PipeOps that perform mostly the same operation during training and prediction. The point of Task preprocessing is often to modify the training data in mostly the same way as prediction data (but in a way that may depend on training data). Consider constant feature removal, for example: The goal is to remove features that have no variance, or only a single factor level. However, what features get removed must be decided during training, and may only depend on training data. Furthermore, the actual process of removing features is the same during training and prediction. A simplification to make is therefore to have a private method .get_state(task) which sets the $state slot during training, and a private method .transform(task), which gets called both during training and prediction. This is done in the PipeOpTaskPreprocSimple class. Just like PipeOpTaskPreproc, one can inherit from this and overload these functions to get a PipeOp that performs preprocessing with very little boilerplate code.

Just like PipeOpTaskPreproc, PipeOpTaskPreprocSimple offers the possibility to instead overload the .get_state_dt(dt, levels) and .transform_dt(dt, levels) methods (and optionally, again, the .select_cols(task) function) to operate on data.table feature data instead of the whole Task.

Even some methods that do not use PipeOpTaskPreprocSimple could work in a similar way: The PipeOpScaleAlways example from above will be shown to also work with this paradigm.

#### 6.3.3.1 Example: PipeOpDropConst

A typical example of a preprocessing operation that does almost the same operation during training and prediction is an operation that drops features depending on a criterion that is evaluated during training. One simple example of this is dropping constant features. Because the mlr3 Task class offers a flexible view on underlying data, it is most efficient to drop columns from the task directly using its $select() function, so the .get_state_dt(dt, levels) / .transform_dt(dt, levels) functions will not get used; instead we overload the .get_state(task) and .transform(task) methods. The .get_state() function’s result is saved to the $state slot, so we want to return something that is useful for dropping features. We choose to save the names of all the columns that have nonzero variance. For brevity, we use length(unique(column)) > 1 to check whether more than one distinct value is present; a more sophisticated version could have a tolerance parameter for numeric values that are very close to each other.

The .transform() method is evaluated both during training and prediction, and can rely on the $state slot being present. All it does here is call the Task$select function with the columns we chose to keep.

The full PipeOp could be written as follows:

PipeOpDropConst = R6::R6Class("PipeOpDropConst",
public = list(
initialize = function(id = "drop.const") {
super$initialize(id = id) } ), private = list( .get_state = function(task) { data = task$data(cols = task$feature_names) nonconst = sapply(data, function(column) length(unique(column)) > 1) list(cnames = colnames(data)[nonconst]) }, .transform = function(task) { task$select(self$state$cnames)
}
)
)

This can be tested using the first five rows of the “Iris” Task, for which one feature ("Petal.Width") is constant:

irishead = task$clone()$filter(1:5)
irishead$data() ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: setosa 1.4 0.2 5.1 3.5 ## 2: setosa 1.4 0.2 4.9 3.0 ## 3: setosa 1.3 0.2 4.7 3.2 ## 4: setosa 1.5 0.2 4.6 3.1 ## 5: setosa 1.4 0.2 5.0 3.6 gr = Graph$new()$add_pipeop(PipeOpDropConst$new())
dropped_task = gr$train(irishead)[[1]] dropped_task$data()
##    Species Petal.Length Sepal.Length Sepal.Width
## 1:  setosa          1.4          5.1         3.5
## 2:  setosa          1.4          4.9         3.0
## 3:  setosa          1.3          4.7         3.2
## 4:  setosa          1.5          4.6         3.1
## 5:  setosa          1.4          5.0         3.6

We can also see that the $state was correctly set. Calling $.predict() with this graph, even with different data (the whole Iris Task!) will still drop the "Petal.Width" column, as it should.

gr$pipeops$drop.const$state ##$cnames
## [1] "Petal.Length" "Sepal.Length" "Sepal.Width"
##
## $affected_cols ## [1] "Petal.Length" "Petal.Width" "Sepal.Length" "Sepal.Width" ## ##$intasklayout
##              id    type
## 1: Petal.Length numeric
## 2:  Petal.Width numeric
## 3: Sepal.Length numeric
## 4:  Sepal.Width numeric
##
## $outtasklayout ## id type ## 1: Petal.Length numeric ## 2: Sepal.Length numeric ## 3: Sepal.Width numeric ## ##$outtaskshell
## Empty data.table (0 rows and 4 cols): Species,Petal.Length,Sepal.Length,Sepal.Width
dropped_predict = gr$predict(task)[[1]] dropped_predict$data()
##        Species Petal.Length Sepal.Length Sepal.Width
##   1:    setosa          1.4          5.1         3.5
##   2:    setosa          1.4          4.9         3.0
##   3:    setosa          1.3          4.7         3.2
##   4:    setosa          1.5          4.6         3.1
##   5:    setosa          1.4          5.0         3.6
##  ---
## 146: virginica          5.2          6.7         3.0
## 147: virginica          5.0          6.3         2.5
## 148: virginica          5.2          6.5         3.0
## 149: virginica          5.4          6.2         3.4
## 150: virginica          5.1          5.9         3.0

#### 6.3.3.2 Example: PipeOpScaleAlwaysSimple

This example will show how a PipeOpTaskPreprocSimple can be used when only working on feature data in form of a data.table. Instead of calling the scale() function, the center and scale values are calculated directly and saved to the $state slot. The .transform_dt() function will then perform the same operation during both training and prediction: subtract the center and divide by the scale value. As in the PipeOpScaleAlways example above, we use .select_cols() so that we only work on numeric columns. PipeOpScaleAlwaysSimple = R6::R6Class("PipeOpScaleAlwaysSimple", inherit = mlr3pipelines::PipeOpTaskPreprocSimple, public = list( initialize = function(id = "scale.always.simple") { super$initialize(id = id)
}
),

private = list(
task$feature_types[type == "numeric", id] }, .get_state_dt = function(dt, levels, target) { list( center = sapply(dt, mean), scale = sapply(dt, sd) ) }, .transform_dt = function(dt, levels) { t((t(dt) - self$state$center) / self$state$scale) } ) ) We can compare this PipeOp to the one above to show that it behaves the same. gr = Graph$new()$add_pipeop(PipeOpScaleAlways$new())
result_posa = gr$train(task)[[1]] gr = Graph$new()$add_pipeop(PipeOpScaleAlwaysSimple$new())
result_posa_simple = gr$train(task)[[1]] result_posa$data()
##        Species Petal.Length Petal.Width Sepal.Length Sepal.Width
##   1:    setosa      -1.3358     -1.3111     -0.89767     1.01560
##   2:    setosa      -1.3358     -1.3111     -1.13920    -0.13154
##   3:    setosa      -1.3924     -1.3111     -1.38073     0.32732
##   4:    setosa      -1.2791     -1.3111     -1.50149     0.09789
##   5:    setosa      -1.3358     -1.3111     -1.01844     1.24503
##  ---
## 146: virginica       0.8169      1.4440      1.03454    -0.13154
## 147: virginica       0.7036      0.9192      0.55149    -1.27868
## 148: virginica       0.8169      1.0504      0.79301    -0.13154
## 149: virginica       0.9302      1.4440      0.43072     0.78617
## 150: virginica       0.7602      0.7880      0.06843    -0.13154
result_posa_simple$data() ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: setosa -1.3358 -1.3111 -0.89767 1.01560 ## 2: setosa -1.3358 -1.3111 -1.13920 -0.13154 ## 3: setosa -1.3924 -1.3111 -1.38073 0.32732 ## 4: setosa -1.2791 -1.3111 -1.50149 0.09789 ## 5: setosa -1.3358 -1.3111 -1.01844 1.24503 ## --- ## 146: virginica 0.8169 1.4440 1.03454 -0.13154 ## 147: virginica 0.7036 0.9192 0.55149 -1.27868 ## 148: virginica 0.8169 1.0504 0.79301 -0.13154 ## 149: virginica 0.9302 1.4440 0.43072 0.78617 ## 150: virginica 0.7602 0.7880 0.06843 -0.13154 ### 6.3.4 Hyperparameters mlr3pipelines uses the paradox package to define parameter spaces for PipeOps. Parameters for PipeOps can modify their behavior in certain ways, e.g. switch centering or scaling off in the PipeOpScale operator. The unified interface makes it possible to have parameters for whole Graphs that modify the individual PipeOp’s behavior. The Graphs, when encapsulated in GraphLearners, can even be tuned using the tuning functionality in mlr3tuning. Hyperparameters are declared during initialization, when calling the PipeOp’s $initialize() function, by giving a param_set argument. The param_set must be a ParamSet from the paradox package; see the tuning chapter or the in-depth paradox chapter for more information on how to define parameter spaces. After construction, the ParamSet can be accessed through the $param_set slot. While it is possible to modify this ParamSet, using e.g. the $add() and $add_dep() functions, after adding it to the PipeOp, it is strongly advised against. Hyperparameters can be set and queried through the $values slot. When setting hyperparameters, they are automatically checked to satisfy all conditions set by the $param_set, so it is not necessary to type check them. Be aware that it is always possible to remove hyperparameter values. When a PipeOp is initialized, it usually does not have any parameter values—$values takes the value list(). It is possible to set initial parameter values in the $initialize() constructor; this must be done after the super$initialize() call where the corresponding ParamSet must be supplied. This is because setting $values checks against the current $param_set, which would fail if the $param_set was not set yet. When using an underlying library function (the scale function in PipeOpScale, say), then there is usually a “default” behaviour of that function when a parameter is not given. It is good practice to use this default behaviour whenever a parameter is not set (or when it was removed). This can easily be done when using the mlr3misc library’s mlr3misc::invoke() function, which has functionality similar to do.call(). #### 6.3.4.1 Hyperparameter Example: PipeOpScale How to use hyperparameters can best be shown through the example of PipeOpScale, which is very similar to the example above, PipeOpScaleAlways. The difference is made by the presence of hyperparameters. PipeOpScale constructs a ParamSet in its $initialize function and passes this on to the super$initialize function: PipeOpScale$public_methods$initialize ## function (id = "scale", param_vals = list()) ## .__PipeOpScale__initialize(self = self, private = private, super = super, ## id = id, param_vals = param_vals) ## <environment: namespace:mlr3pipelines> The user has access to this and can set and get parameters. Types are automatically checked: pss = po("scale") print(pss$param_set)
## <ParamSet:scale>
##                id    class lower upper nlevels        default value
## 1:         center ParamLgl    NA    NA       2           TRUE
## 2:          scale ParamLgl    NA    NA       2           TRUE
## 3:         robust ParamLgl    NA    NA       2 <NoDefault[3]> FALSE
## 4: affect_columns ParamUty    NA    NA     Inf  <Selector[1]>
pss$param_set$values$center = FALSE print(pss$param_set$values) ##$robust
## [1] FALSE
##
## $center ## [1] FALSE pss$param_set$values$scale = "TRUE"  # bad input is checked!
## Error in self$assert(xs): Assertion on 'xs' failed: scale: Must be of type 'logical flag', not 'character'. How PipeOpScale handles its parameters can be seen in its $.train_dt method: It gets the relevant parameters from its $values slot and uses them in the mlr3misc::invoke() call. This has the advantage over calling scale() directly that if a parameter is not given, its default value from the scale() function will be used. PipeOpScale$private_methods$.train_dt ## function (dt, levels, target) ## .__PipeOpScale__.train_dt(self = self, private = private, super = super, ## dt = dt, levels = levels, target = target) ## <environment: namespace:mlr3pipelines> Another change that is necessary compared to PipeOpScaleAlways is that the attributes "scaled:scale" and "scaled:center" are not always present, depending on parameters, and possibly need to be set to default values $$1$$ or $$0$$, respectively. It is now even possible (if a bit pointless) to call PipeOpScale with both scale and center set to FALSE, which returns the original dataset, unchanged. pss$param_set$values$scale = FALSE
pss$param_set$values$center = FALSE gr = Graph$new()
gr$add_pipeop(pss) result = gr$train(task)

result[[1]]$data() ## Species Petal.Length Petal.Width Sepal.Length Sepal.Width ## 1: setosa 1.4 0.2 5.1 3.5 ## 2: setosa 1.4 0.2 4.9 3.0 ## 3: setosa 1.3 0.2 4.7 3.2 ## 4: setosa 1.5 0.2 4.6 3.1 ## 5: setosa 1.4 0.2 5.0 3.6 ## --- ## 146: virginica 5.2 2.3 6.7 3.0 ## 147: virginica 5.0 1.9 6.3 2.5 ## 148: virginica 5.2 2.0 6.5 3.0 ## 149: virginica 5.4 2.3 6.2 3.4 ## 150: virginica 5.1 1.8 5.9 3.0 ## 6.4 Adding new Tuners In this section, we show how to implement a custom tuner for mlr3tuning. The main task of a tuner is to iteratively propose new hyperparameter configurations that we want to evaluate for a given task, learner and validation strategy. The second task is to decide which configuration should be returned as a tuning result - usually it is the configuration that led to the best observed performance value. If you want to implement your own tuner, you have to implement an R6-Object that offers an .optimize method that implements the iterative proposal and you are free to implement .assign_result to differ from the before-mentioned default process of determining the result. Before you start with the implementation make yourself familiar with the main R6-Objects in bbotk (Black-Box Optimization Toolkit). This package does not only provide basic black box optimization algorithms and but also the objects that represent the optimization problem (bbotk::OptimInstance) and the log of all evaluated configurations (bbotk::Archive). There are two ways to implement a new tuner: a ) If your new tuner can be applied to any kind of optimization problem it should be implemented as a bbotk::Optimizer. Any bbotk::Optimizer can be easily transformed to a mlr3tuning::Tuner. b) If the new custom tuner is only usable for hyperparameter tuning, for example because it needs to access the task, learner or resampling objects it should be directly implemented in mlr3tuning as a mlr3tuning::Tuner. ### 6.4.1 Adding a new Tuner This is a summary of steps for adding a new tuner. The fifth step is only required if the new tuner is added via bbotk. 1. Check the tuner does not already exist as a bbotk::Optimizer or mlr3tuning::Tuner in the GitHub repositories. 2. Use one of the existing optimizers / tuners as a template. 3. Overwrite the .optimize private method of the optimizer / tuner. 4. Optionally, overwrite the default .assign_result private method. 5. Use the mlr3tuning::TunerFromOptimizer class to transform the bbotk::Optimizer to a mlr3tuning::Tuner. 6. Add unit tests for the tuner and optionally for the optimizer. 7. Open a new pull request for the mlr3tuning::Tuner and optionally a second one for the bbotk::Optimizer. ### 6.4.2 Template If the new custom tuner is implemented via bbotk, use one of the existing optimizer as a template e.g. bbotk::OptimizerRandomSearch. There are currently only two tuners that are not based on a bbotk::Optimizer: mlr3hyperband::TunerHyperband and mlr3tuning::TunerIrace. Both are rather complex but you can still use the documentation and class structure as a template. The following steps are identical for optimizers and tuners. Rewrite the meta information in the documentation and create a new class name. Scientific sources can be added in R/bibentries.R which are added under @source in the documentation. The example and dictionary sections of the documentation are auto-generated based on the @templateVar id <tuner_id>. Change the parameter set of the optimizer / tuner and document them under @section Parameters. Do not forget to change mlr_optimizers$add() / mlr_tuners$add() in the last line which adds the optimizer / tuner to the dictionary. ### 6.4.3 Optimize method The $.optimize() private method is the main part of the tuner. It takes an instance, proposes new points and calls the $eval_batch() method of the instance to evaluate them. Here you can go two ways: Implement the iterative process yourself or call an external optimization function that resides in another package. #### 6.4.3.1 Writing a custom iteration Usually, the proposal and evaluation is done in a repeat-loop which you have to implement. Please consider the following points: • You can evaluate one or multiple points per iteration • You don’t have to care about termination, as $eval_batch() won’t allow more evaluations then allowed by the bbotk::Terminator. This implies, that code after the repeat-loop will not be executed.
• You don’t have to care about keeping track of the evaluations as every evaluation is automatically stored in inst$archive. • If you want to log additional information for each evaluation of the bbotk::Objective in thebbotk::Archiveyou can simply add columns to thedata.tableobject that is passed to$eval_batch().

#### 6.4.3.2 Calling an external optimization function

Optimization functions from external packages usually take an objective function as an argument. In this case, you can pass inst$objective_function which internally calls $eval_batch(). Check out OptimizerGenSA for an example.

### 6.4.4 Assign result method

The default $.assign_result() private method simply obtains the best performing result from the archive. The default method can be overwritten if the new tuner determines the result of the optimization in a different way. The new function must call the $assign_result() method of the instance to write the final result to the instance. See mlr3tuning::TunerIrace for an implementation of \$.assign_result().

### 6.4.5 Transform optimizer to tuner

This step is only needed if you implement via bbotk. The mlr3tuning::TunerFromOptimizer class transforms a bbotk::Optimizer to a mlr3tuning::Tuner. Just add the bbotk::Optimizer to the optimizer field. See mlr3tuning::TunerRandomSearch for an example.

The new custom tuner should be thoroughly tested with unit tests. mlr3tuning::Tuners can be tested with the test_tuner() helper function. If you added the Tuner via a bbotk::Optimizer, you should additionally test the bbotk::Optimizer with the test_optimizer()` helper function.