The following tuners are available:


Grid search. All kinds of parameter types can be handled. You can either use their correct param type and resolution, or discretize them yourself by always using ParamHelpers::makeDiscreteParam in the par.set passed to tuneParams.


Random search. All kinds of parameter types can be handled.


Evolutionary method mco::nsga2. Can handle numeric(vector) and integer(vector) hyperparameters, but no dependencies. For integers the internally proposed numeric values are automatically rounded.


Model-based/ Bayesian optimization. All kinds of parameter types can be handled.

  same.resampling.instance = TRUE,
  resolution = 10L, = "default",
  final.dw.perc = NULL,
  budget = NULL

  n.objectives = mbo.control$n.objectives,
  same.resampling.instance = TRUE,
  impute.val = NULL,
  learner = NULL,
  mbo.control = NULL,
  tune.threshold = FALSE,
  tune.threshold.args = list(),
  continue = FALSE, = "default",
  final.dw.perc = NULL,
  budget = NULL, = NULL

  same.resampling.instance = TRUE,
  impute.val = NULL, = "default",
  final.dw.perc = NULL,
  budget = NULL,

  same.resampling.instance = TRUE,
  maxit = 100L, = "default",
  final.dw.perc = NULL,
  budget = NULL



Should the same resampling instance be used for all evaluations to reduce variance? Default is TRUE.


Resolution of the grid for each numeric/integer parameter in par.set. For vector parameters, it is the resolution per dimension. Either pass one resolution for all parameters, or a named vector. See ParamHelpers::generateGridDesign. Default is 10.

(function | character(1))
Function used for logging. If set to “default” (the default), the evaluated design points, the resulting performances, and the runtime will be reported. If set to “memory” the memory usage for each evaluation will also be displayed, with character(1) small increase in run time. Otherwise character(1) function with arguments learner, resampling, measures, par.set, control, opt.path, dob, x, y, remove.nas, stage and prev.stage is expected. The default displays the performance measures, the time needed for evaluating, the currently used memory and the max memory ever used before (the latter two both taken from gc). See the implementation for details.


If a Learner wrapped by a makeDownsampleWrapper is used, you can define the value of dw.perc which is used to train the Learner with the final parameter setting found by the tuning. Default is NULL which will not change anything.


Maximum budget for tuning. This value restricts the number of function evaluations. In case of makeTuneMultiCritControlGrid this number must be identical to the size of the grid. For makeTuneMultiCritControlRandom the budget equals the number of iterations (maxit) performed by the random search algorithm. In case of makeTuneMultiCritControlNSGA2 the budget corresponds to the product of the maximum number of generations (max(generations)) + 1 (for the initial population) and the size of the population (popsize). For makeTuneMultiCritControlMBO the budget equals the number of objective function evaluations, i.e. the number of MBO iterations + the size of the initial design. If not NULL, this will overwrite existing stopping conditions in mbo.control.


Number of objectives, i.e. number of Measures to optimize.


If something goes wrong during optimization (e.g. the learner crashes), this value is fed back to the tuner, so the tuning algorithm does not abort. Imputation is only active if on.learner.error is configured not to stop in configureMlr. It is not stored in the optimization path, an NA and a corresponding error message are logged instead. Note that this value is later multiplied by -1 for maximization measures internally, so you need to enter a larger positive value for maximization here as well. Default is the worst obtainable value of the performance measure you optimize for when you aggregate by mean value, or Inf instead. For multi-criteria optimization pass a vector of imputation values, one for each of your measures, in the same order as your measures.


(Learner | NULL)
The surrogate learner: A regression learner to model performance landscape. For the default, NULL, mlrMBO will automatically create a suitable learner based on the rules described in mlrMBO::makeMBOLearner.


(mlrMBO::MBOControl | NULL)
Control object for model-based optimization tuning. For the default, NULL, the control object will be created with all the defaults as described in mlrMBO::makeMBOControl.


Should the threshold be tuned for the measure at hand, after each hyperparameter evaluation, via tuneThreshold? Only works for classification if the predict type is “prob”. Default is FALSE.


Further arguments for threshold tuning that are passed down to tuneThreshold. Default is none.


Resume calculation from previous run using mlrMBO::mboContinue? Requires “save.file.path” to be set. Note that the ParamHelpers::OptPath in the mlrMBO::OptResult will only include the evaluations after the continuation. The complete OptPath will be found in the slot $mbo.result$opt.path.

(data.frame | NULL)
Initial design as data frame. If the parameters have corresponding trafo functions, the design must not be transformed before it is passed! For the default, NULL, a default design is created like described in mlrMBO::mbo.


Further control parameters passed to the control arguments of cmaes::cma_es or GenSA::GenSA, as well as towards the tunerConfig argument of irace::irace.


Number of iterations for random search. Default is 100.


(TuneMultiCritControl). The specific subclass is one of TuneMultiCritControlGrid, TuneMultiCritControlRandom, TuneMultiCritControlNSGA2, TuneMultiCritControlMBO.

See also

Other tune_multicrit: plotTuneMultiCritResult(), tuneParamsMultiCrit()