mlr 2.11: 2017-03-15


  • The internal class naming of the task descriptions have been changed causing probable incompatibilities with tasks generated under old versions.
  • New option on.error.dump to include dumps that can be inspected with the debugger with errors
  • mlr now supports tuning with Bayesian optimization with mlrMBO

functions - general

  • tuneParams: fixed a small and obscure bug in logging for extremely large ParamSets
  • getBMR-operators: now support “drop” argument that simplifies the resulting list
  • configureMlr: added option “on.measure.not.applicable” to handle situations where performance cannot be calculated and one wants NA instead of an error - useful in, e.g., larger benchmarks
  • tuneParams, selectFeatures: removed memory stats from default output for performance reasons (can be restored by using a control object with “” = “memory”)
  • listLearners: change check.packages default to FALSE
  • tuneParams and tuneParamsMultiCrit: new parameter to specify a custom resampling function to use.
  • Deprecated: getTaskDescription, getBMRTaskDescriptions, getRRTaskDescription. New names: getTaskDesc, getBMRTaskDescs, getRRTaskDesc.

functions - new

  • getOOBPreds: get out-of-bag predictions from trained models for learners that store them – these learners have the new “oobpreds” property
  • listTaskTypes, listLearnerProperties
  • getMeasureProperties, hasMeasureProperties, listMeasureProperties
  • makeDummyFeaturesWrapper: fuse a learner with a dummy feature creator
  • simplifyMeasureNames: shorten measure names to the actual measure, e.g. mmce.test.mean -> mmce
  • getFailureModelDump, getPredictionDump, getRRDump: get error dumps
  • batchmark: Function to run benchmarks with the batchtools package on high performance computing clusters
  • makeTuneControlMBO: allows Bayesian optimization

measures - new

  • kendalltau, spearmanrho

learners - general

  • classif.plsdaCaret: added parameter “method”.
  • regr.randomForest: refactored se-estimation code, improved docs and default is now se.method = “jackknife”.
  • regr.xgboost, classif.xgboost: removed “factors” property as these learners do not handle categorical features – factors are silently converted to integers internally, which may misinterpret the structure of the data
  • glmnet: control parameters are reset to factory settings before applying custom settings and training and set back to factory afterwards

learners - removed

  • {classif,regr}.avNNet: no longer necessary, mlr contains a bagging wrapper