makeModelMultiplexer {mlr}

Create model multiplexer for model selection to tune over multiple possible models.

Description

Combines multiple base learners by dispatching on the hyperparameter “selected.learner” to a specific model class. This allows to tune not only the model class (SVM, random forest, etc) but also their hyperparameters in one go. Combine this with tuneParams and makeTuneControlIrace for a very powerful approach, see example below.

The parameter set is the union of all (unique) base learners. In order to avoid name clashes all parameter names are prefixed with the base learner id, i.e. learnerId.parameterName.

The predict.type of the Multiplexer is inherited from the predict.type of the base learners.

The getter getLearnerProperties returns the properties of the selected base learner.

Usage

makeModelMultiplexer(base.learners)

Arguments

base.learners

([list' of Learner)
List of Learners with unique IDs.

Value

(ModelMultiplexer). A Learner specialized as ModelMultiplexer.

Note

Note that logging output during tuning is somewhat shortened to make it more readable. I.e., the artificial prefix before parameter names is suppressed.

See Also

Other multiplexer: makeModelMultiplexerParamSet()

Other tune: TuneControl, getNestedTuneResultsOptPathDf(), getNestedTuneResultsX(), getResamplingIndices(), getTuneResult(), makeModelMultiplexerParamSet(), makeTuneControlCMAES(), makeTuneControlDesign(), makeTuneControlGenSA(), makeTuneControlGrid(), makeTuneControlIrace(), makeTuneControlMBO(), makeTuneControlRandom(), makeTuneWrapper(), tuneParams(), tuneThreshold()

Examples

set.seed(123)

library(BBmisc)
bls = list(
  makeLearner("classif.ksvm"),
  makeLearner("classif.randomForest")
)
lrn = makeModelMultiplexer(bls)
# simple way to contruct param set for tuning
# parameter names are prefixed automatically and the 'requires'
# element is set, too, to make all paramaters subordinate to 'selected.learner'
ps = makeModelMultiplexerParamSet(lrn,
  makeNumericParam("sigma", lower = -10, upper = 10, trafo = function(x) 2^x),
  makeIntegerParam("ntree", lower = 1L, upper = 500L)
)
print(ps)
rdesc = makeResampleDesc("CV", iters = 2L)
# to save some time we use random search. but you probably want something like this:
# ctrl = makeTuneControlIrace(maxExperiments = 500L)
ctrl = makeTuneControlRandom(maxit = 10L)
res = tuneParams(lrn, iris.task, rdesc, par.set = ps, control = ctrl)
print(res)

df = as.data.frame(res$opt.path)
print(head(df[, -ncol(df)]))

# more unique and reliable way to construct the param set
ps = makeModelMultiplexerParamSet(lrn,
  classif.ksvm = makeParamSet(
    makeNumericParam("sigma", lower = -10, upper = 10, trafo = function(x) 2^x)
  ),
  classif.randomForest = makeParamSet(
    makeIntegerParam("ntree", lower = 1L, upper = 500L)
  )
)

# this is how you would construct the param set manually, works too
ps = makeParamSet(
  makeDiscreteParam("selected.learner", values = extractSubList(bls, "id")),
  makeNumericParam("classif.ksvm.sigma", lower = -10, upper = 10, trafo = function(x) 2^x,
    requires = quote(selected.learner == "classif.ksvm")),
  makeIntegerParam("classif.randomForest.ntree", lower = 1L, upper = 500L,
    requires = quote(selected.learner == "classif.randomForst"))
)

# all three ps-objects are exactly the same internally.