Benchmark Multiple Learners on Multiple Tasks

Source: R/benchmark.R

Runs a benchmark of the cross-product of learners, tasks, and resampling strategies (possibly in parallel).

Resamplings which are not already instantiated will be instantiated automatically. Note that these auto-instantiated resamplings will not be synchronized per task, i.e. learners will see different splits of the same task.

To generate exhaustive designs and automatically instantiate resampling strategies per task, see expand_grid().

benchmark(design, ctrl = list())

Arguments

design	(data.frame()): Data frame (or data.table()) with three columns: "task", "learner", and "resampling". Each row defines a set of resampled experiments by providing a Task, Learner and Resampling strategy. The helper function expand_grid() can assist in generating an exhaustive design (see examples).
ctrl	(named list as returned by mlr_control()): Object to control experiment execution. See mlr_control().

Value

BenchmarkResult.

Examples

tasks = mlr_tasks$mget(c("iris", "sonar"))
learners = mlr_learners$mget(c("classif.featureless", "classif.rpart"))
resamplings = mlr_resamplings$mget("holdout")

design = expand_grid(tasks, learners, resamplings)
print(design)
#>             task                     learner          resampling
#> 1: <TaskClassif> <LearnerClassifFeatureless> <ResamplingHoldout>
#> 2: <TaskClassif>       <LearnerClassifRpart> <ResamplingHoldout>
#> 3: <TaskClassif> <LearnerClassifFeatureless> <ResamplingHoldout>
#> 4: <TaskClassif>       <LearnerClassifRpart> <ResamplingHoldout>
bmr = benchmark(design)

# performance for all conducted experiments
head(as.data.table(bmr))
#>                hash          task task_id                     learner
#> 1: a9100fbdd3698662 <TaskClassif>    iris <LearnerClassifFeatureless>
#> 2: 3fe90c093a9fdeab <TaskClassif>    iris       <LearnerClassifRpart>
#> 3: cd3af67cd2425cb1 <TaskClassif>   sonar <LearnerClassifFeatureless>
#> 4: a3b3768bd595fd77 <TaskClassif>   sonar       <LearnerClassifRpart>
#>     learner_id          resampling resampling_id      mmce
#> 1: featureless <ResamplingHoldout>       holdout 0.7200000
#> 2:       rpart <ResamplingHoldout>       holdout 0.0800000
#> 3: featureless <ResamplingHoldout>       holdout 0.5072464
#> 4:       rpart <ResamplingHoldout>       holdout 0.3043478

# aggregated performance values
bmr$aggregated
#>                hash  resample_result task_id  learner_id resampling_id
#> 1: a9100fbdd3698662 <ResampleResult>    iris featureless       holdout
#> 2: 3fe90c093a9fdeab <ResampleResult>    iris       rpart       holdout
#> 3: cd3af67cd2425cb1 <ResampleResult>   sonar featureless       holdout
#> 4: a3b3768bd595fd77 <ResampleResult>   sonar       rpart       holdout
#>         mmce
#> 1: 0.7200000
#> 2: 0.0800000
#> 3: 0.5072464
#> 4: 0.3043478

# Overview of of resamplings that were conducted internally
rrs = bmr$resample_results
print(rrs)
#>                hash task_id  learner_id resampling_id N
#> 1: a9100fbdd3698662    iris featureless       holdout 1
#> 2: 3fe90c093a9fdeab    iris       rpart       holdout 1
#> 3: cd3af67cd2425cb1   sonar featureless       holdout 1
#> 4: a3b3768bd595fd77   sonar       rpart       holdout 1

# Extract first ResampleResult
rr = bmr$resample_result(hash = rrs$hash[1])
print(rr)
#> <ResampleResult> of learner 'iris' on task 'featureless' with 1 iterations
#>  Measure Min. 1st Qu. Median Mean 3rd Qu. Max. Sd
#>     mmce 0.72    0.72   0.72 0.72    0.72 0.72 NA

# Extract predictions of first experiment of this resampling
head(as.data.table(rr$experiment(1)$prediction))
#>    row_id response  truth
#> 1:      2   setosa setosa
#> 2:      4   setosa setosa
#> 3:     10   setosa setosa
#> 4:     13   setosa setosa
#> 5:     19   setosa setosa
#> 6:     21   setosa setosa