For demonstration purposes, let's still try our ridge regression on a one-versus-rest method. To do this, create a MulticlassWrapper for a binary classification method. The classif.penalized.ridge method is from the penalized package, so make sure you have it installed:

    > ovr <- makeMulticlassWrapper("classif.penalized.ridge", 
      mcw.method = "onevsrest")

Now let's go ahead and create a wrapper for our classifier that creates a bagging resample of 10 iterations (it is the default) with replacement, sampling 70% of the observations and all of the input features:

    > bag.ovr = makeBaggingWrapper(ovr, bw.iters = 10, #default of 10
      bw.replace = TRUE, #default
      bw.size = 0.7,
      bw.feats = 1)

This can now be used to train our algorithm. Notice in the code I put mlr:: before train(). The reason is that caret also has a train() function, so we are specifying we want mlr train function and not caret's. Sometimes, if this is not done when both packages are loaded, you will end up with an egregious error:

    > set.seed(317)
    > fitOVR <- mlr::train(bag.ovr, wine.task)
    > predOVR <- predict(fitOVR, newdata = test)

Let's see how it did:

    > head(data.frame(predOVR))
       truth response
    60     2        2
    78     2        2
    79     2        2
    49     1        1
    19     1        1
    69     2        2

    > getConfMatrix(predOVR)
           predicted
    true     1  2  3  -SUM-
      1     58  0  0      0
      2      0 71  0      0
      3      0  0 57      0
      -SUM-  0  0  0      0

Again, it is just too easy. However, don't focus on the accuracy as much as the methodology of creating your classifier, tuning any parameters, and implementing a resampling strategy.