handle failing models

DESCRIPTION

@@ -29,6 +29,7 @@ Suggests:
     randomForest,
     DiceKriging,
     rgenoud,
-    knitr
+    knitr,
+    gbm
 VignetteBuilder: knitr
 RoxygenNote: 6.0.1

R/generateHyperControl.R

@@ -60,7 +60,8 @@ generateHyperControl = function(task, par.config = NULL, learner = NULL, budget.
   } else if (
     all(getParamTypes(par.set) %in% c("numeric", "integer", "numericvector", "integervector")) &&
     getParamNr(par.set) * 4 < budget.evals * 0.75) {
-    mbo.control = mlrMBO::makeMBOControl(final.method = "best.predicted")
+    imputeWorst = function(x, y, opt.path, c = 2) c * max(getOptPathY(opt.path), na.rm = TRUE)
+    mbo.control = mlrMBO::makeMBOControl(final.method = "best.predicted", impute.y.fun = imputeWorst)
     mbo.control = mlrMBO::setMBOControlInfill(mbo.control, crit = mlrMBO::crit.eqi)
     mbo.control = mlrMBO::setMBOControlTermination(mbo.control, max.evals = budget.evals)
     mlr.control = makeTuneControlMBO(mbo.control = mbo.control, mbo.keep.result = TRUE)

R/hyperopt.R

@@ -53,6 +53,8 @@ hyperopt = function(task, learner = NULL, par.config = NULL, hyper.control = NUL
   if ("req.prob" %in% measures[[1]]$properties) {
     learner = setPredictType(learner, "prob")
   }
+  learner$config = insert(learner$config, list(on.learner.error = "warn"))
+
 
   tune.res = tuneParams(
     learner = learner,

tests/testthat/test_hyperopt.R

@@ -39,4 +39,16 @@ test_that("hyperopt works", {
   res3 = hyperopt(mini.task, par.config = par.config)
   expect_class(res3, "TuneResult")
   expect_class(res3$control, "TuneControlGrid")
+})
+
+test_that("hyperopt works with failing learners", {
+  mlr::configureMlr(show.info = FALSE, show.learner.output = FALSE)
+  # should trigger MBO
+  lrn = makeLearner("classif.gbm")
+  task = iris.task
+  res = hyperopt(task = iris.task, learner = lrn)
+  expect_class(res, "TuneResult")
+  expect_class(res$control, "TuneControlMBO")
+  expect_number(res$y, lower = 0, upper = 0.1)
+  expect_true(any(is.na(getOptPathY(res$opt.path))))
 })