some more compatibility with new paradox version

[mb706]

No description provided.

[mb706]

@mllg
this fails with the new paradox version in this test:

mlr3/tests/testthat/test_benchmark.R

Lines 473 to 477 in 78ad4dd

	design = data.table(task = tasks, learner = learners, resampling = resamplings, param_values = list(list(list(cp = 0.1), list(minbucket = 2))))
	bmr = benchmark(design)
	expect_equal(bmr$learners$learner[[1]]$param_set$values, list(xval = 0, minsplit = 12, minbucket = 2))
	expect_equal(bmr$learners$learner[[2]]$param_set$values, list(xval = 0, minsplit = 12, cp = 0.1))

However, I don't get it: the param_values argument in the design is list(cp = 0.1) for the first learner, yet the test expects cp = 0.1 in the second row of the result. Why is that?

[mb706]

is the row order in the benchmark result just coincidental and depends on the hash of objects? in this case, I would adapt this specific test to be row order independent.

[mllg]

Yes, this is a bad test; you should reorder the names (as you already did in this draft) and then not rely on the order, e.g.:

  trained = bmr$learners$learner
  ii = which(map_lgl(trained, function(x) "cp" %in% names(x$param_set$values))) # find learner with cp
  expect_count(ii)

  expect_equal(trained[[ii]]$param_set$values, list(xval = 0, minsplit = 12, cp = 0.1))
  expect_equal(trained[[-ii]]$param_set$values, list(xval = 0, minsplit = 12, minbucket = 2))

[mb706]

@mllg this one can be merged