tests/testthat/test_multiobj_parego.R

In berndbischl/mlrMBO: Bayesian Optimization and Model-Based Optimization of Expensive Black-Box Functions

context("multi-objective: parego")

test_that("mbo parego works", {

  # Test normal run
  learner = makeLearner("regr.km", predict.type = "se", nugget.estim = TRUE)
  ctrl = makeMBOControl(n.objectives = 2L)
  ctrl = setMBOControlTermination(ctrl, iters = 5L)
  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 10)
  ctrl = setMBOControlMultiObj(ctrl, method = "parego", parego.s = 100)
  or = mbo(testfmco1, testdesmco1, learner = learner, control = ctrl)
  expect_output(print(or), "Optimization path")
  expect_matrix(or$pareto.front, mode = "numeric", any.missing = FALSE)

  # Test margin points
  ctrl = makeMBOControl(n.objectives = 2, propose.points = 2L)
  ctrl = setMBOControlTermination(ctrl, iters = 5L)
  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 5)
  ctrl = setMBOControlMultiObj(ctrl, method = "parego", parego.s = 100, parego.use.margin.points = c(TRUE, TRUE))
  or = mbo(testfmco1, testdesmco1, learner = learner, control = ctrl)
  w = as.data.frame(or$opt.path)[-(1:10), c("parego.weight.1", "parego.weight.2")]
  expect_true(all(w == 0 | w == 1))
  expect_equal(1 - w[, 1], w[, 2])

  # Test margin points
  ctrl = makeMBOControl(n.objectives = 2, propose.points = 2L)
  ctrl = setMBOControlTermination(ctrl, iters = 5L)
  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 5)
  ctrl = setMBOControlMultiObj(ctrl, method = "parego", parego.s = 100, parego.use.margin.points = c(TRUE, TRUE))
  or = mbo(testfmco1, testdesmco1, learner = learner, control = ctrl)
  w = as.data.frame(or$opt.path)[-(1:10), c("parego.weight.1", "parego.weight.2")]
  expect_true(all(w == 0 | w == 1))
  expect_equal(1 - w[, 1], w[, 2])

  # Test wrong dimension
  ctrl = makeMBOControl(n.objectives = 3)
  ctrl = setMBOControlTermination(ctrl, iters = 5L)
  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 10L)
  ctrl = setMBOControlMultiObj(ctrl, method = "parego", parego.s = 100)
  expect_error(mbo(testfmco1, testdesmco1, learner = learner, control = ctrl), "Objective function has")

  # Test multippoint
  ctrl = makeMBOControl(n.objectives = 2, propose.points = 5L)
  ctrl = setMBOControlTermination(ctrl, iters = 1L)
  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 10L)
  ctrl = setMBOControlMultiObj(ctrl, method = "parego", parego.s = 100)
  or = mbo(testfmco1, testdesmco1, learner = learner, control = ctrl)

  # check used weights
  w = as.data.frame(or$opt.path)[, c("parego.weight.1", "parego.weight.2")]
  expect_true(all(rowSums(w[-(1:10),]) == 1))
  expect_numeric(w[-(1:10), 1], any.missing = FALSE, unique = TRUE)
  expect_numeric(w[-(1:10), 2], any.missing = FALSE, unique = TRUE)

  # Multipoint with minimization and maximization
  ctrl = makeMBOControl(n.objectives = 2, propose.points = 5L)
  ctrl = setMBOControlTermination(ctrl, iters = 5L)
  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 10L)
  ctrl = setMBOControlMultiObj(ctrl, method = "parego", parego.s = 100)
  or = mbo(testfmco2, testdesmco2, learner = learner, control = ctrl)

  expect_matrix(or$pareto.front, mode = "numeric", any.missing = FALSE)
  # FIXME: make sure to properly test here. minimum and max reached
  # expect_true(all(or$pareto.front)[1,]^2) < 1e-4)
})