added ocba and incumbent replication plus tests

R/intensifyOptState.R

@@ -10,7 +10,6 @@ intensifyOptState = function(opt.state) {
     )
 }
 
-
 intensifyIncumbent = function(opt.state) {
 
   # some intialization

tests/testthat/test_intensify.R

@@ -0,0 +1,93 @@
+# setwd(dir = "C:/Users/jumoo/Desktop/repos/mlrMBO")
+
+# load_all() # load everything 
+
+context("mbo intensification")
+
+test_that("mbo works with incumbent strategy", {
+  ps = makeNumericParamSet("x", 1L, -3, 3)
+  fun = smoof::makeSingleObjectiveFunction(
+	  fn = function(x) (x + 0.5)^2 + 5 * sin(3 * (x + 0.5)) + rnorm(1, sd = 2),
+	  par.set = ps, 
+	  noisy = TRUE
+	)
+
+  ctrl = makeMBOControl()
+  ctrl = setMBOControlTermination(ctrl, iters = 5L)
+	ctrl = setMBOControlNoisy(ctrl, method = "incumbent", incumbent.nchallengers = 2L)
+	or = mbo(fun, control = ctrl)
+	opdf = as.data.frame(or$opt.path)
+
+	# incumbent and proposed point are evaluated once in each iteration
+	expect_true(all(table(opdf$prop.type)[c("incumbent", "infill_cb")] == 5L))
+	# exactly 3 challengers are evaluated in each iteration (1 new point + 2 old points)  
+	expect_true(all(setDT(opdf)[opdf$prop.type == "challenger", .(length(unique(x))), by = "dob"]$V1 == 3L))
+
+	# incumbent of iteration i + 1 should be the best point of iteration i
+	opdfs = setDT(opdf)[, .(ymean = cumsum(y) / (1:.N), dob = dob), by = x]
+	best = opdfs[, .(xmin = x[which.min(ymean)]), by = .(dob)]
+  expect_true(all(best$xmin[1:5] == opdf[opdf$prop.type == "incumbent", ]$x))
+})
+
+
+test_that("mbo works with ocba replication strategy", {
+  ps = makeNumericParamSet("x", 1L, -3, 3)
+  fun = makeSingleObjectiveFunction(
+	  fn = function(x) (x + 0.5)^2 + 5 * sin(3 * (x + 0.5)) + rnorm(1, sd = 2),
+	  par.set = ps, 
+	  noisy = TRUE
+	)
+
+	ctrl = makeMBOControl()
+	ctrl = setMBOControlTermination(ctrl, iters = 5L)
+	ctrl = setMBOControlNoisy(ctrl, method = "ocba", ocba.budget = 12L, ocba.initial = 3L)
+
+	or = mbo(fun, control = ctrl)
+	opdf = as.data.frame(or$opt.path)
+
+	# in each iteration we do 12L replications
+	opdfs = setDT(opdf)[, .(N = sum(prop.type == "OCBA")), by = dob]
+	expect_true(all(opdfs[opdfs$dob > 0, "N"] == 12L))
+
+	# each point is evaluated 3L times at least
+	opdfs = setDT(opdf)[, .N, by = "x"]
+	expect_true(all(opdfs$N >= 3L))
+
+})
+
+
+test_that("mbo with replication works in multidimensional case", {
+
+  fun = makeNoisy(smoof::makeAckleyFunction(5L), noise = function(x) x^2)
+
+  ctrl = makeMBOControl()
+  ctrl = setMBOControlTermination(ctrl, iters = 3L)
+  ctrl = setMBOControlNoisy(ctrl, method = "ocba", ocba.budget = 12L, ocba.initial = 3L)
+
+  or = mbo(fun, control = ctrl)
+  opdf = as.data.frame(or$opt.path)
+
+  # in each iteration we do 12L replications
+  opdfs = setDT(opdf)[, .(N = sum(prop.type == "OCBA")), by = dob]
+  expect_true(all(opdfs[opdfs$dob > 0, "N"] == 12L))
+
+  # each point is evaluated 3L times at least
+  opdfs = setDT(opdf)[, .N, by = eval(paste("x", 1:5, sep = ""))]
+  expect_true(all(opdfs$N >= 3L))
+
+})
+
+
+test_that("per instance aggregation works for different functions", {
+
+  fun = makeNoisy(smoof::makeAckleyFunction(1L), noise = function(x) x^2)
+
+  ctrl = makeMBOControl()
+  ctrl = setMBOControlTermination(ctrl, iters = 3L)
+  ctrl = setMBOControlNoisy(ctrl, method = "incumbent", instance.aggregation = median)
+
+  or = mbo(fun, control = ctrl)
+  opdf = as.data.frame(or$opt.path)
+
+  expect_true(all(length(or$models[[1]]$subset) == length(unique(opdf$x))))
+})