add function to check for constant model while focussearch

R/doc_mbo_OptPath.R

@@ -18,6 +18,7 @@
 #'   }
 #' }
 #' \item{parego.weight}{Weight vector sampled for multipoint ParEGO}
+#' \item{constant.model}{If CMAES or Focussearch is used to optimize the infill crit, they can automatically determine if the model is constant in this iteration.}
 #' \item{...}{Depending on the chosen infill criterion there will be additional columns, e.g. \code{se} and \code{mean} for the Expected Improvement)}
 #' }
 #'

R/getExtras.R

@@ -66,6 +66,8 @@ getExtras = function(n, prop, train.time, control) {
       w = setNames(as.list(weight.mat[i, ]), paste0("parego.weight.", seq_col(weight.mat)))
       ex = c(ex, w)
     }
+    ex$constant.model = isTRUE(attr(prop$prop.points, "constant.model"))
+    # if we use asyn MBO store node information and evaluation starte
     ex$train.time = if (i == 1) train.time else NA_real_
     ex$prop.type = prop$prop.type[i]
     ex$propose.time = NA_real_

R/infillOptCMAES.R

@@ -56,8 +56,11 @@ infillOptCMAES = function(infill.crit, models, control, par.set, opt.path, desig
   if (is.infinite(result$best.fitness)) {
     warningf("Infill optimizer CMA-ES crashed. Random point generated instead.")
     res = t(sampleValue(par.set))
+    constant.model = TRUE
   } else {
     res = t(result$best.param)
+    constant.model = FALSE
   }
-  setColNames(as.data.frame(res), rep.pids)
+  res = setColNames(as.data.frame(res), rep.pids)
+  setAttribute(res, "constant.model", constant.model)
 }

R/infillOptFocus.R

@@ -8,6 +8,7 @@
 # See infillOptCMAES.R for interface explanation.
 infillOptFocus = function(infill.crit, models, control, par.set, opt.path, design, iter, ...) {
   global.y = Inf
+  constant.model = FALSE
 
   # restart the whole crap some times
   for (restart.iter in seq_len(control$infill.opt.restarts)) {
@@ -26,11 +27,15 @@ infillOptFocus = function(infill.crit, models, control, par.set, opt.path, desig
 
 
       # get current best value
-      local.index = getMinIndex(y, ties.method = "random")
+      local.index = getMinIndex(y, ties.method = "random", na.rm = TRUE)
       local.y = y[local.index]
       local.x.df = newdesign[local.index, , drop = FALSE]
       local.x.list = dfRowToList(recodeTypes(local.x.df, ps.local), ps.local, 1)
 
+      # check if the model just gives constant values
+      if (local.iter == 1 && length(unique(y)) == 1) 
+        constant.model = TRUE
+
       # if we found a new best value, store it
       if (local.y < global.y) {
         global.x.df = local.x.df
@@ -64,7 +69,7 @@ infillOptFocus = function(infill.crit, models, control, par.set, opt.path, desig
       })
     }
   }
-  recodeTypes(global.x.df, par.set)
+  setAttribute(recodeTypes(global.x.df, par.set), "constant.model", constant.model)
 }
 
 # as we operate on other types for the learner (ints are nums, logs are factors),

R/makeMBOControl.R

@@ -122,7 +122,7 @@ makeMBOControl = function(n.objectives = 1L,
   assertList(resample.measures, types = "Measure")
 
   assertString(output.num.format)
-
+ 
   control = makeS3Obj("MBOControl",
     n.objectives = n.objectives,
     propose.points = propose.points,

tests/testthat/test_mbo_km.R

@@ -72,3 +72,17 @@ test_that("mbo works with impute and failure model", {
   expect_true(!is.na(op$error.model[13L]))
   expect_true(!is.na(op$error.model[14L]))
 })
+
+test_that("mbo recognizes constant model prediction", {
+  des = testd.fsphere.2d
+  des$y = mean(apply(des, 1, testf.fsphere.2d))
+  # make sure model does not break, and we get a failure model
+  learner = makeLearner("regr.km", predict.type = "se")
+  ctrl = makeMBOControl()
+  ctrl = setMBOControlTermination(ctrl, iters = 4L)
+  ctrl = setMBOControlInfill(ctrl, opt.focussearch.points = 100L)
+  or = mbo(testf.fsphere.2d, des, learner = learner, control = ctrl)
+  expect_equal(getOptPathLength(or$opt.path), 14)
+  op = as.data.frame(or$opt.path)
+  expect_true(sum(op$constant.model)<5)
+})