Implementation of HEBO Optimizer

.github/workflows/r-cmd-check.yml

@@ -54,11 +54,25 @@ jobs:
         with:
           r-version: ${{ matrix.config.r }}
 
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
       - uses: r-lib/actions/setup-r-dependencies@v2
         with:
           extra-packages: any::rcmdcheck
           needs: check
 
+      - name: Create Python virtual environments
+        run: |
+          python3 -m venv ~/.virtualenvs/r-hebo
+          ~/.virtualenvs/r-hebo/bin/pip install --upgrade pip --quiet
+          ~/.virtualenvs/r-hebo/bin/pip install torch --index-url https://download.pytorch.org/whl/cpu --quiet
+          ~/.virtualenvs/r-hebo/bin/pip install hebo --quiet
+          python3 -m venv ~/.virtualenvs/r-smac
+          ~/.virtualenvs/r-smac/bin/pip install --upgrade pip --quiet
+          ~/.virtualenvs/r-smac/bin/pip install smac ConfigSpace --quiet
+
       - uses: mxschmitt/action-tmate@v3
         if: ${{ github.event_name == 'workflow_dispatch' && inputs.debug_enabled }}
         with:

DESCRIPTION

@@ -11,7 +11,8 @@ Authors@R: c(
     person("Bernd", "Bischl", , "bernd_bischl@gmx.net", role = "aut",
            comment = c(ORCID = "0000-0001-6002-6980")),
     person("Martin", "Binder", , "martin.binder@mail.com", role = "aut"),
-    person("Olaf", "Mersmann", , "olafm@statistik.tu-dortmund.de", role = "ctb")
+    person("Olaf", "Mersmann", , "olafm@statistik.tu-dortmund.de", role = "ctb"),
+    person("Alexander", "Winterstetter", "alexanderwinterstetter@gmail.de", role = "ctb")
   )
 Description: Features highly configurable search spaces via the 'paradox'
     package and optimizes every user-defined objective function. The
@@ -35,6 +36,7 @@ Imports:
     R6
 Suggests:
     adagio,
+    callr,
     emoa,
     GenSA,
     irace (>= 4.0.0),
@@ -44,9 +46,13 @@ Suggests:
     processx,
     progressr,
     redux,
+    reticulate,
     RhpcBLASctl,
     rush (>= 1.0.0),
-    testthat (>= 3.0.0)
+    testthat (>= 3.0.0),
+    withr
+Remotes:
+    mlr-org/paradox
 Config/testthat/edition: 3
 Config/testthat/parallel: false
 Encoding: UTF-8
@@ -91,10 +97,12 @@ Collate:
     'OptimizerBatchFocusSearch.R'
     'OptimizerBatchGenSA.R'
     'OptimizerBatchGridSearch.R'
+    'OptimizerBatchHEBO.R'
     'OptimizerBatchIrace.R'
     'OptimizerBatchLocalSearch.R'
     'OptimizerBatchNLoptr.R'
     'OptimizerBatchRandomSearch.R'
+    'OptimizerBatchSmac3.R'
     'Progressor.R'
     'mlr_terminators.R'
     'Terminator.R'
@@ -114,6 +122,7 @@ Collate:
     'bibentries.R'
     'conditions.R'
     'helper.R'
+    'helper_hebo.R'
     'local_search.R'
     'mlr_callbacks.R'
     'mlr_test_functions.R'

NAMESPACE

@@ -48,10 +48,12 @@ export(OptimizerBatchDesignPoints)
 export(OptimizerBatchFocusSearch)
 export(OptimizerBatchGenSA)
 export(OptimizerBatchGridSearch)
+export(OptimizerBatchHEBO)
 export(OptimizerBatchIrace)
 export(OptimizerBatchLocalSearch)
 export(OptimizerBatchNLoptr)
 export(OptimizerBatchRandomSearch)
+export(OptimizerBatchSmac3)
 export(Terminator)
 export(TerminatorClockTime)
 export(TerminatorCombo)

R/OptimizerAsync.R

@@ -20,13 +20,13 @@
 #' There are three types of workers:
 #'
 #' - "mirai": Workers are started with \CRANpkg{mirai} on local or remote machines.
-#'   See `$start_workers()` in [Rush] for more details.
+#'   See `$start_workers()` in [rush::Rush] for more details.
 #'   [mirai::daemons()] must be created before starting the optimization.
 #' - "processx": Workers are started as local processes with \CRANpkg{processx}.
 
-#'   See `$start_local_workers()` in [Rush] for more details.
+#'   See `$start_local_workers()` in [rush::Rush] for more details.
 #' - "script": Workers are started by the user with a custom script.
-#'   See `$create_worker_script()` in [Rush] for more details.
+#'   See `$create_worker_script()` in [rush::Rush] for more details.
 #'
 #' The workers are started when the `$optimize()` method is called.
 #' The main process waits until at least one worker is running.

R/OptimizerBatchHEBO.R

@@ -0,0 +1,156 @@
+#' @title Heteroscedastic evolutionary bayesian optimization
+#'
+#' @export
+#'
+
+OptimizerBatchHEBO = R6Class(
+  "OptimizerBatchHEBO",
+  inherit = OptimizerBatch,
+  public = list(
+    initialize = function() {
+      param_set = ps(
+        # optimizer interface
+        n_suggestions = p_int(lower = 1L, init = 1L), # remove as HEBO only handles single crit
+        n_init = p_int(lower = 1L),
+        seed = p_int(lower = 0L),
+
+        # surrogate model
+        surrogate = p_fct(levels = c("gp", "rf"), init = "gp"),
+        rf_n_estimators = p_int(lower = 1L, depends = quote(surrogate == "rf")),
+        gp_lr = p_dbl(lower = 0, depends = quote(surrogate == "gp")),
+        gp_num_epochs = p_int(lower = 1L, depends = quote(surrogate == "gp")),
+        gp_noise_free = p_lgl(depends = quote(surrogate == "gp")),
+        gp_noise_lb = p_dbl(lower = 0, depends = quote(surrogate == "gp")),
+        gp_pred_likeli = p_lgl(depends = quote(surrogate == "gp")),
+
+        # acquisition function
+        acq_function = p_fct(levels = c("mace", "lcb"), init = "mace"),
+
+        # evolutionary search for acquisition optimization
+        es = p_fct(
+          levels = c(
+            "ga",
+            "brkga",
+            "de",
+            "nelder-mead",
+            "pattern-search",
+            "cmaes",
+            "pso",
+            "nsga2",
+            "rnsga2",
+            "nsga3",
+            "unsga3",
+            "rnsga3",
+            "moead",
+            "ctaea"
+          ),
+          init = "nsga2"
+        )
+      )
+
+      super$initialize(
+        id = "hebo",
+        param_set = param_set,
+        param_classes = c("ParamDbl", "ParamInt", "ParamFct", "ParamLgl"),
+        properties = c("single-crit"),
+        packages = "reticulate",
+        label = "Heteroscedastic Evolutionary Bayesian Optimization",
+        man = "bbotk::mlr_optimizers_hebo"
+      )
+    }
+  ),
+
+  private = list(
+    .optimize = function(inst) {
+      assert_python_packages(c("hebo"))
+      hebo = reticulate::import("hebo")
+
+      pv = self$param_set$values
+
+      search_space = inst$search_space
+      space = paramset_to_hebo_space(search_space)
+      optimizer_args = list(
+        space = space,
+        rand_sample = as.integer(pv$n_init %??% (1L + search_space$length)),
+        es = pv$es %??% "nsga2"
+      )
+
+      # add nn; evtl boosting
+      # add unit tests see SMAC3
+      if (!is.null(pv$surrogate)) {
+        if (pv$surrogate == "rf") {
+          optimizer_args$model_name = "rf"
+          optimizer_args$model_config = list(
+            n_estimators = as.integer(pv$rf_n_estimators %??% 20L)
+          )
+        } else if (pv$surrogate == "gp") {
+          optimizer_args$model_name = "gp"
+          optimizer_args$model_config = list(
+            lr = pv$gp_lr %??% 0.01,
+            num_epochs = as.integer(pv$gp_num_epochs %??% 100L),
+            verbose = pv$gp_verbose %??% FALSE,
+            print_every = as.integer(pv$gp_print_every %??% 10L),
+            noise_free = pv$gp_noise_free %??% FALSE,
+            noise_lb = pv$gp_noise_lb %??% 8e-4,
+            pred_likeli = pv$gp_pred_likeli %??% FALSE
+          )
+        }
+      }
+
+      if (!is.null(pv$acq_function)) {
+        acq_mod = hebo$acquisitions$acq
+        optimizer_args$acq_cls = switch(
+          pv$acq_function,
+          "lcb" = acq_mod$LCB,
+          "mace" = acq_mod$MACE,
+          stopf(
+            "Unsupported HEBO acquisition function '%s'. Supported values are 'mace' and 'lcb'.",
+            pv$acq_function
+          )
+        )
+      }
+
+      if (!is.null(pv$seed)) {
+        optimizer_args$scramble_seed = as.integer(pv$seed)
+      }
+
+      hebo_optimizer = invoke(hebo$optimizers$hebo$HEBO, .args = optimizer_args)
+
+      repeat {
+        if (inst$is_terminated) {
+          break
+        }
+
+        # tryCatch muss noch etwas ausgestaltet werden, siehe folgender Kommentar
+        # hebo$optimizers$hebo$HEBO()$quasi_sample --- default HEBO sampler evtl notwendig, wenn suggest fehlschlägt; useful wenn SM zusammenbricht; wrappen in Trycatch von suggest funktion falls das passiert
+        # ISSUE 1 - HEBO suggest doesnt expose metadata as SMAC3
+        # trial_info exposes paraeters as a pandas DataFrame
+        trial_info = tryCatch(
+          hebo_optimizer$suggest(n_suggestions = as.integer(pv$n_suggestions %??% 1L)),
+          # rausfinden wie HEBO suggest verwendet um mehrere configs oder einzelne configs zu kriegen; gibt mehrere Möglichkeiten das zu machen bei MBO interessant wie HEBO das macht
+          error = function(e) NULL
+        )
+        if (is.null(trial_info)) {
+          break
+        }
+
+        # convert pandas DataFrame to R data.frame-like object
+        trial_info_dt = setDT(reticulate::py_to_r(trial_info))
+
+        # evaluation of that config
+        res = inst$eval_batch(trial_info_dt)
+
+        # evaluated target columns are turned into a numeric matrix
+        y = as.matrix(res[, inst$archive$cols_y, with = FALSE])
+
+        # signs are flipped
+        # ist es möglich bei HEBO auch multi-objective zu machen, also dass da mehrere y zurückkomen; das auch wichtig für denOptimizer ob er nicht nur SingleCrit oder auch MultiCrit kann
+        y = y * matrix(inst$objective_multiplicator, nrow = nrow(y), ncol = ncol(y), byrow = TRUE)
+
+        hebo_optimizer$observe(trial_info, y)
+      }
+    }
+  )
+)
+
+mlr_optimizers$add("hebo", OptimizerBatchHEBO)