add multi objective BOHB optimizer

blackboxopt/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "5.3.8"
+__version__ = "6.0.0"
 
 from parameterspace import ParameterSpace
 

blackboxopt/optimizers/bohb.py

@@ -17,19 +17,22 @@
 from parameterspace import ParameterSpace
 
 from blackboxopt import Objective
+from blackboxopt.optimizers.staged.utils import greedy_promotion
 
 try:
     from blackboxopt.optimizers.staged.bohb import Sampler as BOHBSampler
     from blackboxopt.optimizers.staged.hyperband import create_hyperband_iteration
-    from blackboxopt.optimizers.staged.optimizer import StagedIterationOptimizer
+    from blackboxopt.optimizers.staged.optimizer import (
+        SingleObjectiveStagedIterationOptimizer,
+    )
 except ImportError as e:
     raise ImportError(
         "Unable to import BOHB optimizer specific dependencies. "
         + "Make sure to install blackboxopt[bohb]"
     ) from e
 
 
-class BOHB(StagedIterationOptimizer):
+class BOHB(SingleObjectiveStagedIterationOptimizer):
     def __init__(
         self,
         search_space: ParameterSpace,
@@ -124,11 +127,12 @@ def _create_new_iteration(self, iteration_index):
         `blackboxopt.optimizer.utils.staged_iteration.StagedIteration` object
         """
         return create_hyperband_iteration(
-            iteration_index,
-            self.min_fidelity,
-            self.max_fidelity,
-            self.eta,
-            self.config_sampler,
-            self.objective,
-            self.logger,
+            iteration_index=iteration_index,
+            min_fidelity=self.min_fidelity,
+            max_fidelity=self.max_fidelity,
+            eta=self.eta,
+            config_sampler=self.config_sampler,
+            objectives=[self.objective],
+            config_promotion_function=greedy_promotion,
+            logger=self.logger,
         )

blackboxopt/optimizers/hyperband.py

@@ -17,20 +17,23 @@
 from parameterspace import ParameterSpace
 
 from blackboxopt import Objective
+from blackboxopt.optimizers.staged.utils import greedy_promotion
 
 try:
     from blackboxopt.optimizers.staged.configuration_sampler import RandomSearchSampler
     from blackboxopt.optimizers.staged.hyperband import create_hyperband_iteration
     from blackboxopt.optimizers.staged.iteration import StagedIteration
-    from blackboxopt.optimizers.staged.optimizer import StagedIterationOptimizer
+    from blackboxopt.optimizers.staged.optimizer import (
+        SingleObjectiveStagedIterationOptimizer,
+    )
 except ImportError as e:
     raise ImportError(
         "Unable to import Hyperband optimizer specific dependencies. "
         + "Make sure to install blackboxopt[hyperband]"
     ) from e
 
 
-class Hyperband(StagedIterationOptimizer):
+class Hyperband(SingleObjectiveStagedIterationOptimizer):
     def __init__(
         self,
         search_space: ParameterSpace,
@@ -78,11 +81,12 @@ def _create_new_iteration(self, iteration_index: int) -> StagedIteration:
         `blackboxopt.optimizer.staged.iteration.StagedIteration` object
         """
         return create_hyperband_iteration(
-            iteration_index,
-            self.min_fidelity,
-            self.max_fidelity,
-            self.eta,
-            self.config_sampler,
-            self.objective,
-            self.logger,
+            iteration_index=iteration_index,
+            min_fidelity=self.min_fidelity,
+            max_fidelity=self.max_fidelity,
+            eta=self.eta,
+            config_sampler=self.config_sampler,
+            objectives=[self.objective],
+            config_promotion_function=greedy_promotion,
+            logger=self.logger,
         )

blackboxopt/optimizers/mobohb.py

@@ -0,0 +1,140 @@
+# Copyright (c) 2020 - for information on the respective copyright owner
+# see the NOTICE file and/or the repository https://github.com/boschresearch/blackboxopt
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import logging
+from typing import List, Optional
+
+import numpy as np
+from parameterspace import ParameterSpace
+
+from blackboxopt import Objective
+
+try:
+    from blackboxopt.optimizers.staged.hyperband import create_hyperband_iteration
+    from blackboxopt.optimizers.staged.iteration import Datum
+    from blackboxopt.optimizers.staged.mobohb import Sampler, argsort_nondominated
+    from blackboxopt.optimizers.staged.optimizer import (
+        MultiObjectiveStagedIterationOptimizer,
+    )
+except ImportError as e:
+    raise ImportError(
+        "Unable to import MO-BOHB optimizer specific dependencies. "
+        + "Make sure to install blackboxopt[mobohb]"
+    ) from e
+
+
+def _nondominated_promotion(data: List[Datum], num_configs: int) -> list:
+    """Promote configurations to the next stage by multi-objective nondominated rank.
+
+    Args:
+        data: List with all successful evaluations for this stage. All failed
+            configurations have already been removed.
+        num_configs: Maximum number of configurations to be promoted.
+
+    Returns:
+        List of the config_keys to be evaluated on the next higher fidelity.
+    """
+    if not data:
+        return []
+
+    losses = np.stack([d.losses for d in data])
+    n = min(num_configs, len(data))
+    selected = argsort_nondominated(losses)[:n]
+    return [data[i].config_key for i in selected]
+
+
+class MOBOHB(MultiObjectiveStagedIterationOptimizer):
+    def __init__(
+        self,
+        search_space: ParameterSpace,
+        objectives: List[Objective],
+        min_fidelity: float,
+        max_fidelity: float,
+        num_iterations: int,
+        eta: float = 3.0,
+        top_n_percent: int = 15,
+        min_samples_in_model: Optional[int] = None,
+        num_samples: int = 64,
+        random_fraction: float = 1 / 3,
+        bandwidth_factor: float = 3.0,
+        min_bandwidth: float = 1e-3,
+        seed: Optional[int] = None,
+        logger: Optional[logging.Logger] = None,
+    ):
+        """Multi-objective BOHB Optimizer.
+
+        MO-BOHB extends BOHB to multiple objectives by replacing its single-objective,
+        scalar-loss based good/bad split with a multi-objective one: configurations are
+        ranked by nondominated (Pareto) front and, within the boundary front, by
+        crowding distance (NSGA-II style). The same ranking drives the successive
+        halving promotion between stages. The Bayesian optimization component (kernel
+        density estimators and expected improvement sampling) is reused from BOHB
+        unchanged.
+
+        Args:
+            search_space: [description]
+            objectives: The objectives of the optimization.
+            min_fidelity: The smallest fidelity value that is still meaningful.
+                Must be strictly greater than zero!
+            max_fidelity: The largest fidelity value used during the optimization.
+                Must not be smaller than `min_fidelity`.
+            num_iterations: The number of iterations that the optimizer will run.
+            eta: Scaling parameter to control the aggressiveness of Hyperband's racing.
+            top_n_percent: Determines the percentile of configurations that will be
+                used as training data for the kernel density estimator of the good
+                configurations, e.g. if set to 10 the best 10% configurations (by
+                nondominated rank) will be considered for training.
+            min_samples_in_model: Minimum number of datapoints needed to fit a model.
+            num_samples: Number of samples drawn to optimize EI via sampling.
+            random_fraction: Fraction of random configurations returned.
+            bandwidth_factor: Widens the bandwidth for contiuous parameters for
+                proposed points to optimize EI
+            min_bandwidth: to keep diversity, even when all (good) samples have the
+                same value for one of the parameters, a minimum bandwidth
+                (reasonable default: 1e-3) is used instead of zero.
+            seed: [description]
+            logger: [description]
+        """
+        if min_samples_in_model is None:
+            min_samples_in_model = 3 * len(search_space)
+
+        self.min_fidelity = min_fidelity
+        self.max_fidelity = max_fidelity
+        self.eta = eta
+
+        self.config_sampler = Sampler(
+            search_space=search_space,
+            objectives=objectives,
+            min_samples_in_model=min_samples_in_model,
+            top_n_percent=top_n_percent,
+            num_samples=num_samples,
+            random_fraction=random_fraction,
+            bandwidth_factor=bandwidth_factor,
+            min_bandwidth=min_bandwidth,
+            seed=seed,
+        )
+
+        super().__init__(
+            search_space=search_space,
+            objectives=objectives,
+            num_iterations=num_iterations,
+            seed=seed,
+            logger=logger,
+        )
+
+    def _create_new_iteration(self, iteration_index):
+        """Optimizer specific way to create a new
+        `blackboxopt.optimizer.utils.staged_iteration.StagedIteration` object
+        """
+        return create_hyperband_iteration(
+            iteration_index=iteration_index,
+            min_fidelity=self.min_fidelity,
+            max_fidelity=self.max_fidelity,
+            eta=self.eta,
+            config_sampler=self.config_sampler,
+            objectives=self.objectives,
+            logger=self.logger,
+            config_promotion_function=_nondominated_promotion,
+        )