BNH and TNK (#424)

* add new multiobjective benchmarks

* update notebook

bofire/benchmarks/api.py

@@ -4,7 +4,15 @@
 from bofire.benchmarks.benchmark import Benchmark, GenericBenchmark
 from bofire.benchmarks.detergent import Detergent
 from bofire.benchmarks.hyperopt import Hyperopt
-from bofire.benchmarks.multi import C2DTLZ2, DTLZ2, ZDT1, CrossCoupling, SnarBenchmark
+from bofire.benchmarks.multi import (
+ BNH,
+ C2DTLZ2,
+ DTLZ2,
+ TNK,
+ ZDT1,
+ CrossCoupling,
+ SnarBenchmark,
+)
 from bofire.benchmarks.single import (
  Ackley,
  Branin,
@@ -14,7 +22,9 @@
  MultiTaskHimmelblau,
 )
 
-AnyMultiBenchmark = Union[C2DTLZ2, Detergent, DTLZ2, ZDT1, CrossCoupling, SnarBenchmark]
+AnyMultiBenchmark = Union[
+ C2DTLZ2, Detergent, DTLZ2, ZDT1, CrossCoupling, SnarBenchmark, BNH, TNK
+]
 AnySingleBenchmark = Union[
  Ackley, Branin, Branin30, Hartmann, Himmelblau, MultiTaskHimmelblau
 ]

bofire/benchmarks/multi.py

@@ -28,6 +28,7 @@
  MaximizeObjective,
  MaximizeSigmoidObjective,
  MinimizeObjective,
+ MinimizeSigmoidObjective,
 )
 from bofire.data_models.surrogates.api import SingleTaskGPSurrogate
 from bofire.utils.torch_tools import tkwargs
@@ -128,6 +129,97 @@ def _f(self, candidates: pd.DataFrame) -> pd.DataFrame:
  return Y
 
 
+class BNH(Benchmark):
+ def __init__(self, constraints: bool = True, **kwargs):
+ super().__init__(**kwargs)
+ self.constraints = constraints
+
+ self._domain = Domain(
+ inputs=Inputs(
+ features=[
+ ContinuousInput(key="x1", bounds=(0, 5)),
+ ContinuousInput(key="x2", bounds=(0, 3)),
+ ]
+ ),
+ outputs=Outputs(
+ features=[
+ ContinuousOutput(key="f1", objective=MinimizeObjective(w=1.0)),
+ ContinuousOutput(key="f2", objective=MinimizeObjective(w=1.0)),
+ ]
+ ),
+ )
+ if self.constraints:
+ self._domain.outputs.features.append( # type: ignore
+ ContinuousOutput(
+ key="c1",
+ objective=MinimizeSigmoidObjective(tp=25, steepness=1000),
+ )
+ )
+ self._domain.outputs.features.append( # type: ignore
+ ContinuousOutput(
+ key="c2",
+ objective=MaximizeSigmoidObjective(tp=7.7, steepness=1000),
+ ),
+ )
+
+ def _f(self, candidates: pd.DataFrame) -> pd.DataFrame:
+ experiments = candidates.eval("f1=4*x1**2 + 4*x2**2", inplace=False)
+ experiments = experiments.eval("f2=(x1-5)**2 + (x2-5)**2", inplace=False)
+ experiments["valid_f1"] = 1
+ experiments["valid_f2"] = 1
+ if not self.constraints:
+ return experiments[["f1", "f2", "valid_f1", "valid_f2"]].copy()
+ experiments = experiments.eval("c1=(x1-5)**2 + x2**2", inplace=False)
+ experiments = experiments.eval("c2=(x1-8)**2 + (x2+3)**2", inplace=False)
+ experiments["valid_c1"] = 1
+ experiments["valid_c2"] = 1
+ return experiments[
+ ["f1", "f2", "c1", "c2", "valid_c1", "valid_c2", "valid_f1", "valid_f2"]
+ ].copy()
+
+
+class TNK(Benchmark):
+ def __init__(self, **kwargs):
+ super().__init__(**kwargs)
+ self._domain = Domain(
+ inputs=Inputs(
+ features=[
+ ContinuousInput(key="x1", bounds=(0, math.pi)),
+ ContinuousInput(key="x2", bounds=(0, math.pi)),
+ ]
+ ),
+ outputs=Outputs(
+ features=[
+ ContinuousOutput(key="f1", objective=MinimizeObjective(w=1.0)),
+ ContinuousOutput(key="f2", objective=MinimizeObjective(w=1.0)),
+ ContinuousOutput(
+ key="c1",
+ objective=MaximizeSigmoidObjective(tp=0.0, steepness=500),
+ ),
+ ContinuousOutput(
+ key="c2",
+ objective=MinimizeSigmoidObjective(tp=0.5, steepness=500),
+ ),
+ ]
+ ),
+ )
+
+ def _f(self, candidates: pd.DataFrame) -> pd.DataFrame:
+ experiments = candidates.eval("f1=x1", inplace=False)
+ experiments = experiments.eval("f2=x2", inplace=False)
+ experiments = experiments.eval(
+ "c1=x1**2 + x2**2 -1 -0.1*cos(16*arctan(x1/x2))", inplace=False
+ )
+ experiments = experiments.eval("c2=(x1-0.5)**2+(x2-0.5)**2", inplace=False)
+ experiments["valid_c1"] = 1
+ experiments["valid_c2"] = 1
+ experiments["valid_f1"] = 1
+ experiments["valid_f2"] = 1
+ return experiments[
+ ["f1", "f2", "c1", "c2", "valid_c1", "valid_c2", "valid_f1", "valid_f2"]
+ ].copy()
+
+
 class C2DTLZ2(DTLZ2):
  """Constrained DTLZ2 benchmark function. Taken from
  https://github.com/pytorch/botorch/blob/main/botorch/test_functions/multi_objective.py.

tests/bofire/benchmarks/test_multi.py

@@ -1,6 +1,14 @@
 import pytest
 
-from bofire.benchmarks.multi import C2DTLZ2, DTLZ2, ZDT1, CrossCoupling, SnarBenchmark
+from bofire.benchmarks.multi import (
+ BNH,
+ C2DTLZ2,
+ DTLZ2,
+ TNK,
+ ZDT1,
+ CrossCoupling,
+ SnarBenchmark,
+)
 
 
 @pytest.mark.parametrize(
@@ -24,6 +32,10 @@
  ),
  (C2DTLZ2, True, {"dim": 4}),
  (C2DTLZ2, False, {"dim": 4}),
+ (BNH, False, {"constraints": True}),
+ (BNH, False, {"constraints": False}),
+ (TNK, False, {}),
+ (TNK, True, {}),
  ],
 )
 def test_multi_objective_benchmarks(cls_benchmark, return_complete, kwargs):
@@ -43,6 +55,7 @@ def test_multi_objective_benchmarks(cls_benchmark, return_complete, kwargs):
 
  # Define expected number of output variables
  expected_output_variables = len(benchmark_function.domain.outputs) * 2
+ print(Y.shape, expected_output_variables)
  # Check, whether expected number of output variables match the actual number
  if return_complete:
  assert Y.shape == (