Example for using a mixture use case in a discrete searchspace¶
Example for imposing sum constraints for discrete parameters. The constraints simulate a situation where we want to mix up to three solvents. However, their respective fractions need to sum up to 100. Also, the solvents should never be chosen twice, which requires various other constraints.
This example assumes some basic familiarity with using BayBE.
We thus refer to campaign for a basic example.
Necessary imports for this example¶
import math
import os
import numpy as np
from baybe import Campaign
from baybe.constraints import (
DiscreteDependenciesConstraint,
DiscreteNoLabelDuplicatesConstraint,
DiscretePermutationInvarianceConstraint,
DiscreteSumConstraint,
ThresholdCondition,
)
from baybe.objectives import SingleTargetObjective
from baybe.parameters import NumericalDiscreteParameter, SubstanceParameter
from baybe.searchspace import SearchSpace
from baybe.targets import NumericalTarget
from baybe.utils.dataframe import add_fake_measurements
Experiment setup¶
This parameter denotes the tolerance with regard to the calculation of the sum.
SUM_TOLERANCE = 1.0
SMOKE_TEST = "SMOKE_TEST" in os.environ
# This parameter denotes the resolution of the discretization of the parameters
RESOLUTION = 5 if SMOKE_TEST else 12
dict_solvents = {
"water": "O",
"C1": "C",
"C2": "CC",
"C3": "CCC",
}
solvent1 = SubstanceParameter(name="Solv1", data=dict_solvents, encoding="MORDRED")
solvent2 = SubstanceParameter(name="Solv2", data=dict_solvents, encoding="MORDRED")
solvent3 = SubstanceParameter(name="Solv3", data=dict_solvents, encoding="MORDRED")
Parameters for representing the fraction.
fraction1 = NumericalDiscreteParameter(
name="Frac1", values=list(np.linspace(0, 100, RESOLUTION)), tolerance=0.2
)
fraction2 = NumericalDiscreteParameter(
name="Frac2", values=list(np.linspace(0, 100, RESOLUTION)), tolerance=0.2
)
fraction3 = NumericalDiscreteParameter(
name="Frac3", values=list(np.linspace(0, 100, RESOLUTION)), tolerance=0.2
)
parameters = [solvent1, solvent2, solvent3, fraction1, fraction2, fraction3]
Creating the constraint¶
Since the constraints are required for the creation of the searchspace, we create
them next.
Note that we need a PermutationInvarianceConstraint here.
The reason is that constraints are normally applied in a specific order.
However, the fractions should be invariant under permutations.
We thus require an explicit constraint for this.
perm_inv_constraint = DiscretePermutationInvarianceConstraint(
parameters=["Solv1", "Solv2", "Solv3"],
dependencies=DiscreteDependenciesConstraint(
parameters=["Frac1", "Frac2", "Frac3"],
conditions=[
ThresholdCondition(threshold=0.0, operator=">"),
ThresholdCondition(threshold=0.0, operator=">"),
ThresholdCondition(threshold=0.0, operator=">"),
],
affected_parameters=[["Solv1"], ["Solv2"], ["Solv3"]],
),
)
This is now the actual sum constraint
sum_constraint = DiscreteSumConstraint(
parameters=["Frac1", "Frac2", "Frac3"],
condition=ThresholdCondition(threshold=100, operator="=", tolerance=SUM_TOLERANCE),
)
The permutation invariance might create duplicate labels. We thus include a constraint to remove them.
no_duplicates_constraint = DiscreteNoLabelDuplicatesConstraint(
parameters=["Solv1", "Solv2", "Solv3"]
)
constraints = [perm_inv_constraint, sum_constraint, no_duplicates_constraint]
Creating the searchspace and the objective¶
searchspace = SearchSpace.from_product(parameters=parameters, constraints=constraints)
________________________________________________________________________________
[Memory] Calling baybe.utils.chemistry._smiles_to_mordred_features...
_smiles_to_mordred_features('C')
_______________________________________smiles_to_mordred_features - 0.0s, 0.0min
________________________________________________________________________________
[Memory] Calling baybe.utils.chemistry._smiles_to_mordred_features...
_smiles_to_mordred_features('CC')
_______________________________________smiles_to_mordred_features - 0.0s, 0.0min
objective = SingleTargetObjective(target=NumericalTarget(name="Target_1", mode="MAX"))
Creating and printing the campaign¶
campaign = Campaign(searchspace=searchspace, objective=objective)
print(campaign)
Campaign
Meta Data
Batches done: 0
Fits done: 0
SearchSpace
Search Space Type: DISCRETE
SubspaceDiscrete
Discrete Parameters
Name Type Num_Values Encoding
0 Frac1 NumericalDis... 5 None
1 Frac2 NumericalDis... 5 None
2 Frac3 NumericalDis... 5 None
3 Solv1 SubstancePar... 4 SubstanceEnc...
4 Solv2 SubstancePar... 4 SubstanceEnc...
5 Solv3 SubstancePar... 4 SubstanceEnc...
Experimental Representation
Solv1 Solv2 ... Frac2 Frac3
0 C3 C2 ... 0.0 100.0
1 C3 C2 ... 25.0 75.0
2 C3 C2 ... 50.0 50.0
.. ... ... ... ... ...
31 water C3 ... 25.0 50.0
32 water C3 ... 50.0 25.0
33 water C3 ... 25.0 25.0
[34 rows x 6 columns]
Meta Data
was_recommended: 0/34
was_measured: 0/34
dont_recommend: 0/34
Constraints
Type Affected_Paramet
0 DiscreteNoLa... [Solv1, Solv...
1 DiscreteSumC... [Frac1, Frac...
2 DiscretePerm... [Solv1, Solv...
Computational Representation
Frac1 Frac2 ... Solv3_MORDRED_AT Solv3_MORDRED_AA
0 0.0 0.0 ... 0.000 -36.020
1 0.0 25.0 ... 0.000 -36.020
2 0.0 50.0 ... 0.000 -36.020
.. ... ... ... ... ...
31 25.0 25.0 ... 0.005 -18.091
32 25.0 50.0 ... 0.005 -18.091
33 50.0 25.0 ... 0.005 -18.091
[34 rows x 15 columns]
Objective
Type: SingleTargetObjective
Targets
Type Name ... Upper_Bound Transformation
0 NumericalTarget Target_1 ... inf None
[1 rows x 6 columns]
TwoPhaseMetaRecommender
Initial recommender
RandomRecommender
Compatibility: SearchSpaceType.HYBRID
Recommender
BotorchRecommender
Surrogate
GaussianProcessSurrogate
Supports Transfer Learning: True
Kernel factory: DefaultKernelFactory()
Acquisition function: qLogExpectedImprovement()
Compatibility: SearchSpaceType.HYBRID
Sequential continuous: False
Hybrid sampler: None
Sampling percentage: 1.0
Switch after: 1
Manual verification of the constraint¶
The following loop performs some recommendations and manually verifies the given constraints.
N_ITERATIONS = 2 if SMOKE_TEST else 3
for kIter in range(N_ITERATIONS):
print(f"\n#### ITERATION {kIter+1} ####")
print("## ASSERTS ##")
print(
"No. of searchspace entries where fractions do not sum to 100.0: ",
campaign.searchspace.discrete.exp_rep[["Frac1", "Frac2", "Frac3"]]
.sum(axis=1)
.apply(lambda x: x - 100.0)
.abs()
.gt(SUM_TOLERANCE)
.sum(),
)
print(
"No. of searchspace entries that have duplicate solvent labels: ",
campaign.searchspace.discrete.exp_rep[["Solv1", "Solv2", "Solv3"]]
.nunique(axis=1)
.ne(3)
.sum(),
)
print(
"No. of searchspace entries with permutation-invariant combinations: ",
campaign.searchspace.discrete.exp_rep[["Solv1", "Solv2", "Solv3"]]
.apply(frozenset, axis=1)
.to_frame()
.join(campaign.searchspace.discrete.exp_rep[["Frac1", "Frac2", "Frac3"]])
.duplicated()
.sum(),
)
# The following asserts only work if the tolerance for the threshold condition in
# the constraint are not 0. Otherwise, the sum/prod constraints will remove more
# points than intended due to numeric rounding
print(
f"No. of unique 1-solvent entries (exp. {math.comb(len(dict_solvents), 1)*1})",
(campaign.searchspace.discrete.exp_rep[["Frac1", "Frac2", "Frac3"]] == 0.0)
.sum(axis=1)
.eq(2)
.sum(),
)
print(
f"No. of unique 2-solvent entries (exp."
f" {math.comb(len(dict_solvents), 2)*(RESOLUTION-2)})",
(campaign.searchspace.discrete.exp_rep[["Frac1", "Frac2", "Frac3"]] == 0.0)
.sum(axis=1)
.eq(1)
.sum(),
)
print(
f"No. of unique 3-solvent entries (exp."
f" {math.comb(len(dict_solvents), 3)*((RESOLUTION-3)*(RESOLUTION-2))//2})",
(campaign.searchspace.discrete.exp_rep[["Frac1", "Frac2", "Frac3"]] == 0.0)
.sum(axis=1)
.eq(0)
.sum(),
)
rec = campaign.recommend(batch_size=5)
add_fake_measurements(rec, campaign.targets)
campaign.add_measurements(rec)
#### ITERATION 1 ####
## ASSERTS ##
No. of searchspace entries where fractions do not sum to 100.0: 0
No. of searchspace entries that have duplicate solvent labels: 0
No. of searchspace entries with permutation-invariant combinations: 0
No. of unique 1-solvent entries (exp. 4) 4
No. of unique 2-solvent entries (exp. 18) 18
No. of unique 3-solvent entries (exp. 12) 12
#### ITERATION 2 ####
## ASSERTS ##
No. of searchspace entries where fractions do not sum to 100.0: 0
No. of searchspace entries that have duplicate solvent labels: 0
No. of searchspace entries with permutation-invariant combinations: 0
No. of unique 1-solvent entries (exp. 4) 4
No. of unique 2-solvent entries (exp. 18) 18
No. of unique 3-solvent entries (exp. 12) 12