Example for using the multi target mode for the objective¶
Example for using the multi target mode for the objective. It uses a desirability value to handle several targets.
This example assumes some basic familiarity with using BayBE.
We thus refer to campaign for a basic example.
Necessary imports for this example¶
import pandas as pd
from baybe import Campaign
from baybe.objectives import DesirabilityObjective
from baybe.parameters import CategoricalParameter, NumericalDiscreteParameter
from baybe.searchspace import SearchSpace
from baybe.targets import NumericalTarget
from baybe.utils.dataframe import add_fake_results
Experiment setup and creating the searchspace¶
Categorical_1 = CategoricalParameter("Cat_1", values=["22", "33"], encoding="OHE")
Categorical_2 = CategoricalParameter(
"Cat_2",
values=["very bad", "bad", "OK", "good", "very good"],
encoding="INT",
)
Num_disc_1 = NumericalDiscreteParameter(
"Num_disc_1", values=[1, 2, 3, 4, 6, 8, 10], tolerance=0.3
)
Num_disc_2 = NumericalDiscreteParameter(
"Num_disc_2", values=[-1, -3, -6, -9], tolerance=0.3
)
parameters = [Categorical_1, Categorical_2, Num_disc_1, Num_disc_2]
searchspace = SearchSpace.from_product(parameters=parameters)
Defining the targets¶
The multi target mode is handled when creating the objective object. Thus we first need to define the different targets.
This examples has different targets with different modes. The first target is maximized and while the second one is minimized. Note that in this multi target mode, the user must specify bounds for each target.
Target_1 = NumericalTarget(
name="Target_1", mode="MAX", bounds=(0, 100), transformation="LINEAR"
)
Target_2 = NumericalTarget(
name="Target_2", mode="MIN", bounds=(0, 100), transformation="LINEAR"
)
For each target it is also possible to specify a target_transform function.
A detailed discussion of this functionality can be found at the end of this example.
In this example, define a third target working with the mode MATCH.
We furthermore use target_transform="BELL".
Target_3 = NumericalTarget(
name="Target_3", mode="MATCH", bounds=(45, 55), transformation="BELL"
)
Note that the MATCH mode seeks to have the target at the mean between the two bounds.
For example, choosing 95 and 105 will lead the algorithm seeking 100 as the optimal
value.
Thus, using the bounds, it is possible to control both the match target and
the range around this target that is considered viable.
Creating the objective¶
Now to work with these three targets the objective object must be properly created.
The mode is set to DESIRABILITY and the targets are described in a list.
targets = [Target_1, Target_2, Target_3]
As the recommender requires a single function, the different targets need to be
combined.
Thus, a scalarizer is used to create a single target out of the several targets given.
The combine function can either be the mean MEAN or the geometric mean GEOM_MEAN.
Per default, GEOM_MEAN is used.
Weights for each target can also be specified as a list of floats in the arguments
Per default, weights are equally distributed between all targets and are normalized
internally.
It is thus not necessary to handle normalization or scaling.
objective = DesirabilityObjective(
targets=targets,
weights=[20, 20, 60],
scalarizer="MEAN",
)
print(objective)
[1mObjective[0m
[1mType: [0mDesirabilityObjective
[1mTargets [0m
Type Name Mode Lower_Bound Upper_Bound Transformation \
0 NumericalTarget Target_1 MAX 0.0 100.0 LINEAR
1 NumericalTarget Target_2 MIN 0.0 100.0 LINEAR
2 NumericalTarget Target_3 MATCH 45.0 55.0 BELL
Weight
0 20.0
1 20.0
2 60.0
[1mScalarizer: [0mMEAN
Creating and printing the campaign¶
campaign = Campaign(searchspace=searchspace, objective=objective)
print(campaign)
[1mCampaign[0m
[1mMeta Data[0m
Batches Done: 0
Fits Done: 0
[1mSearch Space[0m
[1mSearch Space Type: [0mDISCRETE
[1mDiscrete Search Space[0m
[1mDiscrete Parameters[0m
Name Type Num_Values Encoding
0 Cat_1 CategoricalParameter 2 CategoricalEncoding.OHE
1 Cat_2 CategoricalParameter 5 CategoricalEncoding.INT
2 Num_disc_1 NumericalDiscreteParameter 7 None
3 Num_disc_2 NumericalDiscreteParameter 4 None
[1mExperimental Representation[0m
Cat_1 Cat_2 Num_disc_1 Num_disc_2
0 22 very bad 1.0 -9.0
1 22 very bad 1.0 -6.0
2 22 very bad 1.0 -3.0
.. ... ... ... ...
277 33 very good 10.0 -6.0
278 33 very good 10.0 -3.0
279 33 very good 10.0 -1.0
[280 rows x 4 columns]
[1mMetadata:[0m
was_recommended: 0/280
was_measured: 0/280
dont_recommend: 0/280
[1mConstraints[0m
Empty DataFrame
Columns: []
Index: []
[1mComputational Representation[0m
Cat_1_22 Cat_1_33 ... Num_disc_1 Num_disc_2
0 1.0 0.0 ... 1.0 -9.0
1 1.0 0.0 ... 1.0 -6.0
2 1.0 0.0 ... 1.0 -3.0
.. ... ... ... ... ...
277 0.0 1.0 ... 10.0 -6.0
278 0.0 1.0 ... 10.0 -3.0
279 0.0 1.0 ... 10.0 -1.0
[280 rows x 5 columns]
[1mObjective[0m
[1mType: [0mDesirabilityObjective
[1mTargets [0m
Type Name Mode Lower_Bound Upper_Bound Transformation \
0 NumericalTarget Target_1 MAX 0.0 100.0 LINEAR
1 NumericalTarget Target_2 MIN 0.0 100.0 LINEAR
2 NumericalTarget Target_3 MATCH 45.0 55.0 BELL
Weight
0 20.0
1 20.0
2 60.0
[1mScalarizer: [0mMEAN
TwoPhaseMetaRecommender(initial_recommender=RandomRecommender(allow_repeated_recomm
endations=False, allow_recommending_already_measured=True), recommender=BotorchRecommender(allow_repeated_recommendations=False, allow_recommending_already_measured=True, surrogate_model=GaussianProcessSurrogate(kernel_factory=DefaultKernelFactory(), _model=None), acquisition_function=qLogExpectedImprovement(), _botorch_acqf=None, acquisition_function_cls=None, sequential_continuous=False, hybrid_sampler=None, sampling_percentage=1.0), switch_after=1)
Performing some iterations¶
The following loop performs some recommendations and adds fake results. It also prints what happens to internal data.
N_ITERATIONS = 3
for kIter in range(N_ITERATIONS):
rec = campaign.recommend(batch_size=3)
add_fake_results(rec, campaign.targets)
campaign.add_measurements(rec)
desirability = campaign.objective.transform(campaign.measurements)
print(f"\n\n#### ITERATION {kIter+1} ####")
print("\nRecommended measurements with fake measured results:\n")
print(rec)
print("\nInternal measurement database with desirability values:\n")
print(pd.concat([campaign.measurements, desirability], axis=1))
#### ITERATION 1 ####
Recommended measurements with fake measured results:
Cat_1 Cat_2 Num_disc_1 Num_disc_2 Target_1 Target_2 Target_3
46 22 bad 6.0 -3.0 89.287695 27.689040 53.620385
188 33 bad 8.0 -9.0 54.714958 32.395289 56.792072
136 22 very good 10.0 -9.0 82.235979 29.393339 53.739874
Internal measurement database with desirability values:
Cat_1 Cat_2 Num_disc_1 Num_disc_2 Target_1 Target_2 Target_3 \
0 22 bad 6.0 -3.0 89.287695 27.689040 53.620385
1 33 bad 8.0 -9.0 54.714958 32.395289 56.792072
2 22 very good 10.0 -9.0 82.235979 29.393339 53.739874
BatchNr FitNr Desirability
0 1 NaN 0.784838
1 1 NaN 0.483119
2 1 NaN 0.759277
#### ITERATION 2 ####
Recommended measurements with fake measured results:
Cat_1 Cat_2 Num_disc_1 Num_disc_2 Target_1 Target_2 Target_3
index
18 22 very bad 6.0 -3.0 52.099147 47.343427 51.813798
74 22 OK 6.0 -3.0 93.488807 41.706375 55.159932
104 22 good 8.0 -9.0 63.316223 84.934914 51.455603
Internal measurement database with desirability values:
Cat_1 Cat_2 Num_disc_1 Num_disc_2 Target_1 Target_2 Target_3 \
0 22 bad 6.0 -3.0 89.287695 27.689040 53.620385
1 33 bad 8.0 -9.0 54.714958 32.395289 56.792072
2 22 very good 10.0 -9.0 82.235979 29.393339 53.739874
3 22 very bad 6.0 -3.0 52.099147 47.343427 51.813798
4 22 OK 6.0 -3.0 93.488807 41.706375 55.159932
5 22 good 8.0 -9.0 63.316223 84.934914 51.455603
BatchNr FitNr Desirability
0 1 1.0 0.784838
1 1 1.0 0.483119
2 1 1.0 0.759277
3 2 NaN 0.771304
4 2 NaN 0.655847
5 2 NaN 0.731868
#### ITERATION 3 ####
Recommended measurements with fake measured results:
Cat_1 Cat_2 Num_disc_1 Num_disc_2 Target_1 Target_2 Target_3
index
2 22 very bad 1.0 -3.0 32.944727 88.076184 56.847457
19 22 very bad 6.0 -1.0 73.640865 70.293208 56.821857
26 22 very bad 10.0 -3.0 5.462638 37.145001 53.003474
Internal measurement database with desirability values:
Cat_1 Cat_2 Num_disc_1 Num_disc_2 Target_1 Target_2 Target_3 \
0 22 bad 6.0 -3.0 89.287695 27.689040 53.620385
1 33 bad 8.0 -9.0 54.714958 32.395289 56.792072
2 22 very good 10.0 -9.0 82.235979 29.393339 53.739874
3 22 very bad 6.0 -3.0 52.099147 47.343427 51.813798
4 22 OK 6.0 -3.0 93.488807 41.706375 55.159932
5 22 good 8.0 -9.0 63.316223 84.934914 51.455603
6 22 very bad 1.0 -3.0 32.944727 88.076184 56.847457
7 22 very bad 6.0 -1.0 73.640865 70.293208 56.821857
8 22 very bad 10.0 -3.0 5.462638 37.145001 53.003474
BatchNr FitNr Desirability
0 1 1.0 0.784838
1 1 1.0 0.483119
2 1 1.0 0.759277
3 2 2.0 0.771304
4 2 2.0 0.655847
5 2 2.0 0.731868
6 3 NaN 0.324641
7 3 NaN 0.443249
8 3 NaN 0.637588
Addendum: Description of transformation functions¶
This function is used to transform target values to the interval [0,1] for MAX/MIN
mode.
An ascending or decreasing LINEAR function is used per default.
This function maps input values in a specified interval [lower, upper] to the interval
[0,1].
Outside the specified interval, the function remains constant, that is, 0 or 1.
For the match mode, two functions are available TRIANGULAR and BELL.
The TRIANGULAR function is 0 outside a specified interval and linearly increases to 1
from both
interval ends, reaching the value 1 at the center of the interval.
This function is used per default for MATCH mode.
The BELL function is a Gaussian bell curve, specified through the boundary values of
the sigma
interval, reaching the maximum value of 1 at the interval center.