Example for using different strategies

This example shows how to create and use recommender objects. Such an object specifies the recommender adopted to make recommendations. It has several parameters one can adjust, depending on the recommender the user wants to follow.

To apply the selected recommender, this object can be specified in the arguments of the campaign. The different parameters the user can change are:

  • The initial recommender

  • The recommender with its surrogate model and its acquisition function

  • Other parameters to allow or not repetition of recommendations

This examples assumes some basic familiarity with using BayBE. We refer to campaign for a more general and basic example.

Necessary imports for this example

from baybe import Campaign
from baybe.objective import Objective
from baybe.parameters import NumericalDiscreteParameter, SubstanceParameter
from baybe.recommenders import (
    RandomRecommender,
    SequentialGreedyRecommender,
    TwoPhaseMetaRecommender,
)
from baybe.searchspace import SearchSpace
from baybe.surrogates import (
    BayesianLinearSurrogate,
    GaussianProcessSurrogate,
    NGBoostSurrogate,
    RandomForestSurrogate,
)
from baybe.targets import NumericalTarget
from baybe.utils.dataframe import add_fake_results

Available recommenders suitable for initial recommendation

For the first recommendation, the user can specify which recommender to use. The following initial recommenders are available. Note that it is necessary to make the corresponding import before using them.

initial_recommenders = [
    "Random",  #: RandomRecommender(),
    "Farthest Point Sampling",  # FPSRecommender(),
    "KMEANS Clustering",  # KMeansClusteringRecommender(),
]

# Per default the initial recommender chosen is a random recommender.
INITIAL_RECOMMENDER = RandomRecommender()

Available surrogate models

This model uses available data to model the objective function as well as the uncertainty. The surrogate model is then used by the acquisition function to make recommendations.

The following are some available basic surrogates Use baybe.surrogates.get_available_surrogates() for a complete list.

available_surrogate_models = [
    GaussianProcessSurrogate(),
    RandomForestSurrogate(),
    NGBoostSurrogate(),
    BayesianLinearSurrogate(),
]

Per default a Gaussian Process is used

SURROGATE_MODEL = GaussianProcessSurrogate()

Acquisition function

This function looks for points where measurements of the target value could improve the model. The following acquisition functions are generally available.

available_acq_functions = [
    "qPI",  # q-Probability Of Improvement
    "qEI",  # q-Expected Improvement
    "qUCB",  # q-upper confidence bound with beta of 1.0
    "PM",  # Posterior Mean,
    "PI",  # Probability Of Improvement,
    "EI",  # Expected Improvement,
    "UCB",  # upper confidence bound with beta of 1.0
]

Note that the qvailability of the acquisition functions might depend on the batch_size:

  • If batch_size is set to 1, all available acquisition functions can be chosen

  • If a larger value is chosen, only those that allow batching. That is, ‘q’-variants of the acquisition functions must be chosen.

The default he acquisition function is q-Expected Improvement.

ACQ_FUNCTION = "qEI"

Other parameters

Two other boolean hyperparameters can be specified when creating a recommender object. The first one allows the recommendation of points that were already recommended previously. The second one allows the recommendation of points that have already been measured. Per default, they are set to True.

ALLOW_REPEATED_RECOMMENDATIONS = True
ALLOW_RECOMMENDING_ALREADY_MEASURED = True

Creating the recommender object

To create the recommender object, each parameter described above can be specified as follows. Note that they all have default values. Therefore one does not need to specify all of them to create a recommender object.

recommender = TwoPhaseMetaRecommender(
    initial_recommender=INITIAL_RECOMMENDER,
    recommender=SequentialGreedyRecommender(
        surrogate_model=SURROGATE_MODEL,
        acquisition_function_cls=ACQ_FUNCTION,
        allow_repeated_recommendations=ALLOW_REPEATED_RECOMMENDATIONS,
        allow_recommending_already_measured=ALLOW_RECOMMENDING_ALREADY_MEASURED,
    ),
)

print(recommender)

TwoPhaseMetaRecommender(allow_repeated_recommendations=None,

allow_recommending_already_measured=None, initial_recommender=RandomRecommender(allow_repeated_recommendations=False, allow_recommending_already_measured=True), recommender=SequentialGreedyRecommender(allow_repeated_recommendations=True, allow_recommending_already_measured=True, surrogate_model=GaussianProcessSurrogate(model_params={}, _model=None), acquisition_function_cls=’qEI’, _acquisition_function=None, hybrid_sampler=’None’, sampling_percentage=1.0), switch_after=1)

Note that there are the additional keywords hybrid_sampler and sampling_percentag. Their meaning and how to use and define it are explained in the hybrid backtesting example. We thus refer to hybrid for details on these.

Example Searchspace and objective parameters

We use the same data used in the campaign example.

dict_solvent = {
    "DMAc": r"CC(N(C)C)=O",
    "Butyornitrile": r"CCCC#N",
    "Butyl Ester": r"CCCCOC(C)=O",
    "p-Xylene": r"CC1=CC=C(C)C=C1",
}
dict_base = {
    "Potassium acetate": r"O=C([O-])C.[K+]",
    "Potassium pivalate": r"O=C([O-])C(C)(C)C.[K+]",
    "Cesium acetate": r"O=C([O-])C.[Cs+]",
    "Cesium pivalate": r"O=C([O-])C(C)(C)C.[Cs+]",
}
dict_ligand = {
    "BrettPhos": r"CC(C)C1=CC(C(C)C)=C(C(C(C)C)=C1)C2=C(P(C3CCCCC3)C4CCCCC4)C(OC)="
    "CC=C2OC",
    "Di-tert-butylphenylphosphine": r"CC(C)(C)P(C1=CC=CC=C1)C(C)(C)C",
    "(t-Bu)PhCPhos": r"CN(C)C1=CC=CC(N(C)C)=C1C2=CC=CC=C2P(C(C)(C)C)C3=CC=CC=C3",
}

solvent = SubstanceParameter("Solvent", data=dict_solvent, encoding="MORDRED")
base = SubstanceParameter("Base", data=dict_base, encoding="MORDRED")
ligand = SubstanceParameter("Ligand", data=dict_ligand, encoding="MORDRED")
temperature = NumericalDiscreteParameter(
    "Temperature", values=[90, 105, 120], tolerance=2
)
concentration = NumericalDiscreteParameter(
    "Concentration", values=[0.057, 0.1, 0.153], tolerance=0.005
)

We collect all parameters in a list.

parameters = [solvent, base, ligand, temperature, concentration]

We create the searchspace and the objective.

searchspace = SearchSpace.from_product(parameters=parameters)

objective = Objective(
    mode="SINGLE", targets=[NumericalTarget(name="yield", mode="MAX")]
)

Creating the campaign

The recommender object can now be used together with the searchspace and the objective as follows.

campaign = Campaign(
    searchspace=searchspace,
    recommender=recommender,
    objective=objective,
)

This campaign can then be used to get recommendations and add measurements:

recommendation = campaign.recommend(batch_size=3)
print("\n\nRecommended experiments: ")
print(recommendation)

Recommended experiments: 
           Solvent               Base                        Ligand  \
182  Butyornitrile     Cesium acetate                 (t-Bu)PhCPhos   
309    Butyl Ester    Cesium pivalate  Di-tert-butylphenylphosphine   
24            DMAc  Potassium acetate                 (t-Bu)PhCPhos   

     Temperature  Concentration  
182         90.0          0.153  
309        105.0          0.057  
24         120.0          0.057  

add_fake_results(recommendation, campaign)
print("\n\nRecommended experiments with fake measured values: ")
print(recommendation)

Recommended experiments with fake measured values: 
           Solvent               Base                        Ligand  \
182  Butyornitrile     Cesium acetate                 (t-Bu)PhCPhos   
309    Butyl Ester    Cesium pivalate  Di-tert-butylphenylphosphine   
24            DMAc  Potassium acetate                 (t-Bu)PhCPhos   

     Temperature  Concentration      yield  
182         90.0          0.153  18.968785  
309        105.0          0.057   6.431328  
24         120.0          0.057  23.587571  

campaign.add_measurements(recommendation)