botorchex

Botorch extention library including custom acquistion functions and surrogate models.

Installation

$ pip install botorchex

Ease to use

Botorch compatible interface.

Implementation List

Custom Acqusition function

• Multi Objective Monte-Carlo Probability Improvement This acquistion function can deacrease more computational resource(wall-time) comparing to other multi objective acqusition function. This performance especially is shown in the more than 3 objctive cases. However, the convergence speed is longer than the others and there is no theoretical background.

from botorch.models.gp_regression import SingleTaskGP
from botorch.models.model_list_gp_regression import ModelListGP

from botorchex.acquisition.multi_objective.monte_carlo import qMultiProbabilityOfImprovement

model1 = SingleTaskGP(train_X, train_Y[0, :])
model2 = SingleTaskGP(train_X, train_Y[1, :])
# we assume the outputs are independent each other.
best_f = train_Y.max(dim=1)
modes = ModelListGP([model1, model2])
qPI = qMultiProbabilityOfImprovement(models, best_f)
qmpi = qMPI(test_X)

If you want to know more examples, you can check the example(multi_objective_bo.ipynb)

Custom Surrogates

• GNN based surrogates?

Referances

• https://botorch.org/
• https://gpytorch.ai/