Custom Sobo Strategy

The CustomSoboStrategy can be used to design custom objectives or objective combinations for optimizations. In this tutorial notebook, it is shown how to use it to optimize a quantity that depends on a combination of an inferred quantity and one of the inputs.

Imports

import torch

import bofire.strategies.api as strategies
from bofire.benchmarks.api import Himmelblau
from bofire.data_models.strategies.api import CustomSoboStrategy
from bofire.utils.torch_tools import tkwargs

Setup the optimization

For the optimization, we want to subtract the inferred quantity by the value of feature x_0.

benchmark = Himmelblau()
experiments = benchmark.f(benchmark.domain.inputs.sample(10), return_complete=True)

strategy_data = CustomSoboStrategy(domain=benchmark.domain)
strategy = strategies.map(strategy_data)


# here we find out what is the index of the input feature in the input tensor `X`
# in the manipulation function below
feature2index, _ = strategy.domain.inputs._get_transform_info(
    strategy.input_preprocessing_specs
)
feat_idx = feature2index["x_1"][0]


# we assign now a torch based function to the strategy which performs the custom manipulation of the objective
# the signature has to be understood in the following way:
# - samples: the samples to evaluate the objective on, these are the predicted Y/output values of the model(s)
# - callables: the botorch callables associated to objectives associated to the features
#   (have a look at `get_objective_callable` in `bofire/utils/torch_tools.py`)
# - weights: the weights associated to the objectives
#   (have a look here: `_callables_and_weights` in `bofire/utils/torch_tools.py`)
# - X: a tensor of input values associated to the output values  samples, associated to the Y/output values (`samples`)


def f(samples, callables, weights, X):
    val = torch.tensor(0.0).to(**tkwargs)
    for c, w in zip(callables, weights):
        val = val + c(samples, None) * w
    # here, you have to implement the custom manipulation of the objective
    # in this example, we subtract the value of the first feature from the objective
    val = val - X[..., feat_idx]
    return val


strategy.f = f

strategy.tell(experiments)
strategy.ask(1)

	x_1	x_2	y_pred	y_sd	y_des
0	-2.993551	1.645679	66.696969	211.481288	-66.696969