def _initialise_gp(self, gp, hyper_grid):
"""
Initialises the Gaussian process surrogate model. If `gp` is `None`
uses the default kernel and Gaussian process:
`kernel = sklearn.gaussian_process.kernels.Matern(nu=2.5)`
`gp = sklearn.gaussian_process.GaussianProcessRegressor(
kernel, alpha=1e-6, normalize_y=True, n_restarts_optimizer=5,
random_state=self.generator)`,
such that `random_state` is always set to the class generator. The
data is always scaled using `sklearn.preprocessing.StandardScaler`.
If `hyper_grid` is not `None` the best fit combination will be used
as a surrogate model (calls `sklearn.model_selection.GridSearchCV`)
with 5-fold cross-validation.
Parameters
----------
gp : None or `sklearn.gaussian_process.GaussianProcessRegressor`
Surrogate model Gaussian process.
hyper_grid : None or dict of dictionaries
Hyperparameter grid to be explored when fitting the Gaussian
process.
"""
# Set up the Gaussian process, pipeline and grid search
if gp is None:
kernel = kernels.Matern(nu=2.5)
gp = GaussianProcessRegressor(kernel,
alpha=1e-6,
normalize_y=True,
n_restarts_optimizer=5,
random_state=self.generator)
elif not isinstance(gp, GaussianProcessRegressor):
raise TypeError(
"`gp` must be of {} type.".format(GaussianProcessRegressor))
else:
# Always overwrite the random state
gp.random_state = self.generator
# Set up the pipeline to scale the data
pipe = Pipeline([('scaler', StandardScaler()), ('gp', gp)])
# Optionally set the hyperparameter grid
if hyper_grid is None:
self._surrogate_model = pipe
else:
self._surrogate_model = GridSearchCV(pipe,
hyper_grid,
n_jobs=self.nthreads)