"""Contains the implementations for experimental design methods using Emukit
Implements the supported experimental design methods using the Emukit library.
"""
import numpy as np
from emukit.core.initial_designs import RandomDesign
from emukit.experimental_design import ExperimentalDesignLoop
from emukit.experimental_design.acquisitions import (
IntegratedVarianceReduction,
ModelVariance,
)
from matplotlib import pyplot as plt
from ..base import EmukitBase
from ..estimators import EmukitEstimator
[docs]
class EmukitExperimentalDesign(EmukitBase):
"""The Emukit experimental design method class."""
[docs]
def execute(self) -> None:
"""Execute the simulation calibration procedure."""
experimental_design_kwargs = self.get_calibration_func_kwargs()
def target_function(X: np.ndarray) -> np.ndarray:
return self.calibration_func_wrapper(
X,
self,
self.specification.observed_data,
self.names,
self.data_types,
experimental_design_kwargs,
)
n_init = self.specification.n_init
X, Y = self.get_X_Y(n_init, target_function)
method_kwargs = self.specification.method_kwargs
estimator = EmukitEstimator(method_kwargs)
estimator.fit(X, Y)
acquisition_name = self.specification.method
acquisitions = dict(
model_variance=ModelVariance,
integrated_variance_reduction=IntegratedVarianceReduction,
)
acquisition_class = acquisitions.get(acquisition_name, None)
if acquisition_class is None:
raise ValueError(
f"Unsupported emulator acquisition type: {acquisition_name}.",
f"Supported acquisition types are {', '.join(acquisitions)}",
)
acquisition = acquisition_class(model=estimator.emulator)
design_loop = ExperimentalDesignLoop(
model=estimator.emulator,
space=self.parameter_space,
acquisition=acquisition,
batch_size=1,
)
n_iterations = self.specification.n_iterations
design_loop.run_loop(target_function, stopping_condition=n_iterations)
self.emulator = estimator
self.design_loop = design_loop
[docs]
def analyze(self) -> None:
"""Analyze the results of the simulation calibration procedure."""
task, time_now, experiment_name, outdir = self.prepare_analyze()
design = RandomDesign(self.parameter_space)
n_samples = self.specification.n_samples
X_sample = design.get_samples(n_samples)
if self.specification.use_shap and outdir is not None:
outfile = self.join(
outdir,
f"{time_now}-{task}-{experiment_name}-param-importances.png",
)
self.calculate_shap_importances(
X_sample,
self.emulator,
self.names,
self.specification.test_size,
outfile,
)
predicted_mu, predicted_std = self.emulator.predict(X_sample, return_std=True)
observed_data = self.specification.observed_data
output_label = self.specification.output_labels[0] # type: ignore[index]
X = np.arange(0, observed_data.shape[0], 1)
fig, axes = plt.subplots(nrows=2, figsize=self.specification.figsize)
axes[0].plot(X, observed_data)
axes[0].set_title(f"Observed {output_label}")
mean_predicted_mu = predicted_mu.mean(axis=0)
mean_predicted_std = predicted_std.mean(axis=0)
axes[1].plot(X, mean_predicted_mu)
for mult, alpha in [(1, 0.6), (2, 0.4), (3, 0.2)]:
axes[1].fill_between(
X,
mean_predicted_mu - mult * mean_predicted_std,
mean_predicted_mu + mult * mean_predicted_std,
color="C0",
alpha=alpha,
)
axes[1].set_title(f"Emulated {output_label}")
self.present_fig(
fig, outdir, time_now, task, experiment_name, f"emulated-{output_label}"
)