bayesdesign 0.5.0

Package for Bayesian optimal experimental design

Bayesian Optimal Experiment Design

Use this package to calculate expected information gain for Bayesian optimal experiment design. For an introduction to this topic, see this interactive notebook. To perform a similar calculation with this package, use:

from bed.grid import Grid, GridStack
from bed.design import ExperimentDesigner

designs = Grid(t_obs=np.linspace(0, 5, 51))
features = Grid(y_obs=np.linspace(-1.25, 1.25, 100))
params = Grid(amplitude=1, frequency=np.linspace(0.2, 2.0, 181), offset=0)

sigma_y=0.1
with GridStack(features, designs, params):
    y_mean = params.amplitude * np.sin(params.frequency * (designs.t_obs - params.offset))
    y_diff = features.y_obs - y_mean
    likelihood = np.exp(-0.5 * (y_diff / sigma_y) ** 2)
    features.normalize(likelihood)

designer = ExperimentDesigner(params, features, designs, likelihood)

prior = np.ones(params.shape)
params.normalize(prior);

designer.calculateEIG(prior)

plt.plot(designs.t_obs, designer.EIG)

Browse the examples folder to learn more about using this package.

To run the examples in google colab, select GitHub and enter https://github.com/dkirkby/bayesdesign.

Installation

Install the latest released version from pypi using:

pip install bayesdesign

The only required dependency is numpy. The optional plot module also requires matplotlib.

The changes with each version are documented here.

Upgrade

To upgrade your pip-installed package to the latest released version use:

pip install bayesdesign --upgrade

Contributing

If you have feedback or would like to contribute to this package, please see our contributor's guide.