Emukit

0.5.1 verified Fri May 01 auth: no python

A Python toolkit for decision-making under uncertainty, including Bayesian optimization, experimental design, and quadrature. Current version 0.5.1, requires Python >=3.10. Released approximately twice per year since 2018.

pip install emukit

Common errors

error ModuleNotFoundError: No module named 'emukit.bayesian_optimization' ↓

cause Import path is wrong; the correct import is from emukit.bayesian_optimization.loops.

fix

Use: from emukit.bayesian_optimization.loops import BayesianOptimizationLoop

error AttributeError: module 'emukit' has no attribute 'model_wrappers' ↓

cause Misspelled 'emukit' or installed the wrong package; also GPy may be missing.

fix

pip install emukit[gpy] or pip install emukit[sklearn] for sklearn wrapper.

error ValueError: The number of points in X and Y must be the same ↓

cause Passed X and Y arrays with mismatched first dimension.

fix

Ensure X.shape[0] == Y.shape[0]; both must be 2D arrays (n_samples, n_dims).

error ImportError: cannot import name 'GPyModel' from 'emukit.model_wrappers' ↓

cause Incorrect class name; should be GPyModelWrapper.

fix

Use: from emukit.model_wrappers import GPyModelWrapper

Warnings

breaking Emukit 0.5.0 dropped support for Python <3.10. Also, numpy>=2.0 compatibility now requires using emukit core without GPy (GPy is pinned to numpy<2.0). ↓

fix Upgrade to Python >=3.10. For numpy>=2.0, install emukit without gpy extra and use sklearn GP wrapper instead.

gotcha The GPy model wrapper in emukit.model_wrappers.GPyModel expects X and Y arrays; do not pass GPy objects directly. ↓

fix Use from emukit.model_wrappers import GPyModelWrapper and pass numpy arrays.

gotcha BayesianOptimizationLoop requires the 'get_next_points' method to receive a list of (X, Y) tuples, not just the model. ↓

fix Pass loop.get_next_points([X, Y]) with X and Y as 2D arrays.

Install

pip install emukit[gpy]

pip install emukit[sklearn]

Imports

BayesianOptimizationLoop

wrong

from emukit.bayesian_optimization import BayesianOptimizationLoop

correct

from emukit.bayesian_optimization.loops import BayesianOptimizationLoop

BayesianOptimizationLoop is in emukit.bayesian_optimization.loops, not the top-level of the module.

create_initial_design

wrong

from emukit.experimental_design import LatinDesign

correct

from emukit.experimental_design.model_free.latin_design import LatinDesign

LatinDesign is in emukit.experimental_design.model_free.latin_design.

GPyModel
wrong
```
from emukit.models import GPyModel
```
correct
```
from emukit.model_wrappers import GPyModel
```
Model wrappers live in emukit.model_wrappers, not emukit.models.

QuadratureCaratheodoryMeasure

wrong

from emukit.quadrature import QuadratureCaratheodoryMeasure

correct

from emukit.quadrature.kernels import QuadratureCaratheodoryMeasure

Quadrature kernel/measure classes are in emukit.quadrature.kernels.

Quickstart

Minimal Bayesian optimization loop using Emukit with a GPy model.

import numpy as np
from emukit.core import ContinuousParameter, ParameterSpace
from emukit.model_wrappers import GPyModelWrapper
from emukit.bayesian_optimization.loops import BayesianOptimizationLoop
from emukit.core.acquisition import ExpectedImprovement

# Define a simple objective
def objective(x):
    return x**2

# Parameter space
space = ParameterSpace([ContinuousParameter('x', -5, 5)])

# Initial data
X_init = np.random.uniform(-5, 5, (3, 1))
Y_init = objective(X_init)

# Model
gpy_model = GPyModelWrapper(X_init, Y_init)

# Acquisition function
ei = ExpectedImprovement()

# Optimization loop
loop = BayesianOptimizationLoop(space, gpy_model, acquisition=ei)

# Run one iteration
new_x = loop.get_next_points([X_init, Y_init])
print('Suggested next point:', new_x)