Skip to content

pablormier/yabox

Repository files navigation

DOI

Yet another black-box optimization library for Python

Description

Yabox is a very small library for black-box (derivative free) optimization of functions that only depends on numpy and matplotlib for visualization. The library includes different stochastic algorithms for minimizing a function f(X) that does not need to have an analytical form, where X = {x1, ..., xN}. The current version of the library includes the Differential Evolution algorithm and a modified version for parallel evaluation.

Example of minimization of the Ackley function (using Yabox and Differential Evolution):

Ackley Function

Installation

Yabox is in PyPI so you can use the following command to install the latest released version:

pip install yabox

Basic usage

Pre-defined functions

Yabox includes some default benchmark functions used in black-box optimization, available in the package yabox.problems. These functions also include 2D and 3D plotting capabilities:

>>> from yabox.problems import Levy
>>> problem = Levy()
>>> problem.plot3d()

Levy Function

A problem is just a function that can be evaluated for a given X:

>>> problem(np.array([1,1,1]))
0.80668910823394901

Optimization

Simple example minimizing a function of one variable x using Differential Evolution, searching between -10 <= x <= 10:

>>> from yabox import DE
>>> DE(lambda x: sum(x**2), [(-10, 10)]).solve()
(array([ 0.]), 0.0)

Example using Differential Evolution and showing progress (requires tqdm)

Optimization example

Yabox includes a parallel version of Differential Evolution. Import PDE instead of DE:

>>> from yabox import PDE
>>> PDE(lambda x: sum(x**2), [(-10, 10)]).solve()
(array([ 0.]), 0.0)

For more examples, check the notebooks included in the project

About

This library is inspired in the scipy's differential evolution implementation. The main goal of Yabox is to include a larger set of stochastic black-box optimization algorithms plus many utilities, all in a small library with minimal dependencies.