sparkle

sparkle is a parametric, gradient-free optimization library that I mostly developed at CEMEF, in the CFL group (public repo from the team is here). It is designed to provide a common interface to various algorithms, and to make numerical experimentation easy.

If you end up using this library for research purpose, please consider citing the following paper (link):

Mixed-variable policy-based optimization
J. Viquerat
arXiv pre-print 2506.13240, 2025

Installation and usage

Clone this repository and install it locally. We recommend to use uv:

git clone git@github.com:jviquerat/sparkle.git
cd sparkle
uv venv
source .venv/bin/activate
uv pip install -r pyproject.toml
uv pip install -e .

Environments are expected to be available locally or present in the path. To train an agent on an environment, a .json case file is required (sample files are available in sparkle/env). Once you have written the corresponding <env_name>.json file to configure your agent, just run:

spk --train <json_file>

Analytical environments

Environment	Default dimension	Description	Illustration
parabola	2	Classic parabola (solved with PBO)
rosenbrock	2	Rosenbrock function (solved with CMAES)
multi1d	1	Multi1D function (solved with EGO)
constraint	2	Parabola with a priori constraints on parameters (solved with CMAES)

Physics-based environments

Environment	Default dimension	Description	Illustration
lorenz	4	Optimizing a control law for the chaotic Lorenz attractor, adapted from this ref (solved with PBO)
n-body	9	Optimizing the initial parameters to find periodic orbits, adapted from this ref (solved with CMAES)
heat-source	14	Optimizing the positions of heat sources to obtain a high temperature distribution with low variance in a target area (solved with CMAES)
packing	26	Finding the best disk packing within a square domain, adapted from this ref (solved with PBO)
emstack	40	Optimizing the reflectance of a dielectric mirror with mixed continuous/discrete variables (solved with PBO)