Rust toolbox for Efficient Global Optimization algorithms inspired from SMT.
egobox is twofold:
1. for developers: a set of Rust libraries useful to implement bayesian optimization (EGO-like) algorithms,
2. for end-users: a Python module, the Python binding of the implemented EGO-like optimizer, named Egor.
egobox Rust libraries consists of the following sub-packages.
| Name | Version | Documentation | Description |
| :---------------------------------------------------- | :---------------------------------------------------------------------------------------------- | :-------------------------------------------------------------------------- | :------------------------------------------------------------------------------ |
| doe | 
| sampling methods; contains LHS, FullFactorial, Random methods |
| gp | 
| gaussian process regression; contains Kriging and PLS dimension reduction |
| moe | 
| mixture of experts using GP models |
| ego | 
| efficient global optimization with basic constraints and mixed integer handling |
Depending on the sub-packages you want to use, you have to add following declarations to your Cargo.toml
[dependencies]
egobox-doe = { version = "0.5.0" }
egobox-gp = { version = "0.5.0" }
egobox-moe = { version = "0.5.0" }
egobox-ego = { version = "0.5.0" }
serializable-gpThe serializable-gp feature enables the serialization of GP models using the serde crate.
persistent-moeThe persistent-moe feature enables save() and load() methods for MoE model to/from a json file using the serde crate.
Examples (in examples/ sub-packages folder) are run as follows:
bash
$ cd doe && cargo run --example samplings --release
bash
$ cd gp && cargo run --example kriging --release
bash
$ cd moe && cargo run --example clustering --release
bash
$ cd ego && cargo run --example ackley --release
egobox relies on linfa project for methods like clustering and dimension reduction, but also try to adopt as far as possible the same coding structures.
As for linfa, the linear algebra routines used in gp, moe ad ego are provided by the pure-Rust linfa-linalg crate, the default linear algebra provider.
Otherwise, you can choose an external BLAS/LAPACK backend available through the ndarray-linalg crate. In this case, you have to specify the blas feature and a linfa BLAS/LAPACK backend feature (more information in linfa features).
Thus, for instance, to use gp with the Intel MKL BLAS/LAPACK backend, you could specify in your Cargo.toml the following features:
[dependencies]
egobox-gp = { version = "0.5.0", features = ["blas", "linfa/intel-mkl-static"] }
or you could run the gp example as follows:
bash
$ cd gp && cargo run --example kriging --release --features blas,linfa/intel-mkl-static
Thanks to the PyO3 project, which makes Rust well suited for building Python extensions, the EGO algorithm written in Rust (aka Egor) is binded in Python. You can install the Python package using:
bash
$ pip install egobox
See the tutorial notebook for usage of the optimizer.
I started this library as a way to learn Rust and see if it can be used to implement algorithms like those in the SMT toolbox[^1]. As the first components (doe, gp) emerged, it appears I could translate Python code almost line by line in Rust (well... after a great deal of borrow-checker fight!) and thanks to Rust ndarray library ecosystem.
This library relies also on the linfa project which aims at being the "scikit-learn-like ML library for Rust". Along the way I could contribute to linfa by porting gaussian mixture model (linfa-clustering/gmm) and partial least square family methods (linfa-pls) confirming the fact that Python algorithms translation in Rust could be pretty straightforward.
While I did not benchmark exactly my Rust code against SMT Python one, from my debugging sessions, I noticed I did not get such a great speed up. Actually, algorithms like doe and gp relies extensively on linear algebra and Python famous libraries numpy/scipy which are strongly optimized by calling C or Fortran compiled code.
My guess at this point is that interest could come from some Rust algorithms built upon these initial building blocks hence I started to implement mixture of experts algorithm (moe) and on top surrogate-based optimization EGO algorithm (ego) which gives its name to the library^2. Aside from performance, such library can also take advantage from the others Rust selling points.
If you happen to find this Rust library useful for your research, you can cite this project as follows:
@Misc{egobox,
author = {Rémi Lafage},
title = {Egobox: efficient global optimization toolbox in Rust},
year = {2020--},
url = "https://github.com/relf/egobox"
}