ganesh (/ɡəˈneɪʃ/), named after the Hindu god of wisdom, provides several common minimization algorithms as well as a straightforward, trait-based interface to create your own extensions. This crate is intended to be as simple as possible. The user needs to implement the Function trait on some struct which will take a vector of parameters and return a single-valued Result ($f(\mathbb{R}^n)\to\mathbb{R}$). Users can optionally provide a gradient function to speed up some algorithms, but a default central finite-difference implementation is provided so that all algorithms will work out of the box.

Part of the difficulty of this project is the lack of well-documented implementations of some of the more useful algorithms, like the BFGS family. While I am not sure, I believe this might be the first pure Rust implementation of L-BFGS-B. If you look through other optimization crates, Python packages, or even C/C++ libraries, a common theme you’ll find is that they mostly just bind the original FORTRAN code written for this algorithm. I wanted to write the algorithm from scratch for two reasons. First, it’s a great way to learn the language, and it’s very convenient to not have to worry about dependencies in external languages like FORTRAN. Second, I had a lot of trouble finding bounded optimizers for Rust. argmin, the most-downloaded optimization crate, currently doesn’t implement any constrained optimizers, although external solvers like egobox and cobyla can be used to accomplish this. nlopt has a Rust wrapper that includes these algorithms (written in C), but I quit using it after one too many C-related exceptions that were all-but-impossible to track down or solve with Rust code.