Descent

A non-linear constrained optimisation (mathematical programming) modelling library with first and second order automatic differentiation and interface to Ipopt.

This is in an early state of development but is working / useable.

Design

It supports operator overloading for the ergonomic expression of terms and the operators: +, -, *, powi, sin, cos. It also allows parameterisation of values and easy solver warmstarting to enable quick model adjustments and resolving.

It follows a similar design philosopy to the C++ / Python library madopt. As they stand the AD routines are not as efficient but still adequate for many applications (be sure to build with --release).

The library could be expanded to link to other solvers.

Dependencies

Ipopt (or Bonmin) must be separately installed before attempting to build as Descent links to the libipopt.so shared library.

It has only been tested on linux, but presumably would also work on macos, and potentially on windows in the right environment.

Example

The following code shows and example of solving the following simple problem: min 2y s.t. y >= x*x - x, x in [-10, 10].

let mut m = IpoptModel::new();

let x = m.add_var(-10.0, 10.0, 0.0);
let y = m.add_var(f64::NEG_INFINITY, f64::INFINITY, 0.0);
m.set_obj(2.0*y);
m.add_con(y - x*x + x, 0.0, f64::INFINITY);

let (stat, sol) = m.solve();

A full example of this with additional details is provided under examples/simple.rs, which can be built and run as follows:

cargo build --release --example simple
./target/release/examples/simple

Benchmarking

For the test problem debug for the release profile should be turned on (in Cargo.toml), and then build and run using:

cargo build --release --example problem
./target/release/examples/problem

Issues

Currently large expressions that require their second derivatives computed are slow and memory intensive. As a work around for separable expressions, split them up manually. In the future should be able to automate this.

License

Apache-2.0 or MIT