Simple(x) Global Optimization

A Rust implementation of the Simple(x) black-box global optimization algorithm.

This algorithm (which should not be confused with the simplex algorithm) is closest to bayesian optimization. Its strengths, compared to bayesian optimization, are the ability to deal with a large number of sample and high dimension efficiently.

Usage

There are two ways to use the algorithm, either use one of the Optimizer::minimize / Optimizer::maximize functions :

let f = |v:&[f64]| v[0] + v[1] * v[2];
let input_interval = vec![(-10., 10.), (-20., 20.), (0, 5.)];
let nb_iterations = 100;

let (max_value, coordinates) = Optimizer::maximize(&f, &input_interval, nb_iterations);
println!("max value: {} found in [{}, {}, {}]", max_value, coordinates[0], coordinates[1], coordinates[2]);

Or use an iterator if you want to set exploration_depth to an exotic value or to have fine grained control on the stopping criteria :

let f = |v:&[f64]| v[0] * v[1];
let input_interval = vec![(-10., 10.), (-20., 20.)];
let should_minimize = true;

// sets `exploration_depth` to be greedy
// runs the search for 30 iterations
// then waits until we find a point good enough
// finally stores the best value so far
let (min_value, coordinates) = Optimizer::new(&f, &input_interval, should_minimize)
                                       .set_exploration_depth(10)
                                       .skip(30)
                                       .skip_while(|(value,coordinates)| value > 1. )
                                       .next().unwrap();

println!("min value: {} found in [{}, {}]", min_value, coordinates[0], coordinates[1]);

Divergences from the reference implementation

• The user defines the search space as an hypercube (which is then mapped to a simplex using this method).

• The exploration_preference (float) parameter has been replaced by an exploration_depth (unsigned integer) parameter. It represents how many split deep the algorithm can search before requiring higher-level exploration (0 meaning grid-search like exploration, 5 being a good default and large values (10+) being very exploitation/greedy focusses).

• There are two ways to call the algorithm, either by calling a single function or via an iterator which gives the user full control on the stopping criteria.

Potential future developements

Do not hesitate to ask for improvements. The list of things that could be done but will probably be left undone unless requested includes :

• Add (optional) serde support to save searches.

• Add an option to introduce noise in the point generation (?).

• Let the user suggest some points to speed-up the search (will require the ability to check wether a point is in a simplex or a triangularization algorithm).

• Let the user explore several points in parallel.

• Let the user perform the search offline.

• We could offer to integrate the project into the argmin optimization framework (to make the algorithm more accesible, future-proof and easier to compare with the state of the art).