Copper

A constraint programming solver.

A solver is a declarative framework, where combinatorial problems are described in terms of decision variables and binding constraints. The algorithms used to find a feasible solution or minimize an expression are abstracted away from the user. Focus on problem definition, create relationships between variables, and Copper will take care of performing the actual search.

Usage

Below is a formulation of the Knapsack problem. It maximizes the value of the items stored in a bag with a fixed maximum weight.

// Extension trait, used here to scale decision variables by a constant (`times` method)
use copper::views::ViewExt;

// Problem parameters
let item_weights = [10, 60, 30, 40, 30, 20, 20, 2];
let item_values = [1, 10, 15, 40, 60, 90, 100, 15];
let weight_max = 102;

// Model object, used to declare decision variables and constraints
let mut m = copper::Model::default();

// Binary decision variables: for each item, do I put it in the bag?
let xs: Vec<_> = m.new_vars_binary(item_weights.len()).collect();

// Sum of the weights of the selected items
let weight = m.sum_iter(xs.iter().zip(item_weights).map(|(x, w)| x.times(w)));

// Ensure the bag does not exceed its maximum weight
m.less_than_or_equals(weight, weight_max);

// Sum the value of the selected items
let value = m.sum_iter(xs.iter().zip(item_values).map(|(x, v)| x.times(v)));

// Find the selection of items that maximizes the bag's value
let solution = m.maximize(value).unwrap();

// Extract assignment for each decision variable: is this item in the optimal bag?
let is_item_in_bag = solution.get_values_binary(&xs);

assert_eq!(is_item_in_bag, vec![false, false, true, false, true, true, true, true]);
assert_eq!(solution[weight], 102);
assert_eq!(solution[value], 280);

You can find the step-by-step guide for a similar problem in the documentation for the Model struct.

Constraint programming or linear solvers?

Most integer programming solvers, be they open source or commercial, are restricted to linear constraints. This unlocks performant methods like the simplex algorithm, which make them much faster than constraint programming solvers on linear problems.

However, it requires users to formulate their constraints in terms of linear equations. This restriction can lead to awkward linearizations, iffy approximations, or existential dread.

Constraint programming imposes no such restriction on the relationships between decision variables. Complex, non-linear constraints can even help the propagation engine prune the search space further, and improve overall performance.

Why Copper?

• Straight and strict API: attention to developer experience makes Copper easy to pick up, and Rust's robust type system lets us expose strict APIs that are harder to misuse.
• Permissive license: Copper is developed under the MIT license, which imposes few limitations on end users. Free as in beer and speech.

Why not Copper?

• Performance: Copper is still quite early in its development, it cannot rival with mature solvers like Gecode or or-tools. Though written in Rust, it is not ⚡️ blazingly fast ™⚡️ (yet).
• Features: Copper currently. supports a limited number of variable types and constraints. You can extend its engine yourself, but other solvers will offer more out of the box.

Disclaimer

Copper is under heavy development. Some exposed APIs are subject to change, including the ones that relate to custom propagators and branching strategies.