This library provides a metaheuristic framework for solving combinatorial optimization problems.
The following metaheuristics are included:
- local search (with several neighborhood exploration strategies)
- parallel local search
(the neighborhood is explored in parallel using
rayon
) - threshold accepting
- simulated annealing
- tabu search (and a faster parallel version)
The framework supports hierarchical objective, i.e., objectives that consists of multiple levels of linear combinations. The first level is first minimized and only for tie-breaks the next level is considered. This is useful to model hard-constraints as high-priority soft-constraints (via a violation measure), such that normally infeasible solutions are considered feasible. The solver than minimizes these constraints first until the violation is zero and then starts to optimize the remaining objective levels.
As an example we provide a simple implementation of the Traveling Salesman Problem (TSP) with the 3-opt neighborhood.
Suppose you have given a combinatorial optimization problem and defined a solution type. To run a local search solver you need to do the following four steps:
- Define the
Objective
for your problem by definingIndicators
and build a hierarchical objective ofLinearCombinations
of these indicators. - Define modifications for your solution type. The solution type should not be mutable, instead a modified clone should be returned.
- Implement the
Neighborhood
for the local search. - Initialize the
LocalSearchSolver
and run it.
We demonstrate these steps on a simple (but totally artificial) example, where the solution type consists of a fixed-size vector of integers.
struct Solution(Vec<i64>);
1. Define the Objective
for your problem.
For the example, we want to find permutation (i.e., 0 to 10 should appear exactly once) where the sum of squared differences between consecutive elements (cyclic) is minimized.
Hence, we define two Indicators
, namely PermutationViolation
and
SquaredDifference
, and build a hierarchical
objective where PermutationViolation
is minimized first and only for tie-breaks
SquaredDifference
is considered.
use rapid_solve::objective::{BaseValue, Indicator, Objective};
struct PermutationViolation;
impl Indicator<Solution> for PermutationViolation {
fn evaluate(&self, solution: &Solution) -> BaseValue {
let violation: i64 = (0..solution.0.len())
.map(|i| (solution.0.iter().filter(|&n| *n == i as i64).count() as i64 - 1).abs())
.sum();
BaseValue::Integer(violation)
}
fn name(&self) -> String {
String::from("PermutationViolation")
}
}
struct SquaredDifference;
impl Indicator<Solution> for SquaredDifference {
fn evaluate(&self, solution: &Solution) -> BaseValue {
let squared_diff: i64 = (0..solution.0.len())
.map(|i| (solution.0[i] - solution.0[(i + 1) % solution.0.len()]).pow(2))
.sum();
BaseValue::Integer(squared_diff)
}
fn name(&self) -> String {
String::from("SquaredDifference")
}
}
fn build_objective() -> Objective<Solution> {
Objective::new_single_indicator_per_level(vec![
Box::new(PermutationViolation),
Box::new(SquaredDifference),
])
}
In our example we use two modifications:
- Changing one entry to a number between 0 and 10.
- Swapping two entries.
The solution type should not be mutable, instead a modified clone should be returned.
For larger solution types the immutable data structures crate im
might increase
performance.
impl Solution {
fn change_entry(&self, index: usize, new_value: i64) -> Self {
let mut new_values = self.0.clone();
new_values[index] = new_value;
Solution(new_values)
}
fn swap(&self, index1: usize, index2: usize) -> Self {
let mut new_values = self.0.clone();
new_values.swap(index1, index2);
Solution(new_values)
}
}
3. Implement the Neighborhood
.
In our example we want to first try to change all entries and then try all swaps.
use rapid_solve::heuristics::local_search::Neighborhood;
struct ChangeEntryThenSwapNeighborhood;
impl Neighborhood<Solution> for ChangeEntryThenSwapNeighborhood {
fn neighbors_of<'a>(
&'a self,
solution: &'a Solution,
) -> Box<dyn Iterator<Item = Solution> + Send + Sync + 'a> {
let change_entry = (0..solution.0.len()).flat_map(move |i| {
(0..10).map(move |new_value| solution.change_entry(i, new_value))
});
let swap = (0..solution.0.len())
.flat_map(move |i| (0..solution.0.len()).map(move |j| solution.swap(i, j)));
Box::new(change_entry.chain(swap))
}
}
4. Initialize the LocalSearchSolver
and run it.
In the example only a local optimum is found, which is worse than the global optimum.
use rapid_solve::heuristics::local_search::LocalSearchSolver;
use std::sync::Arc;
let objective = Arc::new(build_objective());
let neighborhood = Arc::new(ChangeEntryThenSwapNeighborhood);
let solver = LocalSearchSolver::initialize(neighborhood, objective);
let initial_solution = Solution(vec![0; 10]);
let evaluated_local_minimum = solver.solve(initial_solution);
assert_eq!(
*evaluated_local_minimum.objective_value().as_vec(),
vec![BaseValue::Integer(0), BaseValue::Integer(36)]
);
assert_eq!(
*evaluated_local_minimum.solution().0,
vec![1, 0, 2, 4, 5, 7, 9, 8, 6, 3]
);
// one global optimum is [0, 2, 4, 6, 8, 9, 7, 5, 3, 1] with a squared differences of 34.
For a less artificial demonstration, we refer to the tsp-example.