NSGA RS is a Rust framework for solving multi-objective optimisation problems using the NSGA family of multi-objective evolutionary algorithms (MOEAs). It gives you fast, parallel, and fully serialisable optimisation pipelines — from problem definition through to Pareto-front visualisation.

NSGA2 — 2-objective Schaffer	NSGA3 — 3-objective DTLZ1

• Three built-in algorithms — NSGA2 (Deb et al, 2002), NSGA3 (Deb & Jain, 2014), and AdaptiveNSGA3 (Jain et al, 2014)
• Flexible problem definition — minimise or maximise any number of objectives; constrained or unconstrained
• Rich variable types — real, integer, boolean, and categorical (choice) variables, each with optional bounds
• Parallel evaluation — multi-threaded objective and constraint evaluation via Rayon
• Resumable runs — export the full population history as JSON and resume from any checkpoint
• Python bindings — exposes a PyO3 interface so you can wrap the library in your own Python package
• Hypervolume metric — calculate hypervolume directly from individuals, values, or serialised JSON files
• Flexible stopping conditions — stop by generation count, function evaluations, elapsed time, or combinations using Any / All

Add the crate to your Cargo.toml:

[dependencies]
nsga_rs = "*"

The example below solves Schaffer's problem, a classic 2-objective test, using NSGA2.

The problem aims to minimise the following 2 objectives:

• f₁(x) = x²
• f₂(x) = (x - 2)²

The problem has 1 variable (x) bounded to -1000 and 1000. The optional solution is expected to lie in the [0; 2] range.

The problem is implemented below using the SCHProblem struct. When an algorithm runs, it first generates a set of potential solutions for the problem variables (in this case x). It then calculates the objectives (f₁(x) and f₂(x)) in the Evaluator trait exposed by this library.

#[derive(Debug)]
pub struct SCHProblem;

impl SCHProblem {
    /// Create the problem for the optimisation.
    pub fn create() -> Result<Problem, OError> {
        // define the objectives
        let objectives = vec![
            Objective::new("x^2", ObjectiveDirection::Minimise),
            Objective::new("(x-2)^2", ObjectiveDirection::Minimise),
        ];
        // define the variable
        let variables = vec![VariableType::Real(BoundedNumber::new(
            "x", -1000.0, 1000.0,
        )?)];
        // the problem has no constraints
        let constraints = None;

        let e = Box::new(SCHProblem);
        Problem::new(objectives, variables, constraints, e)
    }

    /// The first objective function
    pub fn f1(x: f64) -> f64 {
        x.powi(2)
    }

    /// The second objective function
    pub fn f2(x: f64) -> f64 {
        (x - 2.0).powi(2)
    }
}

// Implement the function to evaluate the objectives and constraints. The `evaluate`
// function below receives the individuals which contain the variables/solutions `x` 
// proposed by the algorithm. The function must return the evaluated objectives and
// constraints in the `EvaluationResult` struct.
impl Evaluator for SCHProblem {
    fn evaluate(&self, i: &Individual) -> Result<EvaluationResult, Box<dyn Error>> {
        let x = i.get_variable_value("x")?.as_real()?;
        let mut objectives = HashMap::new();
        objectives.insert("x^2".to_string(), SCHProblem::f1(x));
        objectives.insert("(x-2)^2".to_string(), SCHProblem::f2(x));
        Ok(EvaluationResult {
            constraints: None,
            objectives,
        })
    }
}

The code below set up the NSGA2 algorithm with 100 individuals and will stop when 250 population generations are reached.

...

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Setup the NSGA2 algorithm
    let args = NSGA2Arg {
        // use 100 individuals and stop the algorithm at 250 generations
        number_of_individuals: 100,
        stopping_condition: StoppingConditionType::MaxGeneration(MaxGeneration(250)),
        // use default options for the SBX and PM operators
        crossover_operator_options: None,
        mutation_operator_options: None,
        // no need to evaluate the objective in parallel
        threads: NumThreads::None,
        // do not export intermediate solutions
        export_history: None,
        resume_from_file: None,
        // to reproduce results
        seed: Some(10),
    };
    let mut algo = NSGA2::new(problem, args)?;
    // run the algorithm
    algo.run()?;

    // Export the final population to JSON
    algo.save_to_json(&PathBuf::from("."), Some("SCH_2obj"))?;
    Ok(())
}

Run it with:

cargo run --example nsga2_sch --release

The algorithm exports a JSON file (e.g. SCH_2obj_NSGA2_gen250.json) containing the full population.

All examples live in the examples/ folder and can be run with cargo run --example <name> --release.

Full API documentation is on docs.rs.

Runs can be stopped by:

• MaxGeneration — a fixed number of generations
• MaxFunctionEvaluations — a total number of objective evaluations
• Elapsed — a wall-clock duration
• Any / All — combine multiple conditions with OR / AND logic

Set export_history in the algorithm args to write JSON snapshots at regular intervals, or call save_to_json once at the end. To resume from a checkpoint, pass the snapshot path via resume_from_file.

With the library, you can:

• plot the Pareto front chart for problems with 2 or 3 objectives from the exported data. Check out the SCH solved with NSGA2 or DTLZ1 solved with NSGA3 or inverted DTLZ1 solved with Adaptive NSGA3.
• plot the hyper-volume metric to track convergence using Hypervolume::plot_from_files. Check out the convergence example.

The library includes a PyO3 interface, so you can bundle it into your own Python package using Maturin. There is no pre-built package on PyPI, the bindings are provided as a starting point for you to build and distribute as needed.

This project is licensed under the MIT License.