NSGA‑II Rust Core

A Rust implementation of the NSGA‑II multi‑objective evolutionary algorithm.
The crate provides the core components required to build deterministic or stochastic evolutionary workflows.

---

Features

• non‑dominated sorting (parallel dominance matrix, tri‑state dominance)
• crowding distance (NaN‑safe, stable ordering)
• simulated binary crossover (SBX)
• polynomial mutation
• tournament selection
• constraint handling
• strict and auto hypervolume
• IGD (Inverted Generational Distance)
• per‑generation Pareto snapshots
• early stopping
• reproducible runs via optional RNG seeding

Objective evaluation and dominance checks run in parallel through Rayon.
Algorithm parameters use a builder‑style API.

---

Modules

The problem module is now split into a directory:

src/problem/
problem_trait.rs — core Problem trait
schaffer.rs — Schaffer N.1
zdt.rs — ZDT1, ZDT2, ZDT3
dtlz.rs — DTLZ1, DTLZ2, DTLZ3
kursawe.rs — Kursawe

All built‑in problems implement the same Problem trait and are re‑exported through problem::.

Other modules:

• evolve — NSGA‑II engine and RunResult
• sort — non‑dominated sorting and crowding distance
• data — individuals, dominance logic, feasibility rules
• metrics — strict HV, auto HV, IGD

All tests for Schaffer, ZDT, DTLZ, and Kursawe pass.

---

Usage

Built‑in Schaffer problem

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::Schaffer;fnmain(){let result = Evolution::new(Schaffer::default(),100,500).evolve();for ind in&result.pareto_front{println!("{:?} -> {:?}", ind.features, ind.objectives);}}

---

Built‑in ZDT problems

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::ZDT1;fnmain(){let result = Evolution::new(ZDT1::new(30),200,400).evolve();println!("Front size: {}", result.pareto_front.len());}

ZDT2 and ZDT3 follow the same pattern.

---

Built‑in DTLZ problems

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::DTLZ2;fnmain(){let result = Evolution::new(DTLZ2::new(3,12),300,600).evolve();println!("Front size: {}", result.pareto_front.len());}

DTLZ1 and DTLZ3 follow the same pattern.

---

Built‑in Kursawe problem

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::Kursawe;fnmain(){let result = Evolution::new(Kursawe::default(),100,200).evolve();println!("Front size: {}", result.pareto_front.len());}

---

Custom problem

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::Problem;structMyProblem{ranges:[(f64,f64);2],}implMyProblem{fnnew() -> Self{Self{ranges:[(0.0,1.0),(0.0,1.0)]}}}implProblemforMyProblem{fnnum_variables(&self) -> usize{2}fnnum_objectives(&self) -> usize{2}fnvariable_ranges(&self) -> &[(f64,f64)]{&self.ranges}fncalculate_objectives(&self,x:&[f64]) -> Vec<f64>{vec![x[0] + x[1],(x[0] - 1.0).powi(2) + (x[1] - 1.0).powi(2)]}}fnmain(){let result = Evolution::new(MyProblem::new(),200,300).evolve();println!("Front size: {}", result.pareto_front.len());}

---

Constrained problem

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::Problem;structConstrainedProblem{ranges:[(f64,f64);2],}implConstrainedProblem{fnnew() -> Self{Self{ranges:[(0.0,5.0),(0.0,5.0)]}}}implProblemforConstrainedProblem{fnnum_variables(&self) -> usize{2}fnnum_objectives(&self) -> usize{2}fnvariable_ranges(&self) -> &[(f64,f64)]{&self.ranges}fncalculate_objectives(&self,x:&[f64]) -> Vec<f64>{vec![x[0], x[1]]}fnconstraint_violations(&self,x:&[f64]) -> Vec<f64>{vec![2.0 - x[0] - x[1]]// x0 + x1 >= 2}}fnmain(){let result = Evolution::new(ConstrainedProblem::new(),100,200).evolve();println!("Front size: {}", result.pareto_front.len());}

---

Hypervolume tracking and early stopping

use rs_nsga2::evolve::Evolution;use rs_nsga2::problem::Schaffer;fnmain(){let result = Evolution::new(Schaffer::default(),100,500).with_reference_point(vec![10.0,10.0]).with_convergence_threshold(10,0.001).evolve();println!("Generations completed: {}", result.generations_completed);}

---

Sorting

Non‑dominated sorting

• tri‑state dominance (IDominatesJ, JDominatesI, None)
• parallel dominance matrix
• sequential front extraction
• feasibility rules:
  • feasible dominates infeasible
  • among infeasible: lower total violation dominates

Crowding distance

• stable, NaN‑safe sorting per objective
• infinite distance for boundary individuals
• finite‑difference distance for interior individuals

---

Metrics

Strict hypervolume

use rs_nsga2::metrics::hypervolume_2d_strict;

Reference point must dominate the front.

Auto hypervolume

use rs_nsga2::metrics::hypervolume_2d_auto;

Reference point is expanded when needed.

IGD

use rs_nsga2::metrics::igd;

Computes the average distance from the true front to the obtained front.

---

Algorithm

Each generation:

1. tournament selection
2. SBX crossover
3. polynomial mutation
4. parallel objective evaluation
5. merge parents and offspring
6. non‑dominated sorting
7. crowding‑distance truncation

Feasible solutions dominate infeasible ones.
Among infeasible solutions, lower total violation dominates.

---

RunResult

Field	Description
pareto_front	final Pareto front
history	per‑generation Pareto snapshots
hypervolume_history	HV per generation (NaN if no reference point)
igd_history	IGD per generation (NaN if no true front)
generations_completed	number of executed generations

---

Benchmarks

Included benchmarks:

• full NSGA‑II loop
• sorting only
• strict vs auto hypervolume
• IGD microbench

Run all:

cargo bench

Run a specific benchmark:

cargo bench --bench igd
cargo bench --bench sorting
cargo bench --bench evolution

Reports are written to target/criterion/.

---

Original authors (Python version)

• Pham Ngo Gia Bao
• Tram Loi Quan
• Quan Thanh Tho
• Akhil Garg

Rust port

• Giorgio