Home/ Expertise/ 04 · Algorithm Design

Algorithm design.

Optimisation is only as good as the search engine behind it. I design and benchmark genetic algorithms, evolutionary methods, and hybrid strategies — including custom NDCI-aware fitness functions — and evaluate them across OEM, operator, and MRO perspectives so the chosen algorithm matches the decision context, not the trend cycle.

§ 01 / Live · six algorithms · DTLZ2 testbed

Pick one. Watch it converge.

Six full implementations on the same DTLZ2 benchmark (M=3, n=12). Switch between them with the dropdown — same population size, same evaluation budget, same metrics (HV via Monte-Carlo, IGD against a 200-point true-front sample). The cyan octant is the analytic Pareto front; the orange spheres are the algorithm's current rank-0 set.

Algorithm / EVOLUTIONARY Deb · Pratap · Agarwal · Meyarivan · 2002
NSGA-II · LIVE / POP 80 · OBJ 3 · VAR 12
GEN 0 · |F₁| 0 · HV 0.000 · IGD —
drag · scroll · double-click reset
Pareto front · F₁
Dominated population
True PF · DTLZ2 octant
CONVERGENCE · LIVE
TARGET HV → 0.74
HV0.000
IGD—
ΔHV+0.000 ΔIGD+0.000 PLATEAUNO
FRONT METRICS
SCHOTT · DEB
Mean f₁ on F₁—
Mean f₂ on F₁—
Mean f₃ on F₁—
Spacing · S(P)—
Norm error · g(x)—
EVAL0 RANK-00 REF(1.1, 1.1, 1.1)
DTLZ2 · Deb, Thiele, Laumanns, Zitzler 2002  ·  NSGA-II · Deb 2002  ·  NSGA-III · Deb & Jain 2014  ·  MOPSO · Coello & Lechuga 2002  ·  MOEA/D · Zhang & Li 2007  ·  NDCI-GA · Suslu, Ali, Jennions 2025  ·  Memetic GA · Ishibuchi 2003
§ 02 / Classical guidance · pursuit & evasion

Closed-form vs. evolutionary.

The cockpit above runs an evolutionary search — population, generations, Pareto sorting. The two simulations below run a classical alternative for the same family of problem: closed-form proportional-navigation guidance. Each engagement is a single trajectory shaped by one analytic control law (acmd = N · Vc · σ̇, where σ̇ is the line-of-sight rate and Vc the closing velocity). PN intercepts cleanly when the target is non-cooperative and slower than the pursuer; it loses the lock when target lateral acceleration exceeds the pursuer's. Same maths, opposite outcomes.

PN PURSUIT · CATCH / N 3.0 · amax 1.4 g
RANGE — · Vc — · tgo —
TRACKING
PURSUERV=1.00 TARGETV=0.55 EVADE0 g
PN EVASION · AVOID / N 3.0 · amax 0.7 g
RANGE — · Vc — · MISS —
TRACKING
PURSUERV=1.00 TARGETV=0.92 EVADE±1.5 g
Proportional navigation · pure-PN form acmd = N · Vc · σ̇, perpendicular to the pursuer velocity vector (Lockheed / Yuan / Adler 1956; Nesline 1962). Closing velocity Vc = −dR/dt. Time-to-go tgo = R / Vc. Evasion: target weaves with sinusoidal lateral acceleration; PN lock collapses when target lateral g exceeds pursuer's by ≳ 2× (Shinar & Steinberg 1977).
§ 03 / Comparative evaluation

Six algorithms, one benchmark.

Each radar plots a method across five quantitative MOEA metrics — HV (hypervolume), 1−IGD (proximity to true PF), 1−Δ (Deb spread / uniformity), Speed (gens to 95% HV), and Scale (M>5 viability). Values are normalised to [0,1] (1 = best) from canonical performance ranges on DTLZ2 (M=3, n=12) reported in the references below.

NSGA-II
EVO · 2-4 OBJ
DEFAULTYES DEB '02●
NSGA-III
EVO · M ≥ 5
MANY-OBJYES DEB '14●
MOPSO
SWARM
FASTESTYES COELLO '02●
MOEA/D
DECOMP.
SHAPE-AWAREYES ZHANG '07●
NDCI-GA
CUSTOM
DIAG.FOCUS SUSLU '25●
GA + LOCAL
MEMETIC
CONSTRAINTSHARD ISHIBUCHI '03●
BENCHMARK · DTLZ2 · M=3 · n=12
30 RUNS · 100 GEN · POP 80
ALGORITHM HV ↑ IGD ↓ Δ (SPREAD) ↓ GENS → 95% HV SCALE M ≥ 5 BEST USE
NSGA-II 0.7401 0.0011 0.420 ≈ 60 limited 2-4 obj default
NSGA-III 0.7395 0.0010 0.380 ≈ 65 5-15 obj many-objective
MOPSO 0.7180 0.0064 0.510 ≈ 35 limited smooth fronts, fast
MOEA/D 0.7415 0.0009 0.330 ≈ 50 moderate known PF shape
NDCI-GA 0.7350 0.0019 0.395 ≈ 70 limited sensor / diagnostic
GA + local 0.7388 0.0014 0.405 ≈ 95 limited hard constraints
Indicative ranges from canonical comparisons on DTLZ2 (M=3): Deb & Jain 2014, Zhang & Li 2007, Coello & Lechuga 2002. Δ = Deb spread metric. HV reference = (1.1, 1.1, 1.1). 95% HV-target = 0.95 × measured asymptote (≈ 0.74).
PROBLEMDTLZ2 METRICSHV · IGD · Δ · S DECISIONCONTEXT-DEPENDENT
§ 04 / Methods
EVOLUTIONARY
NSGA-II

Non-dominated sorting genetic algorithm. Default workhorse for 2–4 objectives.

EVOLUTIONARY
NSGA-III

Reference-point variant for many-objective problems (5+).

SWARM
MOPSO

Multi-objective particle swarm — fast on smooth landscapes.

DECOMP.
MOEA/D

Decomposition-based — strong when objective shape is known.

CUSTOM
NDCI-aware GA

Bespoke fitness using the Normalised Diagnostic Contribution Index.

HYBRID
GA + local search

Memetic methods for ill-conditioned constraint sets.

§ 05 / Benchmark perspectives
OEM
Lifecycle cost · certification weight
Operator
Availability · ops cost · safety
MRO
Diagnosability · turnaround time
§ / Contact

Need the right algorithm?

If your team is reaching for whatever's in the closest library, that's a sign you'd benefit from someone who's benchmarked the alternatives in production conditions.

[email protected] →