How to find a Pareto front with MOGA

Goal: given a fitted surrogate and one or more objectives to minimize, find the set of non-dominated trade-off points (single objective → a single optimum; multiple objectives → a Pareto front).

Assumes a training file in the shape Getting started builds. MOGA searches over uniform-distribution bounds only — see below.

Run the optimization

from itis_sumo.evaluate.funs_evaluate import perform_moga_optimization

distributions = {
    "length": {"distribution": "uniform", "min": 0.0, "max": 1.0},
    "width": {"distribution": "uniform", "min": 0.0, "max": 1.0},
}
moga_kwargs = {
    "populationSize": 32,
    "maxIterations": 100,
    "seed": 42,
}
pareto = perform_moga_optimization(
    run_dir, training_file, ["length", "width"], distributions, ["y1"], moga_kwargs,
)
# pareto["y1"] — objective values on the front
# pareto["length"], pareto["width"] — the input points that produced them

distributions must use {"distribution": "uniform", "min": ..., "max": ...} for every entry in input_vars — MOGA raises ValueError on any non-uniform distribution (normal, constant, etc. aren't supported as search bounds here). Pass more than one name in output_responses to get a genuine multi-objective Pareto front instead of a single optimum.

moga_kwargs options

Key	Default	Notes
populationSize	32	JEGA initial population size
maxIterations	100	Generation budget — the actual evaluation-count control
fitnessType	"layer_rank"	or "domination_count"
replacementType	"elitist"	or "unique_roulette_wheel", "below_limit"
seed	12345	reproducibility
max_function_evaluations	—	accepted but silently ignored — deprecated on the Dakota side; use maxIterations to bound cost, not this

Reading the front

• Single objective: perform_moga_optimization still returns the same shape; treat the returned point(s) as converging toward the single optimum as maxIterations grows.
• Multiple objectives: every point in the returned front is non-dominated by construction (verified in V&V Category G) — no point is simultaneously worse on every objective than another returned point. There's no single "best" point; which one to pick is a decision about your objectives' relative priority, not something MOGA can answer.
• More iterations → a better (dominating) front, not just a bigger one — if you're unsure the front has converged, rerun with a higher maxIterations and check whether the new front dominates the old one.
• All returned points respect the distributions bounds by construction.

Details on the underlying Pareto-dominance utilities and the max_function_evaluations limitation: Reference → MOGA optimization.