---

name: problem-modeling description: "Define, validate, and troubleshoot pymoo optimization problems, built-in test problems, constraints, bounds, and evaluation output shapes." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

problem-modeling

Use this sub-skill when a task asks how to represent an optimization problem in pymoo: built-in benchmark problems, custom vectorized or elementwise objectives, functional objectives, bounds, constraint signs, equality constraints, Pareto front/set methods, evaluation shape validation, or confusing Problem warnings.

Route first

• Algorithm choice, minimize, termination, ask-and-tell, callbacks, and result interpretation belong to optimization-workflows.
• Mixed variables, sampling/crossover/mutation/repair, duplicate elimination, initialization, and custom operators belong to operators-and-variables.
• Vectorization vs parallel runners and optional joblib/dask/ray backends belong to performance-and-parallelization.
• Indicators, reference directions, decomposition, MCDM, convergence, and plots belong to analysis-and-visualization.
• Stay here for problem metadata, _evaluate contracts, objective/constraint sign conventions, built-in get_problem, and validation of F/G/H.

Fast operating checklist

1. Convert the math to pymoo conventions: every objective is minimized. Negate objectives that the source problem describes as maximized. Inequality constraints belong in out["G"] and are feasible when G <= 0; equality constraints belong in out["H"] and should be near zero.
2. Pick the problem style: use vectorized Problem for matrix-friendly NumPy/SciPy objectives; use ElementwiseProblem for one-candidate black-box calls; use FunctionalProblem for concise objective/constraint functions.
3. Set metadata in __init__: n_var, n_obj, n_ieq_constr, n_eq_constr, xl, xu, and optional vtype/vars must match the search space and outputs.
4. Return stable shapes: for vectorized multi-objective outputs, make out["F"] a 2-D matrix with one row per candidate and one column per objective. For constraints, align rows with F.
5. Validate before optimizing: call problem.evaluate(...) on tiny sample inputs, request F, G, and H, assert finite values, and inspect signs.
6. Use built-in problems for smoke tests: get_problem("zdt1"), get_problem("dtlz2", n_obj=3), Sphere(n_var=...), and constrained families are useful baselines before debugging custom code.
7. Treat deprecated n_constr carefully: this version maps it to n_ieq_constr with a warning; new problem code should use n_ieq_constr and n_eq_constr explicitly.

Open the bundled references

• API reference: signatures and output contracts for Problem, ElementwiseProblem, FunctionalProblem, get_problem, and evaluate.
• Problem definition workflows: built-in problems, vectorized/elementwise/functional definitions, constraints, Pareto fronts, bounds, and validation recipes.
• Troubleshooting: fixes for wrong constraint sign, objective maximization, shape errors, bounds warnings, NaNs, deprecated arguments, and missing Pareto fronts.

Bundled script

• scripts/validate_problem_shapes.py: safe CPU-only assertions for a vectorized constrained problem and an elementwise constrained problem.