---

name: analysis-and-visualization description: "Postprocess pymoo results with indicators, Pareto analysis, reference directions, decomposition, MCDM, convergence traces, and headless plots." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

analysis-and-visualization

Use this sub-skill when a pymoo task asks to interpret or postprocess completed optimization results: hypervolume, IGD/GD, epsilon indicators, KKTPM, R-metric, Pareto-front or Pareto-set analysis, reference directions, decomposition functions, MCDM/knee-point selection, convergence from res.history, or saving visualizations in a headless environment.

Route first

• Running minimize, choosing algorithms, termination, callbacks, checkpoints, or ask-and-tell optimization loops belongs to optimization-workflows.
• Defining problems, bounds, objective/constraint shapes, G <= 0, or Pareto front/set methods belongs to problem-modeling.
• Parallel evaluation, optional distributed backends, compiled-extension speed, and long-run resource tuning belongs to performance-and-parallelization.
• Stay here for objective-space quality metrics, post-run selection, reference direction matrices, scalarization of objective matrices, convergence plots, and static/animated visual summaries.

Fast operating checklist

1. Normalize the data contract: treat objective values as a finite 2-D minimization matrix F with shape (n_points, n_obj). Pareto fronts pf need the same number of columns. Hypervolume reference points must be worse than the relevant front in every objective.
2. Choose metrics by available evidence: use GD, GDPlus, IGD, IGDPlus, and epsilon only when a true or accepted approximate Pareto front is available. If no front is available, prefer hypervolume with an explicit reference point, history-based convergence, or KKTPM when differentiable problem gradients are available.
3. Separate final-set quality from convergence: final res.F or res.opt.get("F") scores one set; save_history=True stores snapshots in res.history for curves but can be memory intensive.
4. Use decomposition and MCDM for selection: normalize objective scales, then rank rows with ASF/AASF, weighted sum, Tchebicheff, PBI, pseudo weights, or high-tradeoff point detection. Keep selected indices tied back to the original res.X and res.F rows.
5. Generate reference directions deliberately: uniform/das-dennis use partition counts and only create achievable point counts; energy and reduction can target arbitrary counts; multi-layer, layer-energy, and incremental cover specialized layouts.
6. Make plots headless-safe: set a non-interactive Matplotlib backend such as Agg before importing pymoo visualization helpers, then call .save(...) instead of relying on .show().

Open the bundled references

• Analysis API reference: imports, minimal signatures, input requirements, output meaning, and version caveats for indicators, non-dominated sorting, reference directions, decomposition, MCDM, and visualization helpers.
• Postprocessing workflows: recipes for final indicator calculation, Pareto-front-unknown analysis, MCDM solution selection, convergence/history curves, and headless static/optional video plotting.
• Troubleshooting: fixes for reference-point, ideal/nadir, missing Pareto front, invalid shape, normalization, headless display, optional animation, KKTPM, R-metric, and decomposition mistakes.

Bundled scripts

• scripts/check_indicators.py: deterministic numeric checks for core indicators plus small sanity checks for reference directions, decomposition, and MCDM selection.
• scripts/save_scatter_plot.py: headless Agg scatter-plot save workflow with assertions that the image file was created.