Skip to content

feat(agemoea): AGE-MOEA-II as AGE-MOEA's variant, checked against jMetal - #359

Merged
NicolasRodriguezUribe merged 2 commits into
mainfrom
feat/agemoea-ii
Nov 18, 2026
Merged

NicolasRodriguezUribe merged 2 commits into
mainfrom
feat/agemoea-ii

Conversation

@NicolasRodriguezUribe

Copy link
Copy Markdown
Contributor

Summary

VAMOS runs AGE-MOEA-II (Panichella, GECCO 2022, doi:10.1145/3512290.3528732) as a variant of AGE-MOEA, as Evolver does: AGEMOEAConfig.builder().variant("agemoea_ii"), and replacement: agemoea_ii in the tuning spaces (Evolver's replacement: agemoeaII). The default, "agemoea", is unchanged, so every existing configuration and seeded run gives the same result.

The reason is practical: the reference fronts of the CRE and RCM suites are being regenerated with VAMOS, and their algorithm portfolio, until now run with Evolver, includes AGE-MOEA-II with CDP and with the epsilon-constrained method.

Changes

  • Survival (agemoea/geometry.py). AGE-MOEA-II changes two steps of the first-front score. They are the two that jMetal's AGEMOEA2EnvironmentalSelection, written by the author, overrides:

    • the exponent p: Newton-Raphson from p = 1 (precision 0.001, at most 100 steps) on log(sum |a_i|^p), from the non-extreme point nearest to the ideal point (AGE-MOEA takes the point nearest to the diagonal and a closed form);
    • the diversity: geodesic distances on the L_p manifold. Both points and their midpoint are projected on it, and the distance is the length of the path through the projected midpoint.

    Everything else is AGE-MOEA's, with that p: the extreme points, the normalization, the later fronts, the tournament, the constraint handling (feasibility and the six techniques) and the numba greedy kernel.

  • Configuration. AGEMOEAConfig.variant is a field added after 1.0, so it goes last, and the builder gains variant(...). An unknown name raises ValueError. AGEMOEAVariant, AGEMOEA_VARIANTS and resolve_agemoea_variant live in config/types.py.

  • Tuning (second commit):

    • all six AGE-MOEA spaces gain replacement (agemoea, the default, or agemoea_ii);
    • the bridge sets variant;
    • parameter_names.py exchanges the parameter and the value agemoeaII with Evolver;
    • the parameter-space snapshot was regenerated with tools/update_parameter_spaces.py.
    • agemoea_constrained now passes 100 000 discrete combinations, so it joins _LARGE_SPACES: the exhaustive build runs in the slow tier (it passes) and the fast tier samples it.
  • Docs: the AGE-MOEA page (setting and reference), an "AGE-MOEA-II" section in the algorithm reference, and the tuning guide (the AGE-MOEA spaces and the name table). All carry the "Development after 1.4.0" marker.

Checked against jMetal

tools/agemoea2_reference_values.py and tools/jmetal_values/AgeMoea2Values.java follow the pattern of the MOEA/D criteria (#258). They write tests/engine/data/agemoea2_jmetal_values.json from jMetal's main at 46fb0a61f7, the commit pinned for the MOEA/D fixture, with the AGE-MOEA selection classes compiled from that commit and placed first on the classpath. There are 119 cases:

  • 44 exponents and 19 distance matrices: they agree with jMetal on every case (relative 1e-12). The fronts are points of L_q spheres with q from 0.4 to 6, and include zero coordinates and coordinates above 1.
  • 56 survival fronts, raw objectives with 2 to 5 objectives: the AGE-MOEA-II scores agree on every front where VAMOS's AGE-MOEA itself agrees with jMetal's. That is 31 of the 56 fronts, and the test requires at least half. The other fronts differ in AGE-MOEA's shared steps, not in AGE-MOEA-II's:
    • 20 pick other extreme rows, because VAMOS searches them as PlatEMO and pymoo do;
    • 2 tie in the greedy selection on exactly linear fronts, where jMetal keeps the last tie and VAMOS the first.

Declared departures (also in the docs):

  • p is capped at 20, as in AGE-MOEA and in pymoo's AGE-MOEA-II; jMetal does not cap it.
  • A point at the ideal point is not projected; jMetal would produce NaN.
  • pymoo takes the straight Euclidean distance when 0.95 < p < 1.05; VAMOS, like jMetal, always goes through the midpoint. Both coincide at p = 1.

A behavioural check, not part of the tests. Hypervolume over 10 seeds, population 100, NumPy engine, reference point 1.1, against pymoo 0.6.1.6's AGEMOEA2 with the same operators:

Problem (evaluations) VAMOS agemoea VAMOS agemoea_ii pymoo AGEMOEA2
ZDT1, 30 variables (20 000) 0.8707 ± 0.0002 0.8705 ± 0.0001 0.8690 ± 0.0002
DTLZ2, 3 objectives (30 000) 0.7457 ± 0.0013 0.7336 ± 0.0157 0.7371 ± 0.0098

VAMOS's AGE-MOEA-II is within seed variation of pymoo's. On DTLZ2 both show the same lower mean and larger spread than AGE-MOEA.

Validation

Run on Windows 11 with Python 3.12, following AGENTS.md:

  • Typing: python tools/typecheck.py --scope strict and --scope full pass, with 0 diagnostics and ratchet new=0, in a .[dev]-only environment with constraints/ci.txt.
  • Health and docs:
    • python tools/health.py: all health checks passed;
    • mkdocs build --strict: passes;
    • python tools/check_agent_docs.py: passes;
    • python tools/update_parameter_spaces.py --check: the grammars are valid and the snapshot is current.
  • Lint: ruff check src/vamos tests, ruff format --check on the changed files, compileall on the changed files and git diff --check all pass.
  • Tests:
    • pytest -m "not slow" -n 6: 10 839 passed, 2 failed. Both are fixed in this branch:
      • the tuning guide's name table now lists replacement;
      • tests/experiment/test_cli_ablation.py::test_cli_ablation_runs_and_writes_variants is pre-existing. It fails the same way (SystemExit: 2) on an untouched origin/main under -n 6, and passes alone. It is a MOEA/D ablation, which this branch does not touch.
    • pytest -m slow, without the exhaustive combination builds: 70 passed and 1 failure, the _LARGE_SPACES threshold above, now fixed.
    • The exhaustive and sampled builds of every AGE-MOEA space pass.
    • Not run to the end locally: the exhaustive builds of the other large spaces (NSGA-II, SMS-EMOA and MOEA/D). I stopped them after 40 minutes; this branch does not change those spaces.
    • tests/engine/test_agemoea_ii.py: 88 tests.

Changelog entry

[Unreleased] — Added

  • AGE-MOEA-II (Panichella, GECCO 2022) as a variant of AGE-MOEA:
    • AGEMOEAConfig.builder().variant("agemoea_ii"), and replacement: agemoea_ii in every AGE-MOEA tuning space (Evolver's replacement: agemoeaII);
    • its survival estimates the front's exponent by Newton-Raphson from the point nearest to the ideal point, and measures diversity with geodesic distances on the estimated manifold;
    • its exponents, distances and survival scores are checked against jMetal's AGEMOEA2EnvironmentalSelection;
    • the default, "agemoea", is unchanged.

🤖 Generated with Claude Code

AGEMOEAConfig gains variant ("agemoea", the default, or "agemoea_ii"), with
its builder method. AGE-MOEA-II (Panichella, GECCO 2022) changes two steps of
the first-front score, the two that jMetal's AGEMOEA2EnvironmentalSelection
overrides: the exponent p, found by Newton-Raphson from the non-extreme point
nearest to the ideal point, and the diversity, measured with geodesic
distances through the projected midpoints. Everything else is AGE-MOEA's.

tools/agemoea2_reference_values.py and tools/jmetal_values/AgeMoea2Values.java
record jMetal's exponents, distances and survival scores (main at 46fb0a61f7)
in tests/engine/data/agemoea2_jmetal_values.json; the exponents and distances
agree on every case, and the scores wherever AGE-MOEA's shared steps agree
with jMetal's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…acement

Every AGE-MOEA space (real, constrained, binary, integer, permutation and
mixed) offers replacement, Evolver's choice between AGE-MOEA's survival
(agemoea, the default) and AGE-MOEA-II's (agemoea_ii, Evolver's agemoeaII).
The bridge sets AGEMOEAConfig.variant, and the name table exchanges the
parameter and its values with Evolver. The default still builds the
configuration of VAMOS 1.4.0.

agemoea_constrained now has more than 100 000 discrete combinations, so its
exhaustive build moves to the slow tier and the fast tier samples it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-18T21:41:55.990410Z 74c4231 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@github-actions

Copy link
Copy Markdown
Contributor

VAMOS documentation preview

https://pr-359-vamos-docs-preview.vamos-optimization.workers.dev

Built from 74c4231. Canonical metadata points to the production .org domain; the preview itself remains isolated.

@NicolasRodriguezUribe
NicolasRodriguezUribe added this pull request to the merge queue Nov 18, 2026
Merged via the queue into main with commit a4c8461 Nov 18, 2026
15 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant