Skip to content

feat(resources): DTLZ1-4 and DTLZ7 reference fronts with 4, 6 and 8 objectives - #362

Merged
NicolasRodriguezUribe merged 1 commit into
mainfrom
feat/dtlz-many-objective-fronts
Nov 19, 2026
Merged

NicolasRodriguezUribe merged 1 commit into
mainfrom
feat/dtlz-many-objective-fronts

Conversation

@NicolasRodriguezUribe

Copy link
Copy Markdown
Contributor

Summary

Packages reference fronts of DTLZ1–DTLZ4 and DTLZ7 with 4, 6 and 8 objectives and makes the resolver find them by n_obj (VAMOS-Studio change request R4). Until now resolve_reference_front_path("dtlz2", None, n_obj=m) returned None for every m other than 2 and 3, so IGD, IGD+, epsilon and the Hausdorff distance had no automatic reference on many-objective DTLZ runs. VAMOS-Studio ships its own copies of jMetal's files for these cases and can drop them.

  • New files (src/vamos/resources/reference_fronts/, the default vamos source), written by experiments/scripts/generate_reference_fronts.py (dtlz1.4d … dtlz7.8d):
    • dtlz1.{4,6,8}d.csv and dtlz2.{4,6,8}d.csv: the largest Das-Dennis simplex lattice with at most 10 000 points (H = 37, 13 and 8: 9880, 8568 and 6435 points), scaled to sum(f) = 0.5 (DTLZ1) or projected onto the unit sphere (DTLZ2). This is the lattice of PlatEMO's reference fronts (PROBLEM.optimum = GetOptimum(10000) with UniformPoint's NBI method, which is single-layer for these M and also raises zeros to 1e-6). jMetal's DTLZ2.{4,6,8}D.csv (its DTLZ3 and DTLZ4 files are the same bytes) hold exactly the same points in the same order (checked: maximum difference 0.0), so the default source gives the IGD-like values of jMetal and Evolver for DTLZ2–DTLZ4.
    • dtlz7.{4,6,8}d.csv: on the front, the last objective is 2 (M - sum psi(f_i)) with psi(t) = t (1 + sin(3 pi t)). It separates, so a point is Pareto-optimal exactly when each f_i is a value that no smaller value beats in psi; the front is the product of those values (taken from a grid of 200 001 points of [0, 1]). About 10^5 such combinations are thinned to 5050 points by farthest-point selection, as for the three-objective file.
    • Files with more than three objectives are written with the shortest repr of each value (exact round trip), about half the size of np.savetxt's default format: 4.98 MB in all, the largest file 808 kB.
  • DTLZ3 and DTLZ4 share DTLZ2's front, so they are not packaged again (byte copies above 256 kB would also need hygiene exceptions): the resolver now falls back from a problem without a file of its own to the files of the problem whose front it shares, in either source. That is what it already did for jMetal's DTLZ3 and DTLZ4 files, now written once (_SHARED_FRONTS). The existing two- and three-objective files keep their names and are still the ones resolved (dtlz3.csv, dtlz4.2d.csv, jmetal-dtlz2.3d.csv, ...).
  • Not packaged: DTLZ5 and DTLZ6 (degenerate), other numbers of objectives, and jMetal's own DTLZ1 and DTLZ7 files with 4, 6 and 8 objectives. Those are sparse (183 to 370 points; DTLZ1's are rounded to five decimals and do not reach the front's corners: the smallest coordinate in DTLZ1.4D.csv is 0.02486). The jmetal source therefore keeps two and three objectives, and it never falls back to VAMOS's files.
  • Docs: docs/reference/problems.md: the DTLZ rows of the catalog table, the reference-front notes, a provenance row and a note on PlatEMO's and jMetal's files, all marked Development after 1.4.0:.

Sources: jMetal resources/referenceFronts/DTLZ{1,2,3,4,7}.{4,6,8}D.csv (local clone; identical at commit 1e4b83cae6, the commit VAMOS already cites), compared but not copied; PlatEMO Problems/Multi-objective optimization/DTLZ/DTLZ1.m, DTLZ2.m, Algorithms/Utility functions/UniformPoint.m and Problems/PROBLEM.m (BIMK/PlatEMO at e2d58cb), read for the lattice convention only.

Tests

  • tests/reference/test_reference_fronts.py:
    • every point of the new files lies on its analytic front within 1e-9 (sum(f) = 0.5, the unit sphere, DTLZ7's last objective) and each file has at least 1000 points; the existing mutual non-dominance test covers the new files;
    • test_many_objective_dtlz_fronts_are_reached_by_pareto_optimal_designs: for DTLZ1, DTLZ2, DTLZ3, DTLZ4 and DTLZ7 with 4, 6 and 8 objectives, a design rebuilt from each point of the resolved file, with the distance variables at their optimum, evaluates to that point within 1e-9 (DTLZ4 with its alpha = 100);
    • test_many_objective_dtlz_fronts_are_chosen_by_number_of_objectives: dtlz<k>.<m>d.csv for DTLZ1, DTLZ2 and DTLZ7, dtlz2.<m>d.csv for DTLZ3 and DTLZ4 (also for a key such as dtlz3.4d), None for DTLZ5, DTLZ6 and m + 1 objectives, no DTLZ3 or DTLZ4 many-objective files packaged, and the old names unchanged;
    • the check that a resolved front has the requested number of objectives now also runs every registered problem with 4, 6 and 8 objectives.
  • tests/reference/test_jmetal_reference_fronts.py: the jmetal source still gives None for DTLZ2 and DTLZ3 with more than three objectives, although VAMOS now has a front there.

Compatibility

Additive. New packaged files: a lookup that returned None (DTLZ1–DTLZ4 and DTLZ7 with 4, 6 or 8 objectives) now returns a front, so benchmarks, studies and Studio compute reference-front indicators for those runs where they skipped them before. Every lookup that returned a file returns the same file. No public API, schema or default changes.

Validation

Run from the worktree with PYTHONPATH set to its src (vamos.__file__ checked):

  • python -m pytest -q -n 4 tests/reference/test_reference_fronts.py tests/reference/test_jmetal_reference_fronts.py tests/foundation/test_packaging_data.py tests/ux/test_studio_catalog.py: 972 passed, 18 skipped (existing skips).
  • ruff check and ruff format --check on the changed Python: clean. python -m compileall: clean. git diff --check: clean.
  • python tools/typecheck.py --scope strict and --scope full: 0 diagnostics; ratchet new=0, increased=0.
  • python tools/check_agent_docs.py: passed. python tools/check_repository_hygiene.py with the new files staged: passed.
  • python -m pytest -m "not slow" -n 4 -p no:cacheprovider -rfE: 10876 passed, 253 skipped, 9 xfailed, 1 failed, tests/experiment/test_cli_ablation.py::test_cli_ablation_runs_and_writes_variants, a known failure under xdist on Windows that also fails on untouched main and passes alone.
  • python tools/health.py: all health checks passed.
  • mkdocs build --strict: built.
  • Generation: python experiments/scripts/generate_reference_fronts.py dtlz1.4d dtlz1.6d dtlz1.8d dtlz2.4d dtlz2.6d dtlz2.8d dtlz7.4d dtlz7.6d dtlz7.8d (about two minutes).

Changelog entry

Added (under [Unreleased])

  • Reference fronts: DTLZ1–DTLZ4 and DTLZ7 with 4, 6 and 8 objectives (dtlz1.<m>d.csv, dtlz2.<m>d.csv and dtlz7.<m>d.csv), which resolve_reference_front_path finds by n_obj; DTLZ3 and DTLZ4 resolve to DTLZ2's file. DTLZ1–DTLZ4 use the Das-Dennis lattice of PlatEMO's reference fronts, the points of jMetal's DTLZ2–DTLZ4 files with those numbers of objectives.

Checklist

Architecture health

  • I read the canonical rules in docs/dev/architecture_health.md relevant to this change.
  • I ran the local health command: python tools/health.py.
  • I updated docs/dev/architecture_health.md if I changed guardrails. (No guardrail changed.)

Public API & dependencies

  • Public API changes are intentional; snapshot updated via python tools/update_public_api_snapshot.py if needed. (No public export changes.)
  • No optional/heavy dependencies were added to core [project].dependencies.

Behavior & tests

  • Runtime behavior is unchanged (or explicitly documented as a bugfix). (Only lookups that returned None now return a front.)
  • Tests pass locally (fast tier, apart from the known Windows/xdist ablation failure above).

Changelog

  • CHANGELOG.md is unchanged, and the changelog entry is under a "Changelog entry" heading in this description.

Reports/retention

  • Raw audit/Goal evidence is external or a CI artifact; only durable conclusions are added to maintained docs.
  • Generated output follows docs/dev/repository_hygiene.md and python tools/check_repository_hygiene.py passes.

🤖 Generated with Claude Code

…bjectives

resolve_reference_front_path("dtlz2", None, n_obj=m) gave None for every
m other than 2 and 3, so IGD, IGD+, epsilon and the Hausdorff distance
had no automatic reference on many-objective DTLZ runs (VAMOS-Studio
request R4). VAMOS now packages its own samples with 4, 6 and 8
objectives, written by generate_reference_fronts.py:

- dtlz1.<m>d and dtlz2.<m>d: the largest Das-Dennis lattice with at most
  10 000 points (H = 37, 13 and 8: 9880, 8568 and 6435 points), scaled
  to sum(f) = 0.5 or projected onto the unit sphere. It is the lattice of
  PlatEMO's reference fronts (UniformPoint(10000, M)), and jMetal's
  DTLZ2-DTLZ4 files with those numbers of objectives hold the same
  points in the same order.
- dtlz7.<m>d: the last objective is 2 (M - sum psi(f_i)) with
  psi(t) = t (1 + sin(3 pi t)), which separates, so the front is the
  product of the values of each f_i that no smaller value beats in psi;
  a grid of such values, thinned to 5050 points.

DTLZ3 and DTLZ4 share DTLZ2's front, so their sample is DTLZ2's: the
resolver falls back from a problem without a file of its own to the
files of the problem whose front it shares, in either source, as it
already did for jMetal's files. Their two- and three-objective files
keep their names. Files with more than three objectives are written
with the shortest repr of each value, which reads back exactly. The
jmetal source keeps two and three objectives: jMetal's DTLZ1 and DTLZ7
files with 4, 6 and 8 objectives are sparse (183 to 370 points).

Tests: every point lies on the analytic front within 1e-9 and is the
image of a Pareto-optimal design as VAMOS evaluates it, and the resolver
returns dtlz<k>.<m>d.csv (DTLZ3 and DTLZ4: dtlz2.<m>d.csv) for 4, 6
and 8 objectives and nothing for DTLZ5, DTLZ6 or other numbers.

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 2026-11-18T23:42:32.237209Z 12b0f92 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-362-vamos-docs-preview.vamos-optimization.workers.dev

Built from 12b0f92. Canonical metadata points to the production .org domain; the preview itself remains isolated.

@NicolasRodriguezUribe
NicolasRodriguezUribe added this pull request to the merge queue Nov 19, 2026
Merged via the queue into main with commit a4d2ac3 Nov 19, 2026
15 checks passed
NicolasRodriguezUribe added a commit that referenced this pull request Nov 19, 2026
Keeps both sides of #362 (DTLZ fronts with 4, 6 and 8 objectives) and of
this branch (IDTLZ1 and IDTLZ2): the DTLZ rows of the problem table with
their new fronts and the IDTLZ rows, both provenance rows, both generator
functions and entries, and both sets of front residual checks.

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