VAMOS 1.2.0 release notes¶
Released 2026-11-14.
VAMOS 1.2.0 brings many built-in algorithms, operators, problems, reference fronts, indicators and statistics back to their published definitions. These scientific corrections change numerical results: read which results change before you compare new runs with runs from earlier versions. The release also adds 115 constrained benchmark problems, jMetal's constraint-handling techniques, the CCMO, C-TAEA, ToP and CMOEA-MS algorithms for constrained problems, faster Numba paths for the steady-state algorithms, and stricter checks that reject inputs earlier versions ignored or misapplied. The stable 1.x API, the omitted-engine NumPy default, and the 1.0.0 run and study schemas remain the compatibility baseline.
The package metadata now serves users of VAMOS: the research extra and the
third-party framework integrations are removed, and analysis and all
install only what users need. VAMOS 1.2.0 supports Python 3.10 to 3.14 and is
declared Production/Stable; it is the last minor release that supports
Python 3.10.
Which results change¶
A built-in component that departs from its published definition is a defect, and its fix ships even when it changes results; see scientific corrections. The changelog lists each correction with its pull request, whose description gives the measured effect. Rerun the experiments that use any of the following before comparing them with results of earlier versions.
| Component | What changed | Results affected |
|---|---|---|
| Default SBX | Probability 0.9 instead of 1.0 in NSGA-II, NSGA-III, SMS-EMOA, SPEA2, IBEA and RVEA, as in Deb et al. (2002), jMetal and pymoo | Runs with the default crossover; .crossover("sbx", prob=1.0, eta=...) reproduces 1.1.0 |
| IBEA | Default kappa 0.05; the epsilon indicator is normalized as in adaptive IBEA |
Every IBEA run with the default kappa or the epsilon indicator |
| AGE-MOEA | Mating and survival follow the author's PlatEMO code | Every AGE-MOEA run |
| MOEA/D | Neighbours replaced in random order; exact resume from checkpoints | Every MOEA/D run, and resumed runs |
| NSGA-III | Complete Das-Dennis lattice without divisions (the CLI and specification default); N >= H; offspring offered to the archive |
Runs without explicit divisions, and archives |
| RVEA | One survivor per occupied reference vector; parents drawn by permutation | Every RVEA run; populations can be smaller than pop_size |
| SMPSO | jMetal's coefficient sampling, personal-best rule and leader insertion; optimize, studies and vamos tune return the leaders archive by default |
Every SMPSO run, and SMPSO tuning scores |
| SMS-EMOA | Normalized reference point; a fixed user vector; a generation is pop_size children; offspring offered to the archive |
Default SMS-EMOA runs, HV-stopped runs and archives |
| SPEA2 | Lexicographic truncation and fitness-based filling, as in Zitzler et al. (2001); non-dominated result | Every SPEA2 run, and the knn archive pruning |
| Crowding and tournaments | An objective constant within a front is skipped; larger tournaments keep the best rank with NaN crowding; MooCore survival no longer overwrites earlier results | Runs in which these cases occur |
| Generations and budgets | Generation 0 is notified; NSGA-II immigration and the obl initializer count their evaluations |
hv_convergence stops, callbacks, and runs with immigration or obl |
| Operators | pcx, undx, simplex, linked_polynomial, power_law, insert, scramble, GradientRepair, ReflectRepair, WrappingRepair, scatter and obl follow their definitions; overrides of a default operator keep its other parameters |
Runs that use them or override a default operator |
| Problems | CEC2009 UF3 domain and UF9 distance; LSMOP1–4 and LSMOP9; RWA7 and RWA8; RE24 and RE25 domains; CF1 defaults to 10 variables | Results on these problems |
| Reference fronts | Regenerated ZDT, DTLZ, WFG1, WFG2, LSMOP9, C1-DTLZ3, UF6, UF9, DC2-DTLZ, RE, RWA3, RWA9 and RWA10 samples; fronts by number of objectives | IGD, IGD+, epsilon and normalized HV computed with them |
| Stopping | The ("hv", ...) criterion, --hv-threshold and hv_convergence use fixed references |
HV-stopped runs |
| Selection and constraints | front(), best() and top_k() rank feasible solutions first; archive and leader results keep their constraint values; top_k() defaults to source="auto"; top_k(method="kmeans") is deterministic |
Selected solutions of constrained results |
| CLI configuration | --nsga3-* options, every algorithm-block field and partial stopping or archive overrides reach the run |
CLI runs that set them |
| Statistics | Per-problem, feasible-only references in vamos bench and vamos ablation; ablation overrides applied; rank-sum tests; Nemenyi critical values; IGD+ of infeasible runs and merged fronts in Studio; the knee point; vamos profile timings and HV |
Benchmark and ablation tables, critical-difference diagrams, Studio comparisons, knee selections and profiles |
| Archives and indicators | Exact hypervolume contributions without MooCore; explicit reference points; exact deduplication; SPEA2 truncation in knn pruning; the archive hook stops re-offering pruned rows |
Pruned archives, and HV values without MooCore or without a reference point |
| Tuning | Racing's final choice among survivors and elitist restarts; single-fidelity model backends; SMAC3 and BOHB budgets and incumbents; Optuna pruning; failed evaluations; best_score |
Tuning results, budgets and summaries |
| Run provenance | Wheels inside Git projects, editable installs, packaged data, user problems, re-saved results and CLI requests are recorded correctly | Task IDs and verification outcomes of those runs |
New problems and APIs¶
- 115 constrained problems: LIR-CMOP1–14, DAS-CMOP1–9, CEC 2009 CF2–CF10, ZXH-CF1–16, the CRE and RCM real-world suites, C3-DTLZ4, and eight classic problems as jMetal defines them. The problems catalog lists them with the origin of every packaged front.
OptimizationResult.feasibility(), afeasible_onlykeyword on the selection helpers, andhv_reachedin the results of every algorithm.- NSGA-II and MOEA/D complete the ask/tell protocol with
initialize,should_terminateandresult, and every algorithm accepts the samemax_evaluations,n_genandhvtermination criteria. - The experimental tuning gains an
iracebackend throughRscript, and Studio's default LLM provider gets ageminiextra. - NSGA-II, NSGA-III, SMS-EMOA, AGE-MOEA, RVEA and SMPSO offer jMetal's
constraint-handling techniques through
constraint_handling, and MOEA/D its subproblem update criteria throughsubproblem_update_criterion, with matching*_constrainedtuning spaces. - Four algorithms for constrained problems follow jMetal's ports and match
their results: CCMO (
ccmo), the coevolutionary framework of Tian et al. (2021); C-TAEA (ctaea), the two-archive algorithm of Li et al. (2019); ToP (top), the two-phase framework of Liu and Wang (2019); and CMOEA-MS (cmoeams), the multi-stage algorithm of Tian et al. (2022). They are new in 1.2.0 and stay outside the stable API until a later release lists them.
Stricter validation¶
Typed algorithm configurations validate their values when they are built.
optimize, study planning and the CLI reject arguments, configuration keys,
dimensions, operators, checkpoints and undeclared constraint output that 1.1.0
ignored, applied in part or failed on later, and they check each algorithm's minimum population and number
of objectives before running. Stochastic components used directly without a
numpy.random.Generator raise ConfigurationError instead of drawing from an
unseeded one. Every error of the stable API derives from VAMOSError. Code
that passed such inputs to 1.1.0 now receives a ConfigurationError or a CLI
usage error that names the field to correct.
Performance¶
With engine="numba", the one-offspring steps of steady-state NSGA-II,
SMS-EMOA and MOEA/D run as compiled kernels, three to six times faster per
evaluation on ZDT1, with trajectories identical to the NumPy engine. The NumPy
engine sorts two-objective fronts in O(N log N), and import vamos loads the
experiment layer only on first use.
Optional dependencies¶
- The
researchextra and the experimental third-party NSGA-II baselines (--include-external,--external-problem-source, and thepymoo_nsga2,jmetalpy_nsga2andpygmo_nsga2algorithm names) are removed. VAMOS no longer depends on another optimization framework; install one yourself to compare against it. analysisinstalls only what VAMOS imports (pandas, matplotlib, plotly and scikit-learn). Notebooks need[analysis,examples].allinstalls the user extras without thedevanddocstooling.
Python support¶
VAMOS 1.2.0 is tested on Python 3.10 to 3.14, including Python 3.13 and 3.14 on Linux, Windows and macOS, and its package metadata declares the development status Production/Stable. Python 3.10 reached its end of life in October 2026: VAMOS 1.2 is the last minor release that supports it, and VAMOS 1.3.0 will require Python 3.11 or newer.
Deprecations and removals¶
best("knee")andtop_k(method="knee")warn; usebalanced_sum, the rule they always applied, orvamos.ux.api.knee_point_scores. The alias will be removed in VAMOS 2.0.allow_ref_expandis ignored with aFutureWarningin the HV termination criterion and inhypervolume().- Internal modules without callers and experimental tuning helpers without internal use are removed; none was part of the public API.
Installation and remaining scope¶
python -m pip install --upgrade "vamos-optimization==1.2.0"
python -m pip install --upgrade "vamos-optimization[compute]==1.2.0"
Multi-worker study coordination, full-source typing debt, independent
conformance validation of RVEA and SMPSO, and the welded_beam variant remain
documented limitations. See known limitations,
stability and versioning,
the 1.1.0 release notes, and
release verification.