VAMOS 1.5.0 release notes¶
Released 2026-11-19.
VAMOS 1.5.0 is a minor release of VAMOS 1.4.0. It adds AGE-MOEA-II and the authors' reading of RVEA as variants, one source of reference vectors for MOEA/D, NSGA-III and RVEA, the CTP and IDTLZ test problems, DTLZ reference fronts with up to eight objectives, and the approximate hypervolume in the experimental analysis facade. It corrects NSGA-III's parent selection, which changes the results of every NSGA-III run, and the packaged MOEA/D weight files, which changes MOEA/D's results with four objectives and 100 to 119 individuals. Read which results change before comparing results with those of earlier versions. The stable 1.x API, the omitted-engine NumPy default, and the 1.0.0 run and study schemas remain the compatibility baseline.
NSGA-III selects its parents as configured¶
NSGA-III ran a binary tournament on rank and crowding distance, as NSGA-II,
whatever its configured selection, "random" included. Crowding distance is
not part of NSGA-III, so VAMOS's NSGA-III was not the published algorithm. Its
parents now follow the configuration:
selection("random")is the default ofNSGAIIIConfig.default(), the command line and the tuning spaces: parents are drawn at random, as in Deb and Jain (2014). With constraints and feasibility first, each parent wins a binary tournament on the constraint violation, decided at random between feasible solutions, as in Jain and Deb's constrained NSGA-III and in pymoo.selection("tournament", size=k)is jMetal's NSGA-III tournament on rank, with ties drawn at random.- Any other selection method or option raises
ConfigurationError, and--selection-pressureno longer applies to NSGA-III. - The NSGA-III tuning spaces choose
selection,randomortournamentwithtournament_size, as Evolver's space does; assignments of earlier spaces still build their tournament.
On ZDT4 with 100 individuals and 25,000 evaluations (seeds 0 to 9), the mean hypervolume of the default configuration goes from 0.863 to 0.836. On DTLZ2 with three objectives, VAMOS's mean hypervolume is within 0.1 % of pymoo's NSGA-III. See variants and their effect.
Reference and weight vectors¶
MOEA/D, NSGA-III and RVEA take their vectors from one kind of option, as jMetal and Evolver configure them; see reference and weight vectors.
- MOEA/D's
weight_vectors, NSGA-III'sreference_directionsand RVEA's newreference_vectorsaccept a file, with whitespace or commas between the components, a directory with jMetal'sW<m>D_<n>.datfiles, the lattice ofdivisions, or, with the newinner_divisions, Deb and Jain's two-layer lattice (156 vectors with eight objectives, 3 and 2 divisions). - The vectors are the population, so RVEA can run 100 individuals with three
objectives from
W3D_100.dat, and the three algorithms can start from the same vectors. vamos.algorithms.moead_populations(n_obj)lists the populations that MOEA/D runs without weight options.- Every packaged weight file loads, also from the command line, where 14 of the
31 failed.
W4D_100.datheld its 100 vectors twice and was skipped, so MOEA/D with four objectives and 100 to 119 individuals ran 84; it now runs the 100 vectors of the file. - A run that names a file records the file it read and the
sha256of its vectors, and replay refuses the file if its vectors have changed, where VAMOS 1.4.0 replayed whatever the file then held. The run still verifies without the file. - Validation is stricter: a path that does not exist, an unknown key in
weight_vectorsorreference_directions, and a two-layer lattice whose layers repeat a direction raiseConfigurationError. NSGA-III generated a lattice and saved it at a missing path, and MOEA/D created a missing directory for its lattice.
Algorithm variants¶
- AGE-MOEA-II (Panichella, 2022) is a variant of AGE-MOEA:
AGEMOEAConfig.builder().variant("agemoea_ii"), andreplacement: agemoea_iiin every AGE-MOEA tuning space (Evolver'sagemoeaII). Its survival estimates the front's exponent by Newton-Raphson from the point nearest to the ideal point and measures diversity with geodesic distances. Its exponents, distances and survival scores are checked against jMetal'sAGEMOEA2EnvironmentalSelection. The default,"agemoea", does not change; see AGE-MOEA-II. - RVEA's
variant("platemo")runs the authors' reading of the survival, that of PlatEMO and jMetal: the number of objectives in the angle penalty also with two objectives, translation by each generation's minimum, and adaptation of the vectors to the population's range. The default,"pymoo", gives the results of VAMOS 1.4.0, bit for bit; see RVEA.
Problems and reference fronts¶
- CTP1–CTP7 (Deb, Pratap and Meyarivan, 2001), registered as
ctp1–ctp7with the paper's definition and parameters and two-objective reference fronts generated from it;vamos.problems.CTPbuilds other instances. Their constraint is the paper's difference, not the ratio of the NSGA-II code that pymoo uses. - IDTLZ1 and IDTLZ2, the inverted DTLZ1 and DTLZ2 of Jain and Deb (2014) as
PlatEMO defines them, registered as
idtlz1andidtlz2with any number of objectives, and reference fronts for two and three objectives. - DTLZ reference fronts with 4, 6 and 8 objectives for DTLZ1–DTLZ4 and
DTLZ7, which the resolver finds by
n_obj, so IGD, IGD+, epsilon and the Hausdorff distance have an automatic reference on these runs. DTLZ1–DTLZ4 use the Das-Dennis lattice of PlatEMO's reference fronts, whose points are those of jMetal's DTLZ2–DTLZ4 files.
The problems catalog lists 260 registered problems.
Approximate hypervolume¶
The experimental vamos.ux.api exports HVIndicator, HVApproxIndicator,
HV_APPROX_METHODS and IndicatorResult. HVApproxIndicator.compute passes
seed and method to MooCore's hv_approx and records the samples, the seed
and the method in its details; without them it computes the value it computed
before. The exact hypervolume can take minutes from about six objectives; see
hypervolume, exact and approximate.
Faster unbounded archive¶
An unbounded external archive keeps the rows that each generation offers and merges them when the archive is read, instead of refiltering itself every generation. The rows, their order and every result are the same. A run that reads its archive only at the end filters it once: 0.28 s instead of 85 s for a 112,000-row archive over 200,000 evaluations.
Which results change¶
The changelog lists each change with its pull request. Rerun or redo the following before comparing them with results of earlier versions.
| Area | Change | Rerun |
|---|---|---|
| NSGA-III | Parents are drawn at random by default, or by a tournament on rank when configured; NSGA-III ran a tournament on rank and crowding distance whatever its selection. | Every NSGA-III run, by default and with a tournament: Python API, command line, studies and tuning. |
| MOEA/D with four objectives | A population of 100 to 119 runs the 100 vectors of W4D_100.dat instead of the 84 of the six-division lattice, and vamos tune gives 100 as the effective population. |
MOEA/D runs and tuning with four objectives and 100 to 119 individuals. |
| Tuning spaces | The NSGA-III spaces choose selection (random, the default, or tournament with tournament_size), and the AGE-MOEA spaces offer replacement (agemoea, the default, or agemoea_ii). |
Seeded tuning runs of the NSGA-III and AGE-MOEA spaces draw different trials. Stored assignments build the configurations they built. |
| Vector options | A path that does not exist, an unknown key and a two-layer lattice whose layers repeat a direction raise ConfigurationError; NSGA-III generated a lattice at a missing path, and MOEA/D created a missing directory. |
Configurations that relied on these fallbacks fail; name an existing file or directory. |
The other algorithms, AGE-MOEA and RVEA with their default variants, MOEA/D with other populations, runs with an unbounded external archive, and the approximate hypervolume without the new options give the same results as with VAMOS 1.4.0.
Dependencies¶
The declared dependencies do not change. CI builds the documentation with mkdocs-material 9.8.1.
Known limitations¶
See known limitations.