Skip to content

VAMOS 1.4.0 release notes

Released 2026-11-17.

VAMOS 1.4.0 is a minor release of VAMOS 1.3.0. Its built-in tuning spaces become grammars: YAML files with a default for every parameter, which the tuning backends and Studio's algorithm forms read. MOEA/D and SMS-EMOA gain the choices that Evolver offers for them, so that a configuration tuned or published with Evolver can be expressed in VAMOS, as VAMOS-Studio requested. As VAMOS 1.3.0 announced, VAMOS Assist, Studio's Problem Builder and the LLM extras are removed. The default configurations of the algorithms give the results of 1.3.0, but seeded tuning runs draw different trials and persistent Optuna studies of 1.3.0 are not resumed: read which results change. The stable 1.x API, the omitted-engine NumPy default, and the 1.0.0 run and study schemas remain the compatibility baseline.

Tuning spaces as grammars

The VAMOS-CONFIG design turns each built-in tuning space into a grammar under the schema vamos.parameter-space 1.0.0. Tuning is Experimental, so these changes ship in a minor release.

  • Schema and validator. vamos.engine.tuning.grammar reads a space from YAML or JSON and rejects one that breaks a rule of the schema, naming the rule and the place. check_default_configuration checks that a space's defaults run as its algorithm's default() configuration. A configuration of a space has three forms that read back into it: the document vamos.parameter-configuration 1.0.0, the flat assignment that config_from_assignment reads, and the command line of Evolver's VAR_CONF.txt, with VAMOS's names or Evolver's.
  • Shipped grammars. The 53 built-in spaces are YAML files packaged in vamos/resources/parameter_spaces. Every parameter has a default; the spaces no longer tune pop_size, which is a run setting as in Evolver; the external archive is one choice, external_archive; operator rates hang from their operators; and the mutation probability is a factor of 1/n wherever its default is 1/n. Assignments saved from the spaces of VAMOS 1.3.0 still build the configurations they built.
  • Backends. ConfigSpace (SMAC3, BOHB), Optuna, irace, racing and random search read the grammar: they sample only active parameters, and Optuna, irace and racing try the default configuration first.
  • Studio. The algorithm forms are the grammar of the tuning space: they follow its tree, start from its defaults and edit each parameter within the range that tuning uses.

Evolver's choices for MOEA/D and SMS-EMOA

New options of MOEA/D, each off by default, with the definitions of jMetal 7.8, which Evolver uses (see the MOEA/D configuration):

  • subproblem_order="cyclic" visits the subproblems in index order (Evolver's cyclicIntegerSequence); the default stays the random permutation.
  • The DE crossover takes a scheme, rand_1_bin (the default), rand_1_exp, rand_2_bin or rand_2_exp, and a base: the current solution (the default, as MOEA/D-DE) or one more random parent, as Evolver's MOEA/D.
  • The aggregations augmented_tchebycheff, inverted_pbi (Sato, 2014) and inverse_weight_tchebycheff, which is jMetal's and Evolver's modified Tchebycheff, and normalize=True with epsilon for any aggregation, which normalizes the objectives with the ideal point and jMetal's estimate of the nadir point.

Every MOEA/D tuning space offers these options too: the order of the subproblems, the three aggregations with their parameters, and the normalization of the objectives, with the DE scheme and base under differential evolution in the real-coded spaces. They are off by default, so a space's defaults still build the configuration of 1.3.0.

The real-coded MOEA/D and SMS-EMOA tuning spaces, constrained ones included, offer what Evolver gives these algorithms and VAMOS runs: Sobol, opposition-based (Evolver's oppositionBased) and Halton initialization, the Lévy flight and power-law mutations, the Laplace and fuzzy crossovers (with BLX-αβ and whole arithmetic in SMS-EMOA), MOEA/D's choice between differential evolution and crossover and mutation, and distribution indices in [5, 400].

The tuning name table now reads Evolver's modifiedTschebyscheff as inverse_weight_tchebycheff: it read it as VAMOS's modified_tchebycheff, a different function. It reads Evolver's inverted PBI with normalization and its DE with a random base, as Evolver runs them, and to_evolver refuses the combinations that Evolver cannot express.

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
Tuning Every built-in space is a grammar: pop_size is no longer tuned, every parameter has a default, the backends draw only active parameters and try the default configuration first, and the MOEA/D spaces, the real-coded SMS-EMOA spaces and the mixed spaces offer more choices. Every seeded tuning run draws different trials. Stored assignments build the configurations they built.
Persistent Optuna studies vamos tune names them after the hash of the space it searches (__vamos_cli_final_population_hv_<hash>, SMPSO: __vamos_cli_leaders_hv_<hash>). A study of VAMOS 1.3.0 or earlier is not resumed; a changed space starts a new study.
Mixed variation MOEA/D, NSGA-III, SMS-EMOA and SPEA2 run the segment settings of their configuration (perm_crossover, real_mutation_prob, int_mutation_step, ...); they ran the default mixed operators whatever these said. Mixed-encoding runs of those algorithms that set a segment. Default configurations do not change.
Constrained tuning spaces The default epsilon_decay_fraction is 0.8, Takahama and Sakai's; it was 0.2. Tuning runs that rely on that default.
Configurations exchanged with Evolver Evolver's modifiedTschebyscheff reads as inverse_weight_tchebycheff, its inverted PBI with normalization and its DE with de_base: random; to_evolver refuses modified_tchebycheff, which Evolver does not have, and the combinations that it cannot express. MOEA/D configurations read from Evolver with these aggregations or with DE.
Studio MOEA/D's form applies mtch_rho to the modified Tchebycheff aggregation; it ignored it. Studio MOEA/D runs with that aggregation and another mtch_rho.

The default configurations of the algorithms, runs that set none of these, and replays of stored runs give the same results as before. MOEA/D's new options leave a run that does not set them unchanged, bit for bit, as tests pin.

Removed

As announced with VAMOS 1.3.0, the Experimental features that depended on LLM providers are removed:

  • vamos.assist and the vamos assist command, which now fails as unrecognized arguments (exit status 2);
  • Studio's Problem Builder page, which ran Python typed into Studio or drafted by an LLM provider; Studio runs registered problems only;
  • the openai, anthropic and gemini extras, also from all.

Define a problem with make_problem or vamos create-problem, and run it from Python, the command line or Studio.

Dependencies

  • PyYAML (pyyaml>=6.0) is a dependency of VAMOS, which reads its YAML tuning spaces at run time; it was in the dev extra.
  • With NumPy 2.4, the Numba engine keeps its compiled SBX, polynomial mutation and steady-state steps: NumPy 2.4 computes np.power(x, 0.5) as sqrt(x) and np.power(x, -1) as 1 / x, and the compiled power now takes the same shortcuts. VAMOS 1.3.0 fell back to the NumPy operators there, with the same results but slower. A NumPy build whose np.power uses vendor SIMD code still keeps the NumPy operators (see the known limitations).
  • The release CI tests NumPy 2.4.6 and SciPy 1.17.1 on every Python, 3.11 to 3.14. The declared minimums do not change.

Known limitations

See known limitations.