Skip to content

VAMOS 1.3.0 release notes

Released 2026-11-16.

VAMOS 1.3.0 is a minor release of VAMOS 1.2.1. It requires Python 3.11 or newer, as VAMOS 1.2.0 announced. It corrects operators and settings that departed from their published definitions, which changes the results of affected experiments: read which results change before you compare new runs with earlier ones. It rebuilds the built-in tuning spaces and racing's statistics, gives every algorithm one live-callback contract, makes durable studies robust to interruptions, and adds a wall-clock time limit, problem options by name, jMetal's reference fronts and MOEA/D configurations as options, a final subset of the external archive, and an access token for VAMOS Studio. The stable 1.x API, the omitted-engine NumPy default, and the 1.0.0 run and study schemas remain the compatibility baseline.

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 change with its pull request. Rerun the experiments that use any of the following before comparing them with results of earlier versions.

Area Change Rerun
Non-uniform mutation non_uniform is Michalewicz's mutation, as in jMetal: a variable moves towards a random bound by a fraction of its distance to it and never leaves the bounds. Runs with mutation="non_uniform".
Gaussian and Cauchy mutations sigma and gamma are fractions of each variable's range, as in pymoo, and gaussian defaults to sigma = 0.1. On a range of 1 the draws are the same. Runs with gaussian or cauchy on variables whose range is not 1.
Integer and mixed crossovers The integer SBX rounds instead of truncating towards zero, the integer arithmetic crossover and the real part of the mixed crossover give two complementary children instead of the parents' mean, and blx_alpha_beta defaults to alpha = beta = 0.5. Mixed problems with the default crossover, integer runs with SBX or arithmetic, and blx_alpha_beta without explicit parameters.
mutation_prob_factor MOEA/D, NSGA-III, SMS-EMOA, SPEA2, AGE-MOEA and RVEA apply it, as NSGA-II and IBEA did. Runs of those six algorithms that set it.
Command-line NSGA-III vamos --algorithm nsgaiii and configuration files use SBX with distribution index 30, as optimize and Deb and Jain (2014) do. Command-line NSGA-III runs.
Tuning spaces and racing The spaces drop dimensions that never reached the algorithm, take Evolver's names and ranges, and add choices; racing eliminates with F-race. Seeded tuning runs draw other configurations; stored assignments build the same configurations.
Live callbacks Every algorithm sends the same payload, the whole population and stats["evals"], and MOEA/D reports one generation per pop_size evaluations instead of one per evaluation. Code that reads live-callback payloads.
Archive and monitor streams An external archive's ref_dirs pruning and the HV convergence monitor's bootstrap derive their streams from the run seed, unless rng_seed is set. Runs that use them without an explicit seed.

Runs with explicit settings that none of these rows names, and replays of stored runs, give the same results as before.

New options

  • optimize(termination=("max_time", seconds)) stops a run on wall-clock time. Such a run is recorded as not reproducible.
  • A registered problem takes the options its factory declares by name, through optimize(name, problem_kwargs=...), and replay rebuilds the same problem.
  • MOEADConfig.default(variant="de") gives jMetal's MOEA/D-DE and variant="sbx" its MOEA/D with SBX; without variant the default does not change. VAMOS 2.0 will make "de" the default.
  • SMSEMOAConfig.builder().reference_point(rule="jmetal") measures the hypervolume contributions from jMetal's reference point.
  • jMetal's reference fronts of ZDT and DTLZ are packaged as an optional source, vamos --hv-reference-front jmetal; VAMOS's own fronts stay the default. DTLZ1–DTLZ7 also gain two-objective reference fronts.
  • ExternalArchiveConfig.final_selection and final_size report a subset of the external archive when the run ends, as jMetal's and Evolver's results do, while the saved run keeps the whole archive. A saved SPEA2 run whose result does not hold its external archive, as with a final subset or result_mode="population", now stores that archive too; VAMOS 1.2.1 cannot read such a run.
  • Every algorithm with an external archive takes live_callback_mode, and population_archive sends the archive to a live view, which ends with what the run returns.
  • select_top_k_farthest(..., normalize=True) makes the farthest-point subset independent of the objectives' units.
  • MOEA/D's resolved specification records where its weight vectors come from, and vamos tune records the population a MOEA/D trial ran. Its results do not change, but a MOEA/D run stored by an earlier version, which lacks that record, is no longer verified as exact.

Studies

  • A Ctrl+C or a Studio Stop no longer leaves a study that cannot be loaded.
  • load_study(path, verify="metadata") inspects a study whose runs are damaged, and run_limits= and larger StudyLoadLimits apply to every later reload.
  • After Ctrl+C, the CLI no longer suggests resuming a cancelled study.

VAMOS Studio

vamos studio requires the access token of its launch: it prints a link with the token, and the browser keeps a cookie. With a non-loopback address, Studio no longer offers the pages that run code, unless --allow-remote-code-execution is given.

Deprecations

Stable configuration fields that have no effect emit a FutureWarning when they are set and are planned for removal in VAMOS 2.0; the changelog lists them.

VAMOS 1.4.0 removes three Experimental features that depend on LLM providers: VAMOS Assist with the vamos assist command, Studio's Problem Builder, and the openai, anthropic and gemini extras. Define a problem with make_problem or vamos create-problem, and run it from Python, the command line or Studio.

Python support

VAMOS 1.3.0 requires Python 3.11 or newer and is tested on Python 3.11 to 3.14. Python 3.10 reached its end of life in October 2026; VAMOS 1.2.1 is the last release that installs on it. The declared minimum of NumPy rises to 1.23.2, the first NumPy release for Python 3.11, and VAMOS no longer depends on typing-extensions, which only Python 3.10 needed.

Known limitations

In the mixed-encoding tuning spaces of MOEA/D, NSGA-III, SMS-EMOA and SPEA2, the per-type segment settings do not reach the run, so tuning them has no effect. The next minor release fixes them with the grammar-based spaces. See known limitations for the rest.