Skip to content

VAMOS 1.1.0 release notes

Released 2026-10-23.

VAMOS 1.1.0 fixes constraint handling and IBEA's hypervolume indicator, both of which could produce scientifically invalid results, and expands the experimental Studio workflow. The stable 1.x API, omitted-engine NumPy default, Python 3.10 support, and 1.0.0 run and study schemas remain the compatibility baseline.

Constraint corrections

  • cec2009_cf1 and welded_beam now declare the constraints they emit, so registry-created problems enforce them. CF1 uses [0, 1]^n and negates the original CEC2009 constraint to follow the VAMOS G <= 0 convention.
  • IBEA removes infeasible solutions before feasible ones, in decreasing order of violation. Its previous penalty had the wrong sign.
  • NSGA-III uses feasibility-first survival and parent selection. Its ask/tell path rejects missing or misaligned offspring constraints without consuming the pending offspring, so the caller can correct and retry them.
  • Non-dominated result selection respects feasibility across the built-in algorithms. When no feasible solution exists, the least violated candidates remain available for inspection.
  • SPEA2 accepts constraint_mode="none" on constrained problems without the previous archive shape failure. All algorithms reject unsupported constraint modes; only "feasibility" and "none" are implemented.

Scientific impact: affected constrained runs can differ from 1.0.2 even with the same seed. Reassess conclusions based on those runs and rerun affected experiments. Unconstrained regression and replay checks are retained. Exact replay requires a matching material environment; keep the original environment when reproducing historical artifacts.

IBEA indicator corrections

  • With indicator="hypervolume", IBEA used the hypervolume-difference indicator with the opposite orientation and kept dominated solutions; on ZDT1 and DTLZ2 its final populations never reached the reference box. It now uses Zitzler and Künzli's I_HD with their adaptive scaling, as jMetal does, and its final fronts match jMetal 7.9's IBEA on ZDT1, DTLZ2 and DTLZ1.
  • An exponential indicator contribution beyond the float64 range counted as zero and made a far-dominated solution look best. Selection then works in the log domain and recomputes each fitness after every removal, so overflow no longer changes which solutions survive.

Scientific impact: discard hypervolume-based IBEA results from earlier versions and rerun them. Epsilon-based IBEA runs without overflow are unchanged; runs whose objectives differed by more than about 709 * kappa can change.

Tuning conditional parameters

The experimental model-based tuning backends optuna, bohb_optuna, smac3 and bohb now honor the conditions of the tuning space, and random and racing also leave out descendants of inactive parameters. Trials contain only active parameters, Optuna suggests parents first, and SMAC3 and BOHB receive native ConfigSpace conditions. Unsupported conditions, unknown parents and dependency cycles fail before sampling. mutation_eta is active only for polynomial mutations in real-valued operator spaces. Native BOHB no longer fails on an unsupported seed argument, and a single-worker native BOHB search is reproducible for a seed.

Reproducibility: inactive parameters never changed the candidate algorithm, so scores of persisted Optuna studies stay comparable and keep their namespace. Seeded random and racing searches propose the same candidates as 1.0.2, except for MOEA/D, whose mutation_eta moved into the polynomial-mutation block. See conditional parameters.

Studio

  • The launcher serves the Panel application and honors --study-dir.
  • Run Experiment and Problem Builder apply the selected algorithm settings, run in the background, report live progress, and support stopping.
  • Batch Comparison creates persistent StudyManifest studies with planning, progress, stop/resume, retry, and access to stored results.
  • Statistical Comparison adds normalized HV and available IGD+, corrected pairwise tests, effect sizes, ranks, plots, and exports.
  • The problem and algorithm catalogs, environment panel, feasible/objective filters, and run-again workflow support inspecting and repeating experiments.
  • Stop requests target a unique worker token. The matching worker exits itself; Studio does not signal a possibly reused process ID or confirm a stop without the worker's acknowledgment after file cleanup. Interrupted atomic writes do not strand temporary attempt files that prevent resume. Wildcard remote binding accepts the computer's host names and addresses; proxies and aliases have an explicit origin option. Final generations are recorded once in the live history.

Studio remains experimental and intended for trusted local use. Customized stored configurations may require vamos reproduce instead of the run-again form. Stopping a batch loses the in-progress run; resume reconciles it.

Automatic backend selection and dependencies

Explicit engine="auto" prefers loadable, JIT-enabled Numba for all nine built-in algorithms, with NumPy fallback before execution. Omitting engine still selects NumPy. Explicit engine requests and persisted resolved engines never silently substitute. Numba is optional and compilation can dominate short runs.

The reviewed dependency set uses mypy 2.4.0 and preserves the compatible PyMdown/mkdocstrings pair on Python 3.10 while using PyMdown 12.3 on Python 3.11+.

Installation and remaining scope

python -m pip install --upgrade "vamos-optimization==1.1.0"
python -m pip install --upgrade "vamos-optimization[studio,compute]==1.1.0"

The welded-beam variant, UF3 domain, independent algorithm validation, full-source typing debt and multi-worker study coordination remain documented limitations. This release does not claim those research goals are complete. See known limitations, stability and versioning, the 1.0.2 release notes, and release verification.