Skip to content

DESDEO 2.8.0

Choose a tag to compare

@gialmisi gialmisi released this 13 Oct 07:43
· 56 commits to master since this release

[2.8.0] - 13.10.2026

Highlights

This release fixes two defects that users should be aware of. The infix parser
grouped chains of mixed operators wrongly, so a / b * c was parsed as
a / (b * c). This affects every function given as a string, so problems
defined with such expressions may evaluate differently after upgrading. In the
Web API, a decision maker could use another user's interactive session or
parent state by passing its id; this is now refused.

Core logic

Added

  • Added DifferentialEvolutionCrossover (DE/rand/1/bin, Storn and Price 1997)
    and ParentCentricCrossover (PCX, Deb, Anand and Joshi 2002), with
    DifferentialEvolutionCrossoverOptions and ParentCentricCrossoverOptions,
    and the SCALING_FACTOR, SIGMA_ZETA and SIGMA_ETA crossover message
    topics. Both recombine more than two parents and take their step size from the
    population.
  • Added UniformCrossover and UniformCrossoverOptions, a discrete crossover
    for continuous problems in which every decision variable is inherited whole
    from one parent or the other, with pair_xover_probability and
    uniform_xover_probability parameters.
  • Added swap_uncrossed_variables to simulated binary crossover, which
    exchanges the variables not selected for crossover between the two offspring,
    to reproduce jMetal (Java). Defaults to False.
  • Added uniform_xover_probability to the bounded exponential crossover, the
    same uniform-crossover component simulated binary crossover already has.
    Defaults to 0.0, which reproduces the previous offspring exactly.
  • Added an include_perturbed keyword to rpm_solve_solutions. When False,
    only the solution for the given reference point is computed and returned,
    without the k solutions for the perturbed reference points. Defaults to
    True.
  • Added the D-PHI and PHI+ unary indicators to desdeo/tools/indicators_unary.py
    (d_phi_indicator, d_phi_batch, phi_plus_indicator, phi_plus_batch,
    desirability_values, and the DPhiIndicators and PhiPlusIndicator
    result models) (#595).
  • Added desdeo/gdm/grp_subproblem.py for aggregating the reference points of a
    group fairly in NAUTILI: build_grp_subproblem assembles the Group Reference
    Point subproblem from a local preference model
    (additive_preference_constraints or
    symmetric_cones_preference_constraints) and a max-min fairness model
    (maxmin_fairness_constraints, maxmin_fairness_objective). All are exported
    from desdeo.gdm (#572).
  • Added a cluster_extremes field to SCOREBandsResult, which reports the
    minimum and maximum value attained by the solutions of each cluster, as
    opposed to the limits of the band drawn from the cluster.
  • Added the test problems re31 (two bar truss), re32 (welded beam), re33
    (disc brake) and re42 (conceptual marine design), and exported them, along
    with re34, re37, re41 and re61, from desdeo.problem.testproblems.
  • Added the General Aviation Aircraft product family design problem, gaa
    (gaa_problem.py): ten objectives, 27 continuous variables and a single
    aggregated constraint, with coefficients checked bit for bit against the MOEA
    Framework implementation.
  • Added the microalloyed steel design problem of Saini et al. (2023),
    metallurgical_application (metallurgical_application_problem.py). Four
    objectives come from surrogate models that are trained on first use and
    cached; the datasets are read from the repository or downloaded from GitHub.
  • Added forest_problem_income_carbon_habitat (forest_problem.py), the
    discrete forest management problem with income, carbon storage and habitat
    suitability objectives used in R-XIMO and INFRINGER, with its dataset
    datasets/forest_income_carbon_habitat.csv. non_dominated_only=True keeps
    only the 1247 non-dominated rows of the 1728.

Changed

  • Subscribers now register themselves with their publisher once they have been
    constructed, through a metaclass that calls the new Subscriber.__post_init__.
    The EA factories in EAs.py and emo_constructor no longer subscribe the
    components by hand. Publisher.subscribe and Publisher.register_topics are
    now idempotent, so existing code that still subscribes explicitly does not
    receive each message twice.
  • RVEASelector now uses the provisional ideal point, when one is given, until
    the population has a feasible solution, and only then builds the ideal point
    from feasible solutions.
  • RVEASelector now adapts its reference vectors periodically under the
    function evaluation based strategy, instead of only once.
  • CompositeTerminator now forwards messages to its children and exposes their
    topics, so that a limit inside a composite counts. Children the publisher
    already delivers to are skipped.
  • Renamed MaxTimeTerminatorOptions.max_time to max_time_in_seconds, to match
    the MaxTimeTerminator constructor.
  • dmitry_forest_problem_disc now removes duplicate objective vectors, and its
    objectives carry full names and units.
  • Documented that r2_indicator returns the utility form of R2 (higher is
    better, always negative), the negation of the distance form reported by
    PlatEMO and jMetal, and pinned the relationship with tests.
  • The MCWB problem module docstring now names its source, Kenny, Ray and Singh
    (EMO 2025), and the authors' reference implementation.

Fixed

  • Fixed the infix parser grouping mixed operator chains to the right: when the
    operator changed mid-chain, the rest of the chain became the right operand,
    so a / b * c parsed as a / (b * c). Chains are now folded from the left.
    Chains of a single operator and parenthesized expressions were not affected.
  • Fixed the six MCWB welded beam problems (mcwb_*_problem): the parser defect
    inverted their polar moment, which silenced the shear-stress constraint and
    made the tapered deflection NaN, and the formulations had drifted from the
    authors' code (extra and missing constraints, a lost factor of twelve in the
    tapered inertia, a wrong channel bound, mixed units in the Ragsdell variant).
    Every objective and constraint now agrees with the reference implementation
    to machine precision.
  • Fixed RVEASelector crashing on a generation with no feasible member, since
    the ideal point was updated from an empty selection.
  • Fixed the decomposition-based selectors raising TypeError on construction
    when ReferenceVectorOptions had non_preferred_solutions set: the
    norm_order argument of interactive_adapt_2 was passed as ord.
  • Fixed adaptive IBEA dividing by zero when the population collapses onto one
    objective vector. IBEASelector now falls back to the configured kappa when
    the adaptive scale is zero; runs that completed before are unchanged.
  • Fixed terminator_constructor for composite, external check and maximum time
    terminators, none of which could previously be built from options.
  • Fixed the verbose state messages of the EMO evaluator, generators and
    selectors emitting a polars DeprecationWarning on every message, by using
    DataFrame.hstack instead of pl.concat(how="horizontal").
  • Fixed plot_score using matplotlib.cm.get_cmap, which was removed in recent
    matplotlib versions.

Web API

Added

  • Added include_perturbed to RPMSolveRequest, passed on to
    rpm_solve_solutions, so that the reference point method endpoint can return
    a single solution. Defaults to True.

Fixed

  • Fixed the reference point method router not persisting its states. It now
    creates them with StateDB.create, links them to their parent state, and
    stores the preference in RPMState.

Security

  • Fixed SessionContextGuard accepting other users' interactive sessions and
    parent states. fetch_interactive_session now checks that the session belongs
    to the user, and fetch_parent_state only accepts a state in one of the
    user's own sessions or, for a state outside any session, a state of a problem
    the user may access. Analysts and admins are unaffected. A state the user may
    not access is reported as not found.

Documentation

Added

  • Added a how-to guide on solving problems with the web GUI on a local machine
    (docs/howtoguides/webui_solving_problems.md) (#579).
  • Added a how-to guide on solving the RE problems
    (docs/howtoguides/re_problems.ipynb).
  • Replaced the NAUTILI how-to guide with an executable notebook
    (docs/howtoguides/nautili.ipynb) (#572).

Changed

  • Updated the guide on using uv with conda (#579).
  • The installation guide now clones the repository over HTTPS instead of SSH.
  • Moved the solver check in the OpenShift deployment guide after the API
    deployment has rolled out (#576, #577).
  • Listed D-PHI and PHI+ in the features page, and updated the crossover
    explanation notebook for the new operators.
  • Fixed typos throughout the documentation (#588).
  • Re-executed the cached EA how-to notebooks so that they no longer show the
    polars deprecation warnings.

Tooling, CI, and deployment

Added

  • Added command line options to run_fullstack.py: --host, --port,
    --frontend-port, --no-open, --no-reload and --log-level.

Changed

  • run_fullstack.py now locates npm on PATH, so it also works on Windows,
    and reports when it cannot be found (#579).
  • Reformatted pyproject.toml to two-space indentation and updated the lock
    file.