What’s new#
Version 0.4.0#
Requires
numpy>=2.1.vorob_t()returns aNamedTupleinstead of a dictionary.is_nondominated()is up to 10x faster in some inputs thanks to a customized radixsort implementation.New shapes
"concave-cliff"and"convex-cliff"added togenerate_ndset().New function
generate_sequence()to generate sequences of dominated and nondominated points.pygmo is now included in the Nondominated Sorting (Pareto ranking) and Exact computation of hypervolume benchmarks.
New unbounded nondominated archive
UnboundedArchive().New archiving benchmarks Archiving Benchmarks.
Version 0.3.2 (11/07/2026)#
Fix crash in
pareto_rank()with duplicated points (#74).The maximum number of dimensions supported by most functions is now raised to 255, except for hypervolume and hypervolume approximation, where it remains 31. Functions
igd(),igd_plus(),avg_hausdorff_dist(),epsilon_additive()andepsilon_mult()use a slower implementation when reaching that limit. Other functions now give a clear error to the user when the limits are reached (#72).
Version 0.3.1 (04/05/2026)#
Fix potential undefined behavior in
is_nondominated().
Version 0.3.0 (03/05/2026)#
moarchiving, paretobench and patatune are now included in the benchmarks.
r2_exact()implements the exact computation of the R2 indicator for bi-objective solution sets.hypervolume()is significantly faster for more than four dimensions (Andreia P. Guerreiro).hypervolume()now handles 1D inputs and provides a clear error for 0D inputs (#58).is_nondominated()andfilter_dominated()implement the best-known \(O(n(\log_2 n)^{m-2})\) algorithm, thus they are significantly faster for 4D and above. See Identifying nondominated points benchmarks.pareto_rank()is also significantly faster for for 4D and above. See Nondominated Sorting (Pareto ranking) benchmarks.New shapes
"inverted-simplex"and"concave-simplex"added togenerate_ndset(). Shape"convex-simplex"is now equivalent togenerate_ndset(..., method="simplex") ** 2, which is slightly more uniform than the previous approach.The default method of
hv_approx()is now"Rphi-FWE+", which is typically as accurate as the other methods, but significantly faster.
Version 0.2.0 (10/01/2026)#
Requires
numpy>=1.24.Optuna is now included in the benchmarks.
New Nondominated Sorting (Pareto ranking) benchmarks.
pareto_rank()is faster in 3D.HV3D+ is slightly faster with repeated coordinates (#41).
igd(),igd_plus(),avg_hausdorff_dist()are faster.Fix wrong values returned by
epsilon_mult()when mixing minimization and maximization.Fix documentation of Epsilon metric (@leandrolanzieri)
hypervolume()uses the inclusion-exclusion algorithm for small inputs of up to 15 points, which is significantly faster.
Version 0.1.10 (24/11/2025)#
is_nondominated(): Fix wrong assert (#38).
Version 0.1.9 (31/10/2025)#
hv_contributions()ignores dominated points by default. Setignore_dominated=Falseto restore the previous behavior. The 3D case uses the HVC3D algorithm.New function
any_dominated()to quickly detect if a set is nondominated.New function
generate_ndset()to generate random nondominated sets with different shapes.New example Sampling Random Nondominated Sets.
is_nondominated(),any_dominated(),pareto_rank()now handle single-objective inputs correctly (#27, #29).Ranks returned by
pareto_rank()are 0-based.is_nondominated()andfilter_dominated()are faster for dimensions larger than 3.moocorewheels are now built foraarch64(ARM64) in Linux and Windows. See the installation instructions.is_nondominated()andfilter_dominated()are now stable in 2D and 3D with!keep_weakly, that is, only the first of duplicated points is marked as nondominated.
Version 0.1.8 (15/07/2025)#
Correct license to LGPL v2.1 or later.
Bump dependencies to
cffi>=1.17.1andsetuptools>=77.0.3.eaf(),vorob_t()andvorob_dev()take the set indices as a separate argumentsetsfollowing the API of the R package.New example Empirical Attainment Function.
Document EAF and Vorob’ev expectation and deviation in more detail.
New online dataset:
DTLZLinearShape.8d.front.60pts.10(seeget_dataset()).New default method in
hv_approx(). Computation is now done in C, so it is much faster.hv_contributions()is much faster for 2D inputs.
Version 0.1.7 (04/06/2025)#
hypervolume()now uses the HV3D+ algorithm for the 3D case and the HV4D+ algorithm for the 4D case. For dimensions larger than 4, the recursive algorithm uses HV4D+ as the base case, which is significantly faster.read_datasets()is significantly faster for large files.is_nondominated()andfilter_dominated()are faster for 3D inputs.New function:
hv_contributions().New online datasets:
test2D-200k.inp.xzandran.1000pts.3d.10(seeget_dataset()).
Version 0.1.6 (14/05/2025)#
New function:
largest_eafdiff().New class:
RelativeHypervolume.New dataset
tpls50x20_1_MWT.csv.Extended example Computing Multi-Objective Quality Metrics.
vorobT()andvorobDev()were renamed tovorob_t()andvorob_dev()to follow Python convention.get_dataset_path()andget_dataset()can download large datasets from a remote repository.
Version 0.1.4 (30/10/2024)#
Improved example Using moocore with Pandas to work in Pandas version >= 2.
Changed behavior of
apply_within_sets(). The previous behavior could lead to subtle bugs.
Version 0.1.3 (28/10/2024)#
New:
Hypervolume: Object-oriented API for hypervolume indicator.New:
apply_within_sets(): Utility function to apply operations to individual datasets.New:
is_nondominated_within_sets(): Utility function to identify nondominated points within sets.New example using
pandas.DataFramein Using moocore with Pandas.Fix bug in
normalise()when the input ispandas.DataFrameor some other non-contiguous array.
Version 0.1.2 (18/09/2024)#
New:
hv_approx()Documentation improvements.
New gallery examples.