Class UNDXCrossover

Namespace

OptunaSharp.Samplers.Nsgaii

Assembly

OptunaSharp.dll

Unimodal Normal Distribution Crossover used by NSGAIISampler. Port of optuna/samplers/nsgaii/_crossovers/_undx.py: UNDXCrossover (experimental v3.0.0).

Generates a child from three parents using a multivariate normal distribution.

QR sign parity (n > 1). The orthonormal basis orthogonal to the primary search line comes from np.linalg.qr (LAPACK dgeqrf/dorgqr). QR is unique only up to the signs of R's diagonal, and those signs propagate into the basis vectors: get them wrong and, with two or more numerical parameters, the child values differ from Optuna by O(1), which then cascades through the crowding-distance/non-domination selection and diverges the whole study trajectory. So the basis is computed through QrQ(double[,]) (OpenBLAS/LAPACK — column signs byte-identical to NumPy) whenever the native backend is present, which is the normal case and gives full Optuna parity here (to the ~1e-14 managed-vs-LAPACK tolerance the downstream selection is insensitive to).

[Misaligned] no-native (Tier N) fallback. Without OpenBLAS the basis falls back to a managed Householder QR (HouseholderQ(double[,], int)) that matches numpy's magnitudes to ~1e-14 but not its per-column signs (its accumulated Q = Q·Hₖ flips signs vs orgqr). UNDX still samples correctly — etas ~ N(0, σ²) is symmetric, so a sign-flipped basis is a valid draw and the study stays in RNG lock-step within OptunaSharp — but its trajectory can sign-diverge from Optuna for ≥2-numerical-parameter spaces. A single numerical parameter is bit-exact on either path (the basis is unused).

public class UNDXCrossover : BaseCrossover

Inheritance

object

BaseCrossover

UNDXCrossover

Inherited Members

BaseCrossover.ToString()

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

Constructors

UNDXCrossover(double, double?)

Configures normal offspring variation along and perpendicular to the primary parent axis.

public UNDXCrossover(double sigmaXi = 0.5, double? sigmaEta = null)

Parameters

sigmaXi double

Parametrizes the normal distribution from which xi is drawn.

sigmaEta double?

Parametrizes the normal distribution from which etas are drawn. Defaults to 0.35 / sqrt(len(search_space)).

Properties

NParents

Number of parent individuals required to perform crossover.

public override int NParents { get; }

Property Value

int

Methods

Crossover(double[][], NumPyRandom, Study, double[,])

Perform crossover of selected parent individuals.

public override double[] Crossover(double[][] parentsParams, NumPyRandom rng, Study study, double[,] searchSpaceBounds)

Parameters

parentsParams double[][]

A num_parents x num_parameters array. Represents a parameter space for each parent individual (continuous for numerical parameters). Row i is parent i.

rng NumPyRandom

An instance of numpy.random.RandomState.

study Study

Target study object.

searchSpaceBounds double[,]

A len_search_space x 2 array of numerical distribution bounds constructed from the transformed search space. Column 0 is lower, column 1 is upper.

Returns

double[]

A 1-dimensional array containing the new parameter combination.