Feature catalogue

This page is generated from the NumPy feature registry on every documentation build. It records declared semantics and costs; see the Torch guide for the supported tensor profile.

Definition identifiers describe the underlying feature. Objective normalization is a separate preprocessing step, applied before the feature formula. Minimum counts are necessary conditions; rank, variation, and selection requirements can impose further restrictions.

For fitness-distance features the minimum applies to selected observations. Read the fitness-distance guide for selection and estimator conventions.

ela_distr

ela_distr.kurtosis

Type-3 excess sample kurtosis of objective observations.

Property	Value
Definition	flacco-type3-v1
Kind	landscape
Requires	y
Minimum observations	4
Cost tier	sample_only
CPU	O(n)
Memory	O(n) shared
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_distr.kurtosis

Intermediates: y.sum2, y.sum4.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	paired finite observations
y_translation	invariant	finite y with non-zero variance
y_positive_scaling	invariant	finite positive scale and non-zero y variance
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)

ela_distr.skewness

Type-3 sample skewness of objective observations under minimization convention.

Property	Value
Definition	flacco-type3-v1
Kind	landscape
Requires	y
Minimum observations	3
Cost tier	sample_only
CPU	O(n)
Memory	O(n) shared
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_distr.skewness

Intermediates: y.sum2, y.sum3.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	paired finite observations
y_translation	invariant	finite y with non-zero variance
y_positive_scaling	invariant	finite positive scale and non-zero y variance
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)

ela_meta

ela_meta.lin_simple.adj_r2

Adjusted R-squared of an ordinary linear model with an intercept.

Property	Value
Definition	ols-adjusted-r2-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	O(n d^2 + d^3)
Memory	O(n p)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_meta.lin_simple.adj_r2

Intermediates: ela_meta.fit.lin_simple.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant
variable_permutation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale and non-constant objective values
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)
• flacco 1.8, ELA meta-model reference implementation (source)

ela_meta.lin_simple.intercept

Intercept of an ordinary linear model fitted by least squares.

Property	Value
Definition	ols-intercept-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n d^2 + d^3)
Memory	O(n d)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_meta.lin_simple.intercept

Intermediates: ela_meta.fit.lin_simple.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant
variable_permutation	invariant
y_translation	equivariant		The fitted intercept changes by the same additive constant.
y_positive_scaling	equivariant		The fitted intercept changes by the same positive factor.
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)
• flacco 1.8, ELA meta-model reference implementation (source)

ela_meta.lin_w_interact.adj_r2

Adjusted R-squared of a linear model with pairwise variable interactions.

Property	Value
Definition	ols-adjusted-r2-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	O(n d^4 + d^6)
Memory	O(n p)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_meta.lin_w_interact.adj_r2

Intermediates: ela_meta.fit.lin_w_interact.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant
variable_permutation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale and non-constant objective values
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)
• flacco 1.8, ELA meta-model reference implementation (source)

ela_meta.quad_simple.adj_r2

Adjusted R-squared of a model containing linear and squared variable terms.

Property	Value
Definition	ols-adjusted-r2-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	O(n d^2 + d^3)
Memory	O(n p)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_meta.quad_simple.adj_r2

Intermediates: ela_meta.fit.quad_simple.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant
variable_permutation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale and non-constant objective values
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)
• flacco 1.8, ELA meta-model reference implementation (source)

ela_meta.quad_w_interact.adj_r2

Adjusted R-squared of a complete degree-2 polynomial model.

Property	Value
Definition	complete-quadratic-adjusted-r2-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	O(n d^4 + d^6)
Memory	O(n p)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ela_meta.quad_w_interact.adj_r2

Intermediates: ela_meta.fit.quad_w_interact.

Conventions and edge cases

• Unlike flacco and pflacco, this corrected definition does not interact already-squared columns and therefore introduces no cubic or quartic terms.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant
variable_permutation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale and non-constant objective values
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Mersmann et al. (2011), Exploratory Landscape Analysis (source)
• flacco 1.8, ELA meta-model reference implementation (source)

fitness_distance

fitness_distance.distance_mean

Mean distance to the selected reference observation.

Property	Value
Definition	best-fraction-distance-mean-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n + k*d)
Memory	O(n + k*d)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	fitness_distance.distance_mean

Intermediates: fitness_distance.distances.

Conventions and edge cases

• Select round(n * proportion_of_best) best canonical objectives; default 0.1.
• Normalize objectives over the full sample before computing selected statistics.
• Selection and reference ties use the first original row; tied samples may be row-order dependent.
• Euclidean distances use raw X and the best selected observation as reference.
• Covariance divides by k; standard deviations use ddof=1.
• fd_correlation is (k-1)/k times Pearson correlation; zero variance is invalid.
• At least two selected observations are required. k is the selected count.

Declared transformations

Transformation	Behavior	Conditions	Notes
objective_sense_reversal	invariant	negate y and reverse objective sense together

References

• pflacco, calculate_fitness_distance_correlation (source)

fitness_distance.distance_std

Sample standard deviation of reference distances (ddof=1).

Property	Value
Definition	best-fraction-distance-std-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n + k*d)
Memory	O(n + k*d)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	fitness_distance.distance_std

Intermediates: fitness_distance.distances.

Conventions and edge cases

• Select round(n * proportion_of_best) best canonical objectives; default 0.1.
• Normalize objectives over the full sample before computing selected statistics.
• Selection and reference ties use the first original row; tied samples may be row-order dependent.
• Euclidean distances use raw X and the best selected observation as reference.
• Covariance divides by k; standard deviations use ddof=1.
• fd_correlation is (k-1)/k times Pearson correlation; zero variance is invalid.
• At least two selected observations are required. k is the selected count.

Declared transformations

Transformation	Behavior	Conditions	Notes
objective_sense_reversal	invariant	negate y and reverse objective sense together

References

• pflacco, calculate_fitness_distance_correlation (source)

fitness_distance.fd_correlation

Population fitness-distance covariance divided by sample deviations.

Property	Value
Definition	best-fraction-fd-correlation-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n + k*d)
Memory	O(n + k*d)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	fitness_distance.fd_correlation

Intermediates: fitness_distance.distances.

Conventions and edge cases

• Select round(n * proportion_of_best) best canonical objectives; default 0.1.
• Normalize objectives over the full sample before computing selected statistics.
• Selection and reference ties use the first original row; tied samples may be row-order dependent.
• Euclidean distances use raw X and the best selected observation as reference.
• Covariance divides by k; standard deviations use ddof=1.
• fd_correlation is (k-1)/k times Pearson correlation; zero variance is invalid.
• At least two selected observations are required. k is the selected count.

Declared transformations

Transformation	Behavior	Conditions	Notes
objective_sense_reversal	invariant	negate y and reverse objective sense together

References

• pflacco, calculate_fitness_distance_correlation (source)

fitness_distance.fd_cov

Population covariance of selected fitness and reference distance.

Property	Value
Definition	best-fraction-fd-cov-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n + k*d)
Memory	O(n + k*d)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	fitness_distance.fd_cov

Intermediates: fitness_distance.distances.

Conventions and edge cases

• Select round(n * proportion_of_best) best canonical objectives; default 0.1.
• Normalize objectives over the full sample before computing selected statistics.
• Selection and reference ties use the first original row; tied samples may be row-order dependent.
• Euclidean distances use raw X and the best selected observation as reference.
• Covariance divides by k; standard deviations use ddof=1.
• fd_correlation is (k-1)/k times Pearson correlation; zero variance is invalid.
• At least two selected observations are required. k is the selected count.

Declared transformations

Transformation	Behavior	Conditions	Notes
objective_sense_reversal	invariant	negate y and reverse objective sense together

References

• pflacco, calculate_fitness_distance_correlation (source)

fitness_distance.fitness_mean

Mean of the best objective observations.

Property	Value
Definition	best-fraction-fitness-mean-v1
Kind	landscape
Requires	y
Minimum observations	2
Cost tier	sample_only
CPU	O(n)
Memory	O(n)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	fitness_distance.fitness_mean

Intermediates: fitness_distance.selection.

Conventions and edge cases

• Select round(n * proportion_of_best) best canonical objectives; default 0.1.
• Normalize objectives over the full sample before computing selected statistics.
• Selection and reference ties use the first original row; tied samples may be row-order dependent.
• Euclidean distances use raw X and the best selected observation as reference.
• Covariance divides by k; standard deviations use ddof=1.
• fd_correlation is (k-1)/k times Pearson correlation; zero variance is invalid.
• At least two selected observations are required. k is the selected count.

Declared transformations

Transformation	Behavior	Conditions	Notes
objective_sense_reversal	invariant	negate y and reverse objective sense together

References

• pflacco, calculate_fitness_distance_correlation (source)

fitness_distance.fitness_std

Sample standard deviation of the best objective observations (ddof=1).

Property	Value
Definition	best-fraction-sample-std-v1
Kind	landscape
Requires	y
Minimum observations	2
Cost tier	sample_only
CPU	O(n)
Memory	O(n)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	fitness_distance.fitness_std

Intermediates: fitness_distance.selection.

Conventions and edge cases

• Keep round(n * proportion_of_best) smallest canonical objective values; default 0.1.
• Objective normalization uses the full sample before selecting the best fraction.
• Use proportion_of_best=1.0 for the full sample; no optimum or distances are needed.
• Fewer than two selected observations are invalid; a constant selection returns zero.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant
y_translation	invariant
y_positive_scaling	equivariant		Multiplying input objectives by a positive factor scales the output by it.
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• pflacco, calculate_fitness_distance_correlation (source)

ic

ic.eps_max

Epsilon at maximum information entropy.

Property	Value
Definition	deterministic-nn-flacco-grid-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	expected O(n log n + e); worst O(n^2 d)
Memory	O(nd + kn + e), with k=20
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ic.eps_max, ic.eps.max

Intermediates: ic.entropy.

Conventions and edge cases

• Uses a lexicographically anchored nearest-neighbour tour over duplicate-aggregated X.
• Uses the fixed flacco epsilon grid with e=1001.
• Epsilon-indexed curves are accumulated from per-slope threshold events, so the grid-by-slope symbol matrix is never materialised.
• Nearest-neighbour construction respects the explicit compute workers setting; the default is one worker.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	unique lexicographic start and deterministic distance ties
x_translation	invariant	finite translation preserving pairwise distances
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Muñoz, Kirley, and Halgamuge (2015), ELA using information content (source)
• flacco 1.8 information-content implementation (source)

ic.eps_ratio

Log10 half-partial-information sensitivity.

Property	Value
Definition	deterministic-nn-flacco-grid-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	expected O(n log n + e); worst O(n^2 d)
Memory	O(nd + kn + e), with k=20
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ic.eps_ratio, ic.eps.ratio

Intermediates: ic.partial.

Conventions and edge cases

• Uses a lexicographically anchored nearest-neighbour tour over duplicate-aggregated X.
• Uses the fixed flacco epsilon grid with e=1001.
• Epsilon-indexed curves are accumulated from per-slope threshold events, so the grid-by-slope symbol matrix is never materialised.
• Nearest-neighbour construction respects the explicit compute workers setting; the default is one worker.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	unique lexicographic start and deterministic distance ties
x_translation	invariant	finite translation preserving pairwise distances
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Muñoz, Kirley, and Halgamuge (2015), ELA using information content (source)
• flacco 1.8 information-content implementation (source)

ic.eps_s

Log10 settling sensitivity of the fitness sequence.

Property	Value
Definition	deterministic-nn-flacco-grid-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	expected O(n log n + e); worst O(n^2 d)
Memory	O(nd + kn + e), with k=20
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ic.eps_s, ic.eps.s

Intermediates: ic.entropy.

Conventions and edge cases

• Uses a lexicographically anchored nearest-neighbour tour over duplicate-aggregated X.
• Uses the fixed flacco epsilon grid with e=1001.
• Epsilon-indexed curves are accumulated from per-slope threshold events, so the grid-by-slope symbol matrix is never materialised.
• Nearest-neighbour construction respects the explicit compute workers setting; the default is one worker.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	unique lexicographic start and deterministic distance ties
x_translation	invariant	finite translation preserving pairwise distances
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Muñoz, Kirley, and Halgamuge (2015), ELA using information content (source)
• flacco 1.8 information-content implementation (source)

ic.h_max

Maximum information entropy of the fitness sequence.

Property	Value
Definition	deterministic-nn-flacco-grid-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	expected O(n log n + e); worst O(n^2 d)
Memory	O(nd + kn + e), with k=20
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ic.h_max, ic.h.max

Intermediates: ic.entropy.

Conventions and edge cases

• Uses a lexicographically anchored nearest-neighbour tour over duplicate-aggregated X.
• Uses the fixed flacco epsilon grid with e=1001.
• Epsilon-indexed curves are accumulated from per-slope threshold events, so the grid-by-slope symbol matrix is never materialised.
• Nearest-neighbour construction respects the explicit compute workers setting; the default is one worker.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	unique lexicographic start and deterministic distance ties
x_translation	invariant	finite translation preserving pairwise distances
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Muñoz, Kirley, and Halgamuge (2015), ELA using information content (source)
• flacco 1.8 information-content implementation (source)

ic.m0

Initial partial information content at epsilon zero.

Property	Value
Definition	deterministic-nn-flacco-grid-v1
Kind	landscape
Requires	x, y
Minimum observations	3
Cost tier	sample_only
CPU	expected O(n log n + e); worst O(n^2 d)
Memory	O(nd + kn + e), with k=20
Additional objective evaluations	0
Deterministic	Yes
Legacy names	ic.m0

Intermediates: ic.slopes.

Conventions and edge cases

• Uses a lexicographically anchored nearest-neighbour tour over duplicate-aggregated X.
• Uses the fixed flacco epsilon grid with e=1001.
• Epsilon-indexed curves are accumulated from per-slope threshold events, so the grid-by-slope symbol matrix is never materialised.
• Nearest-neighbour construction respects the explicit compute workers setting; the default is one worker.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	unique lexicographic start and deterministic distance ties
x_translation	invariant	finite translation preserving pairwise distances
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Muñoz, Kirley, and Halgamuge (2015), ELA using information content (source)
• flacco 1.8 information-content implementation (source)

nbc

nbc.dist_ratio.coeff_var

Coefficient of variation of nearest/nearest-better distance ratios.

Property	Value
Definition	exact-euclidean-blockwise-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n^2 d)
Memory	O(n b)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	nbc.dist_ratio.coeff_var

Intermediates: nbc.graph.

Conventions and edge cases

• Exact Euclidean search with bounded distance-block memory.
• Strictly better candidates are preferred; equal-fitness candidates are the fallback.
• Exact distance ties choose the lowest current observation index.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	no exact equal-distance candidate ties
variable_permutation	invariant
x_translation	invariant
x_orthogonal_rotation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Kerschke et al. (2015), Detecting funnel structures by means of ELA (source)
• flacco 1.8 nearest-better implementation (source)

nbc.nb_fitness.cor

Correlation of nearest-better indegree and canonical fitness.

Property	Value
Definition	exact-euclidean-blockwise-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n^2 d)
Memory	O(n b)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	nbc.nb_fitness.cor

Intermediates: nbc.graph.

Conventions and edge cases

• Exact Euclidean search with bounded distance-block memory.
• Strictly better candidates are preferred; equal-fitness candidates are the fallback.
• Exact distance ties choose the lowest current observation index.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	no exact equal-distance candidate ties
variable_permutation	invariant
x_translation	invariant
x_orthogonal_rotation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Kerschke et al. (2015), Detecting funnel structures by means of ELA (source)
• flacco 1.8 nearest-better implementation (source)

nbc.nn_nb.cor

Correlation of nearest and nearest-better distances.

Property	Value
Definition	exact-euclidean-blockwise-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n^2 d)
Memory	O(n b)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	nbc.nn_nb.cor

Intermediates: nbc.graph.

Conventions and edge cases

• Exact Euclidean search with bounded distance-block memory.
• Strictly better candidates are preferred; equal-fitness candidates are the fallback.
• Exact distance ties choose the lowest current observation index.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	no exact equal-distance candidate ties
variable_permutation	invariant
x_translation	invariant
x_orthogonal_rotation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Kerschke et al. (2015), Detecting funnel structures by means of ELA (source)
• flacco 1.8 nearest-better implementation (source)

nbc.nn_nb.mean_ratio

Ratio of nearest and nearest-better mean distances.

Property	Value
Definition	exact-euclidean-blockwise-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n^2 d)
Memory	O(n b)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	nbc.nn_nb.mean_ratio

Intermediates: nbc.graph.

Conventions and edge cases

• Exact Euclidean search with bounded distance-block memory.
• Strictly better candidates are preferred; equal-fitness candidates are the fallback.
• Exact distance ties choose the lowest current observation index.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	no exact equal-distance candidate ties
variable_permutation	invariant
x_translation	invariant
x_orthogonal_rotation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Kerschke et al. (2015), Detecting funnel structures by means of ELA (source)
• flacco 1.8 nearest-better implementation (source)

nbc.nn_nb.sd_ratio

Ratio of nearest and nearest-better sample deviations.

Property	Value
Definition	exact-euclidean-blockwise-v1
Kind	landscape
Requires	x, y
Minimum observations	2
Cost tier	sample_only
CPU	O(n^2 d)
Memory	O(n b)
Additional objective evaluations	0
Deterministic	Yes
Legacy names	nbc.nn_nb.sd_ratio

Intermediates: nbc.graph.

Conventions and edge cases

• Exact Euclidean search with bounded distance-block memory.
• Strictly better candidates are preferred; equal-fitness candidates are the fallback.
• Exact distance ties choose the lowest current observation index.

Declared transformations

Transformation	Behavior	Conditions	Notes
row_permutation	invariant	no exact equal-distance candidate ties
variable_permutation	invariant
x_translation	invariant
x_orthogonal_rotation	invariant
y_translation	invariant
y_positive_scaling	invariant	finite positive scale
objective_sense_reversal	invariant	negate y and reverse the declared objective sense together

References

• Kerschke et al. (2015), Detecting funnel structures by means of ELA (source)
• flacco 1.8 nearest-better implementation (source)