Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

MMD-Newton Method for Multi-objective Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proposes minimizing the maximum mean discrepancy between a Pareto approximation set and a reference set, and shows that a Newton method on that objective both pulls the set onto the front and keeps it spread, beating three…

desk verdict MMD-Newton is a real but modest contribution: the analytical derivatives are correct and the hybrid idea is sensible, but the reference-set robustness story is oversold and Lemma 1 has a gap. read the letter →

arxiv 2505.14610 v1 pith:KRG2XNUJ submitted 2025-05-20 cs.LG

Hao Wang , Chenyu Shi , Angel E. Rodriguez-Fernandez , Oliver Schütze This is my paper · ORCID
classification cs.LG MSC 90C2990C53
keywords multi-objectiveoptimizationmaximummeandiscrepancyNewtonmethodParetofrontapproximationkernelembeddingevolutionaryalgorithmreferencesetHessianeigenspectrum
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that maximum mean discrepancy (MMD) is a suitable distance between a Pareto approximation set and a reference set for continuous multi-objective optimization, and that minimizing it with a second-order method refines fronts faster than evolutionary search alone. Treating the two finite sets as empirical measures, the authors derive the gradient and Hessian of $\mathrm{MMD}^2$ with respect to the decision variables and build a Newton method, called MMDN, with theoretical conditions on stationarity and on the Hessian's eigenvalues. Since squared MMD contains the RKHS norm of the approximation set's embedding, minimizing it also spreads the set, so refinement works even when the reference does not cover the whole front. Hybridized with an MOEA, which supplies a warm start and a reference, MMDN produced better average Hausdorff distances than the MOEA alone under the same evaluation budget in 28 of 33 benchmark configurations.

What carries the argument

The central object is squared maximum mean discrepancy between two empirical measures, $$\mathrm{MMD}^2(Y,R)=\|\psi(Y)-\psi(R)\|_{\mathcal H}^2=\frac{1}{\$mu^{2}$}\sum_{i,j}k(y_i,y_j)+\frac{1}{\$lambda^{2}$}\sum_{i,j}k(r_i,r_j)-\frac{2}{\mu\$\lambda$}\sum_{i,j}k(r_i,y_j),$$ where $\psi$ is the kernel mean embedding into an RKHS with a Gaussian or Matérn kernel. The carrying mechanism is the Newton step for the KKT system of minimizing $\mathrm{MMD}^2$ over stacked decision variables $X\in\mathbb{R}^{\mu n}$ subject to constraints, using the derived gradient and Hessian and an active-set treatment of inequalities. Because the RKHS norm $\|\psi(Y)\|_{\mathcal H}^2$ grows when points move apart under a decreasing kernel, the method spreads $Y$ while pulling it toward $R$; because the Hessian is not guaranteed positive definite, the algorithm preconditions it and uses a backtracking line search.

What would settle it

Take a two- or three-objective problem whose Pareto front has two well-separated components; run the warm-start MOEA for 300 iterations and stop it before any population member enters the second component; build the reference set from that population and run MMDN for five iterations; if the final approximation still has no points on the second component, the claimed ability to recover from an imperfect reference is contradicted in that setting.

Watch

Extended reading notes

Core claim

On the paper's own terms, the core discovery is that a Newton method on $\mathrm{MMD}^2(Y,R)$ can refine a Pareto approximation set $Y=F[X]$ against a reference set $R$, and the diversity term hidden inside MMD is what makes the refinement robust to an imperfect $R$. The authors provide closed-form first and second derivatives of $\mathrm{MMD}^2$ with respect to each decision point, characterize the first-order stationary condition in the frequency domain, and bound the Hessian spectrum, showing it can be indefinite and therefore must be preconditioned. Empirically, warm-starting MMDN with the output of NSGA-II, NSGA-III, or MOEA/D and running five Newton steps improves the approximation accuracy compared with running the MOEA alone under the same function-evaluation budget, with one clear failure on DTLZ7 when the warm-start population is concentrated on a subregion.

Load-bearing premise

The method's success depends on the reference set built from the MOEA's own starting population; if that population covers only a piece of the Pareto front, the reference inherits the gap and the refinement cannot reliably recover the missing piece.

Editorial extensions

If this is right

  • MMDN can be used as a cheap post-processor: five Newton steps after a 300-iteration MOEA run improve the Pareto approximation under the same total evaluation budget.
  • The diversity term in MMD means that a reference set that does not cover the full front does not necessarily cap the final coverage; the method can spread beyond it.
  • The Hessian spectral analysis explains why preconditioning is necessary and identifies the spread-dependent terms that control the spectrum.
  • The hybrid's accuracy is not uniform: on problems where the warm-start population is concentrated, such as DTLZ7 with MOEA/D, the refinement can be worse than the MOEA alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the diversity mechanism is as general as the Figure 1 example suggests, the same MMD minimization could serve as a post-processing step for any set-generating optimizer, including Bayesian or gradient-based multi-objective methods, not just MOEAs.
  • The stationary-condition analysis predicts that the converged location of each point is tied to the kernel length-scale and to the reference set's center of mass, so choosing $\theta$ tunes a trade-off between fitting the reference and maximizing spread.
  • A natural stress test is disconnected Pareto fronts: MMD should preserve coverage only if the reference or the initial set samples every component, and the paper's failure case on DTLZ7 suggests the recovery claim has a coverage threshold.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes using maximum mean discrepancy (MMD) between a finite Pareto approximation set Y=F[X] and a reference set R as an objective for numerical multi-objective optimization. It derives analytical gradient and Hessian expressions for MMD with respect to the decision variables, formulates a set-oriented Newton method (MMDN), and analyzes a first-order stationary condition and Hessian eigenvalue bounds. The authors then hybridize MMDN with MOEAs by running an MOEA for 300 iterations, constructing R from the MOEA's final population Y0 via interpolation, k-means clustering, and a utopian shift, and running five MMDN iterations. They benchmark the hybrid against the same MOEA alone under an equivalent function-evaluation budget on 11 problems with three MOEAs, reporting 28 wins, 4 ties, and 1 loss. The central claimed advantage is that MMD's self-term maintains diversity, so the method can improve coverage even when R is imperfect.

Significance. If the central claims hold, the paper would make a useful contribution: it provides explicit derivative formulas for a set-based MMD objective, a Newton-type refinement strategy, and a practical hybrid that appears to improve several MOEA baselines. The analytical gradient/Hessian derivations in Section 4.1 are standard but useful, and the reported code link supports reproducibility. The empirical comparison is extensive for a paper of this length, and the win/tie/loss pattern across 33 configurations is concrete evidence that the hybrid often helps. However, the main theoretical lemma on stationary conditions is stated as an equivalence that is not actually proven, and the core robustness claim about 'imperfect reference sets' is undermined by the fact that R is constructed from the very population being refined. These issues are load-bearing for the paper's stated contributions, so the manuscript needs substantial revision before the claims can be accepted as stated.

major comments (4)
  1. [Section 4.2, Eq. (24), and Appendix A.1] Lemma 1's 'if and only if' direction is not established. From the stationarity equation ∂/∂y_l MMD²(Y,R) = (2/µ²)∫_{R^k} ω B(ω) G(dω) = 0, with B(ω)=Σ_{α≠l} sin⟨ω,y_α-y_l⟩ − Σ_β sin⟨ω,r_β-y_l⟩, one cannot conclude that B(ω)=0 for every ω. A nondegenerate, or even strictly positive, spectral measure G does not prevent cancellations: an odd integrand can integrate to zero over a symmetric measure without vanishing pointwise. Only the 'if' direction (pointwise identity implies zero gradient) follows directly. The lemma, and Remark 1 built on it, should be weakened to a sufficient condition or augmented with an additional assumption that rules out such cancellations.
  2. [Algorithm 1 lines 5-9, Section B, Section 7, and Table 1] The central claim that MMD 'can improve the Pareto approximation set even when the reference R is imperfect' is not supported by the experimental protocol, because R is generated from the MOEA's own final population Y0: Y0 is interpolated, reduced by k-means, and shifted into the utopian region. If Y0 is confined to a subregion of the Pareto front, R inherits that span; the cross-term −2⟨ψ(Y),ψ(R)⟩ attracts points toward R's support, and the self-term repels points but does not direct them toward the missing part of the front. The paper's only loss in Table 1, MOEA/D on DTLZ7, is precisely this failure mode, and Section 7 concedes that the reference set and starting points are of low quality. To substantiate the diversity-recovery claim, the authors should test MMDN with a reference set independent of the initial population, or artificially restrict Y0 to a subregion and show that MMDN recovers full coverage. Without such an ablation, the observed wins are consistent with ordinary local Newton refinement of a good warm start rather than with the claimed diversity mechanism.
  3. [Section 4.3, Corollary 1, Eq. (17), and Algorithm 1 line 12] The eigenspectrum analysis applies to the objective-space block Hessian ∂²/∂y_m∂y_l MMD²(Y,R), but the Newton step and the preconditioning step in Algorithm 1 operate on the decision-space Hessian ∇² MMD²(X,R) defined in Eq. (17). The decision-space Hessian contains additional terms involving DF(x_m) and D²F(x_l), so its definiteness is not governed directly by the bounds in Corollary 1. The statement that preconditioning is necessary because 'the MMD Hessian is indefinite' therefore concerns a different matrix than the one being preconditioned. The authors should either prove corresponding bounds for the decision-space Hessian, or state explicitly that Corollary 1 concerns the objective-space MMD landscape and explain how that justifies the preconditioning of Eq. (17).
  4. [Section 6, 'Fair comparison of MMDN to MOEA' and Section C.3] The claim of comparison 'with the same computation budget' is only approximate. The conversion from five MMDN iterations to an equivalent number of function evaluations uses measured CPU-time ratios for Jacobian and Hessian calls, but the reference set generation (DBSCAN, Delaunay triangulation, k-means, shifting) and the Hessian preconditioning/line-search overhead are not included in the budget. If these costs are non-negligible, the comparison is not strictly fair. At minimum, the wall-clock overhead of these steps should be reported, and the budget-conversion procedure should account for them or justify that they are negligible.
minor comments (5)
  1. [Algorithm 1, line 11] The while-loop condition reads 'while ||∇MMDM²(X_k)||₂ ≤ ε ∧ k < N2', which would execute the loop when the gradient norm is already below the tolerance; this should presumably be '> ε' or '≥ ε' if the intent is to stop once the gradient is small.
  2. [Eq. (17)] The Hessian formula mixes row-vector and tensor notation: the term (∂/∂y_l MMD²) D²F(x_l) δ_m^l should specify which index of the (1,2)-tensor is contracted, and the dimensions of the resulting block should be stated explicitly.
  3. [Section 6 and Table 1] The averaged Hausdorff metric ∆_p used for performance assessment is not defined in the main text; since the paper cites [29] in the introduction, adding the definition or a one-line formula would make the results self-contained.
  4. [Section C.2] The heuristic that selects the kernel and length-scale by minimizing the condition number of the MMD Hessian is evaluated on the same problem instances used for the final comparison; it would be helpful to state whether this selection is performed on a separate validation run or is part of the reported algorithm, to avoid any appearance of tuning on the test set.
  5. [Section A.2 and Example 2] In Theorem 1, the definition of λ as a ratio of integrals requires the denominator integral to be nonzero; the proof should discuss when this holds and what happens in the degenerate case, especially since Example 2 then approximates the ratio by m2/m1.

Circularity Check

1 steps flagged · score 4.0 of 10

The MMD gradient/Hessian derivation is self-contained, but the headline robustness claim that MMD recovers coverage from an imperfect reference is partially self-referential because the reference set R is always constructed from the very warm-start population Y0 that MMDN is supposed to refine.

  1. self definitional [Algorithm 1 lines 4-9; Section 5; Section B; Section 7]
    "//execute MOEA for N1 iteration to get the initial approximate set X0 for MMDN ... X0 <- A(F, h, g, mu, N1); //generate reference set R. see Section B for detail. Y0 <- F(X0); interpolate Y0; R <- use k-means clustering to select mu points from Y0; Compute the unit normal vector eta to the convex hull Conv(Y0); for r in R do r <- r+delta eta"

    The reference set R is a filled, k-means-reduced, utopian-shifted copy of the initial approximation: R = shift_eta(kmeans(interpolate(Y0))). Algorithm 1 then minimizes MMD^2(F[X], R), so the target against which discrepancy is measured is a smoothed, shifted version of the very set being refined. The advertised advantage, that MMD improves the Pareto approximation set even when the reference R is imperfect, is therefore only exercised in the regime where R's imperfection is inherited from Y0. If Y0 misses a front segment, R misses it too; the diversity self-term can spread points but supplies no information about the missing segment.

full rationale

The core mathematical contribution, the analytical gradient and Hessian of MMD with respect to the decision variables (Section 4), the first-order stationarity characterization (Lemma 1), the Hessian eigenspectrum bounds (Theorem 2, Corollary 1), and the Newton system (Eq. 10), is derived from the definition of MMD and does not reduce to its inputs. The empirical comparison with MOEAs is budget-matched and evaluated against known Pareto fronts, so the headline win/loss statistics are not circular by construction. However, the paper's central conceptual claim that MMD's diversity term rescues imperfect reference sets is weakened by the self-referential reference construction: every R used in the hybrid is a transformed copy of the initial MOEA population Y0. Consequently, the method is never tested against an independent reference that could reveal front regions absent from Y0; the one reported loss (MOEA/D on DTLZ7) is precisely the predicted failure mode. This warrants a moderate partial-circularity score, but the Newton derivation itself remains independent, so a score of 4 is appropriate rather than 6 or higher.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The method adds two main free parameters (kernel choice and length-scale θ, selected per problem) and relies on the heuristic that a reference set derived from the initial population is good enough. Several standard mathematical assumptions (C^2 functions, Bochner representation, finite spectral moments) underpin the derivation. No new physical entities are introduced.

free parameters (3)
  • Kernel type and length-scale θ = Table 2: e.g., Gaussian θ=2000 for ZDT1, Laplace θ=500 for DTLZ2, varying per problem and baseline MOEA
    Selected per benchmark problem via grid search over {Gaussian, Laplace} × {10^-2,...,5000} to minimize the condition number of the preconditioned Hessian (Section C.2). This per-problem tuning affects the empirical results and is a free parameter of the method.
  • Shift distance δ = 0.08 (Algorithm 1, line 8)
    Controls how far the reference set is shifted into the utopian region. A fixed hand-chosen constant taken from prior work [35], but it influences the reference set quality.
  • Active-set tolerance tol and line-search constants = ε=1e-6, tolerances from [35]
    Algorithmic hyperparameters; minor influence on convergence but part of the method specification.
assumptions (6)
  • domain assumption Objective function F is C^2 with C^2 constraints
    Section 1 states F ∈ C^2(R^n) and h_i, g_j are C^2, which is required for the Newton method and Hessian computations; excluded problems such as ZDT5 and ZDT6 do not satisfy this.
  • standard math Bochner's spectral representation of MMD for bounded stationary kernels
    Eq. (2) relies on Bochner's theorem to write MMD as an integral over the spectral measure G; this is standard for positive definite stationary kernels.
  • domain assumption Spectral measure G has finite second and fourth moments
    Theorem 2 assumes finite C, m2, m4 to bound eigenvalues; this limits the kernel class (e.g., Gaussian satisfies it).
  • ad hoc to paper The reference set R derived from Y0 by clustering and utopian shifting approximates the true Pareto front well enough
    Algorithm 1 lines 5-9 and Section B construct R from Y0; the success of MMDN depends on this heuristic assumption, which fails for MOEA/D on DTLZ7 as the authors acknowledge in Section 7.
  • ad hoc to paper MMD minimization with an imperfect reference set also improves Pareto front diversity
    The paper's core advantage claim that "MMD also maximizes diversity" is only illustrated in Fig. 1 and not proven; the stationary condition analysis (Lemma 1) is not sufficient to establish this.
  • ad hoc to paper G non-degenerating implies pointwise equality in the stationary condition
    In the proof of Lemma 1 (Section A.1), the authors assert that the gradient being zero forces the integrand bracket to vanish for all ω, which does not follow from the vanishing of an integral of ω times the bracket; this is an unproved and generally false step.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MMD-Newton Method for Multi-objective Optimization." pith.science (2026). https://pith.science/paper/KRG2XNUJ

@misc{pith2026250514610,
  author       = {Pith},
  title        = {Pith review of: MMD-Newton Method for Multi-objective Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KRG2XNUJ}},
  note         = {Machine review of arXiv:2505.14610}
}
read the original abstract

Maximum mean discrepancy (MMD) has been widely employed to measure the distance between probability distributions. In this paper, we propose using MMD to solve continuous multi-objective optimization problems (MOPs). For solving MOPs, a common approach is to minimize the distance (e.g., Hausdorff) between a finite approximate set of the Pareto front and a reference set. Viewing these two sets as empirical measures, we propose using MMD to measure the distance between them. To minimize the MMD value, we provide the analytical expression of its gradient and Hessian matrix w.r.t. the search variables, and use them to devise a novel set-oriented, MMD-based Newton (MMDN) method. Also, we analyze the theoretical properties of MMD's gradient and Hessian, including the first-order stationary condition and the eigenspectrum of the Hessian, which are important for verifying the correctness of MMDN. To solve complicated problems, we propose hybridizing MMDN with multiobjective evolutionary algorithms (MOEAs), where we first execute an EA for several iterations to get close to the global Pareto front and then warm-start MMDN with the result of the MOEA to efficiently refine the approximation. We empirically test the hybrid algorithm on 11 widely used benchmark problems, and the results show the hybrid (MMDN + MOEA) can achieve a much better optimization accuracy than EA alone with the same computation budget.

Figures

Figures reproduced from arXiv: 2505.14610 by the authors.

Figure 1
Figure 1. Example of MMD-Newton (MMDN) achieving uniform coverage of the Pareto front with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Example of the Pareto optimal condition of MMD with the Gaussian kernel for [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Quantum Algorithm for Protein Side-Chain Optimisation: Comparing Quantum to Classical Methods

    quant-ph 2025-07 conditional novelty 5.0 of 10

    The authors show that QAOA with a local XY mixer finds ground-state rotamer configurations for small peptides with a milder fitted exponential scaling than their simulated annealing baseline, suggesting a crossover ar...

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    Nicola Beume, Boris Naujoks, and Michael T. M. Emmerich. SMS-EMOA: Multiobjective selection based on dominated hypervolume.Eur. J. Oper. Res., 181(3):1653–1669, 2007

    work page 2007

  2. [2]

    Pymoo: Multi-Objective Optimization in Python.IEEE Access, 8:89497–89509, 2020

    Julian Blank and Kalyanmoy Deb. Pymoo: Multi-Objective Optimization in Python.IEEE Access, 8:89497–89509, 2020

    work page 2020

  3. [3]

    Monotone Funktionen, Stieltjessche Integrale und harmonische Analyse

    Salomon Bochner. Monotone Funktionen, Stieltjessche Integrale und harmonische Analyse. Mathematische Annalen, 108(1):378–410, 1933

    work page 1933

  4. [4]

    Borgwardt, Arthur Gretton, Malte J

    Karsten M. Borgwardt, Arthur Gretton, Malte J. Rasch, Hans-Peter Kriegel, Bernhard Schölkopf, and Alexander J. Smola. Integrating structured biological data by kernel maximum mean discrepancy. InProceedings 14th International Conference on Intelligent Systems for Molecular Biology 2006, Fortaleza, Brazil, August 6-10, 2006, pages 49–57, 2006

    work page 2006

  5. [5]

    JAX: composable transformations of Python+NumPy programs, 2018

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018

    2018

  6. [6]

    A Kernel-Based Indicator for Multi/Many-Objective Optimization.IEEE Trans

    Xinye Cai, Yushun Xiao, Zhenhua Li, Qi Sun, Hanchuan Xu, Miqing Li, and Hisao Ishibuchi. A Kernel-Based Indicator for Multi/Many-Objective Optimization.IEEE Trans. Evol. Comput., 26(4):602–615, 2022

    work page 2022

  7. [7]

    Coello Coello and Margarita Reyes Sierra

    Carlos A. Coello Coello and Margarita Reyes Sierra. A Study of the Parallelization of a Coevo- lutionary Multi-objective Evolutionary Algorithm. In Raúl Monroy, Gustavo Arroyo-Figueroa, Luis Enrique Sucar, and Juan Humberto Sossa Azuela, editors,MICAI 2004: Advances in Artifi- cial Intelligence, Third Mexican International Conference on Artificial Intelli...

    work page 2004

  8. [8]

    Meyarivan

    Kalyanmoy Deb, Samir Agrawal, Amrit Pratap, and T. Meyarivan. A fast and elitist multiobjec- tive genetic algorithm: NSGA-II.IEEE Trans. Evol. Comput., 6(2):182–197, 2002

    work page 2002

Show all 41 references
  1. [9]

    An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints.IEEE Trans

    Kalyanmoy Deb and Himanshu Jain. An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints.IEEE Trans. Evol. Comput., 18(4):577–601, 2014

    2014

  2. [10]

    Scalable Test Prob- lems for Evolutionary Multiobjective Optimization

    Kalyanmoy Deb, Lothar Thiele, Marco Laumanns, and Eckart Zitzler. Scalable Test Prob- lems for Evolutionary Multiobjective Optimization. InProceedings of the 2002 Congress on Evolutionary Computation (CEC’2002), pages 825–830. IEEE Press, 2002

    2002

  3. [11]

    Delaunay

    B. Delaunay. Sur la sphère vide.Bulletin de lAcadémie des Sciences de l’URSS. Classe des sciences mathématiques et na, 1934(6):793–800, 1934

    1934

  4. [12]

    Eichfelder.Adaptive Scalarization Methods in Multiobjective Optimization

    G. Eichfelder.Adaptive Scalarization Methods in Multiobjective Optimization. Springer, 2008

    2008

  5. [13]

    Deutz, Hao Wang, and Carlos M

    Michael Emmerich, Kaifeng Yang, André H. Deutz, Hao Wang, and Carlos M. Fonseca. A Multicriteria Generalization of Bayesian Global Optimization. In Panos M. Pardalos, Anatoly Zhigljavsky, and Julius Zilinskas, editors,Advances in Stochastic and Deterministic Global Optimizatio...

    2016

  6. [14]

    A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. In Evangelos Simoudis, Jiawei Han, and Usama M. Fayyad, editors,Proceedings of the Second International Conference on Knowle...

    1996

  7. [15]

    Steepest descent methods for multicriteria optimization

    Jörg Fliege and Benar Fux Svaiter. Steepest descent methods for multicriteria optimization. Math. Methods Oper. Res., 51(3):479–494, 2000

    2000

  8. [16]

    Marc G. Genton. Classes of Kernels for Machine Learning: A Statistics Perspective.J. Mach. Learn. Res., 2:299–312, 2001

    2001

  9. [17]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander J. Smola. A Kernel Two-Sample Test.J. Mach. Learn. Res., 13:723–773, 2012

    2012

  10. [18]

    Griewank.Evaluating derivatives - principles and techniques of algorithmic differentiation, volume 19 ofFrontiers in Applied Mathematics

    A. Griewank.Evaluating derivatives - principles and techniques of algorithmic differentiation, volume 19 ofFrontiers in Applied Mathematics. SIAM, 2000

    2000

  11. [19]

    Modified Distance Calculation in Generational Distance and Inverted Generational Distance

    Hisao Ishibuchi, Hiroyuki Masuda, Yuki Tanigaki, and Yusuke Nojima. Modified Distance Calculation in Generational Distance and Inverted Generational Distance. In António Gaspar- Cunha, Carlos Henggeler Antunes, and Carlos A. Coello Coello, editors,Evolutionary Multi- Criterion...

    2015

  12. [20]

    Himanshu Jain and Kalyanmoy Deb. An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point Based Nondominated Sorting Approach, Part II: Handling Constraints and Extending to an Adaptive Approach.IEEE Trans. Evol. Comput., 18(4):602–622, 2014

    2014

  13. [21]

    A survey on evolutionary constrained multiobjective optimization.IEEE Transactions on Evolutionary Computation, 27(2):201–221, 2023

    Jing Liang, Xuanxuan Ban, Kunjie Yu, Boyang Qu, Kangjia Qiao, Caitong Yue, Ke Chen, and Kay Chen Tan. A survey on evolutionary constrained multiobjective optimization.IEEE Transactions on Evolutionary Computation, 27(2):201–221, 2023

    2023

  14. [22]

    Margossian

    Charles C. Margossian. A review of automatic differentiation and its efficient implementation. WIREs Data Mining Knowl. Discov., 9(4), 2019

    2019

  15. [23]

    Martín and O

    A. Martín and O. Schütze. Pareto tracer: A predictor-corrector method for multi-objective optimization problems.Engineering Optimization, 50(3):516–536, 2018

    2018

  16. [24]

    Springer Science & Business Media, 2012

    Kaisa Miettinen.Nonlinear multiobjective optimization, volume 12. Springer Science & Business Media, 2012

    2012

  17. [25]

    Sriperumbudur, and Bernhard Schölkopf

    Krikamol Muandet, Kenji Fukumizu, Bharath K. Sriperumbudur, and Bernhard Schölkopf. Kernel Mean Embedding of Distributions: A Review and Beyond.Found. Trends Mach. Learn., 10(1-2):1–141, 2017

    2017

  18. [26]

    Wright.Numerical Optimization

    Jorge Nocedal and Stephen J. Wright.Numerical Optimization. Springer, 1999

    1999

  19. [27]

    Rodriguez-Fernandez, Hao Wang, and Oliver Schütze

    Angel E. Rodriguez-Fernandez, Hao Wang, and Oliver Schütze. Reference Set Generator: A Method for Pareto Front Approximation and Reference Set Generation.Mathematics, 13(10), 2025

    2025

  20. [28]

    The Pareto tracer for the treatment of degenerated multi- objective optimization problems.Engineering Optimization, pages 1–26, 2024

    Oliver Schütze and Oliver Cuate. The Pareto tracer for the treatment of degenerated multi- objective optimization problems.Engineering Optimization, pages 1–26, 2024

    2024

  21. [29]

    Coello Coello

    Oliver Schütze, Xavier Esquivel, Adriana Lara, and Carlos A. Coello Coello. Using the Averaged Hausdorff Distance as a Performance Measure in Evolutionary Multiobjective Optimization. IEEE Trans. Evol. Comput., 16(4):504–522, 2012

    2012

  22. [30]

    Sinha and J

    A. Sinha and J. Wallenius. MCDM, EMO and hybrid approaches: Tutorial and review.Mathe- matical and Computational Applications, 27(6), 2022

    2022

  23. [31]

    Deutz, and Michael Emmerich

    Víctor Adrián Sosa-Hernández, Oliver Schütze, Hao Wang, André H. Deutz, and Michael Emmerich. The Set-Based Hypervolume Newton Method for Bi-Objective Optimization.IEEE Trans. Cybern., 50(5):2186–2196, 2020. 11

    2020

  24. [32]

    Sriperumbudur, Arthur Gretton, Kenji Fukumizu, Bernhard Schölkopf, and Gert R

    Bharath K. Sriperumbudur, Arthur Gretton, Kenji Fukumizu, Bernhard Schölkopf, and Gert R. G. Lanckriet. Hilbert Space Embeddings and Metrics on Probability Measures.J. Mach. Learn. Res., 11:1517–1561, 2010

    2010

  25. [33]

    World Scientific, 2008

    Christiane Tretter.Spectral theory of block operator matrices and applications. World Scientific, 2008

    2008

  26. [34]

    Controlling Wasserstein Distances by Kernel Norms with Application to Compressive Statistical Learning.J

    Titouan Vayer and Rémi Gribonval. Controlling Wasserstein Distances by Kernel Norms with Application to Compressive Statistical Learning.J. Mach. Learn. Res., 24:149:1–149:51, 2023

    2023

  27. [35]

    Rodriguez-Fernandez, Lourdes Uribe, André H

    Hao Wang, Angel E. Rodriguez-Fernandez, Lourdes Uribe, André H. Deutz, Oziel Cortés-Piña, and Oliver Schütze. A Newton Method for Hausdorff Approximations of the Pareto Front within Multi-objective Evolutionary Algorithms.CoRR, abs/2405.05721, 2024

    2024 arXiv

  28. [36]

    Probability Distribution of Hypervolume Improvement in Bi-objective Bayesian Optimization

    Hao Wang, Kaifeng Yang, and Michael Affenzeller. Probability Distribution of Hypervolume Improvement in Bi-objective Bayesian Optimization. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024

    2024

  29. [37]

    Deutz, and Thomas Bäck

    Kaifeng Yang, Michael Emmerich, André H. Deutz, and Thomas Bäck. Efficient computation of expected hypervolume improvement using box decomposition algorithms.J. Glob. Optim., 75(1):3–34, 2019

    2019

  30. [38]

    MOEA/D: A multiobjective evolutionary algorithm based on decomposition.IEEE Trans

    Qingfu Zhang and Hui Li. MOEA/D: A multiobjective evolutionary algorithm based on decomposition.IEEE Trans. Evol. Comput., 11(6):712–731, 2007

    2007

  31. [39]

    Comparison of multiobjective evolutionary algorithms: Empirical results.Evolutionary computation, 8(2):173–195, 2000

    Eckart Zitzler, Kalyanmoy Deb, and Lothar Thiele. Comparison of multiobjective evolutionary algorithms: Empirical results.Evolutionary computation, 8(2):173–195, 2000. A Proofs A.1 Proof of Lemma 1 Proof.We take the spectral representation of MMD (Eq. (2)): MMD2(Y, R) = Z Rk |...

    2000

  32. [40]

    We use Gershgorin’s circle theorem for block matrices [33]

    Applying Weyl’s inequality to H m ℓ , i.e., σmin(H m ℓ )≥σ min(H1)−σ max(H2) and σmax(H)≥σ max(H1)−σ min(H2) leads to: 2 µ σmin (C)−m 2 − m4 2 ¯D2(⃗ yℓ, R) ≤σ(H m ℓ )≤ 2 µ m2 −σ min (C) + m4 2 ¯D2(⃗ yℓ, Y) .(27) A.4 Proof of Corollary 1 Proof. We use Gershgorin’s circle theore...

  33. [41]

    After filling, we perform b),reduction, to return to exactly µ elements

    to the set, and fill each triangle uniformly at random with a number of points proportional to its area. After filling, we perform b),reduction, to return to exactly µ elements. This is achieved by applying k-means clustering to the filled set and selecting the centroids of th...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.