Pith. sign in

REVIEW 4 major objections 8 minor 39 references

Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks

T0 review · 4 major / 8 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read SVH-PSL claims that adding Stein Variational Gradient Descent repulsion to hypernetwork-based Pareto set learning avoids surrogate-induced pseudo-local optima and substantially improves the quality of the learned Pareto set for expensive…

desk verdict Plausible empirical method with a load-bearing kernel that is underspecified and likely asymmetric as written; worth reviewing, but the SVGD theory needs a fix. read the letter →

arxiv 2412.17312 v3 pith:ZWTNL7SJ submitted 2024-12-23 cs.LG cs.NEstat.ML

Minh-Duc Nguyen , Phuong Mai Dinh , Quang-Huy Nguyen , Long P. Hoang , Dung D. Le This is my paper · ORCID
classification cs.LGcs.NEstat.ML MSC 90C2968T07
keywords Paretosetlearningexpensivemulti-objectiveoptimizationSteinVariationalGradientDescenthypernetworksBayesianGaussianprocesssurrogateskernelrepulsionlocal
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

Expensive multi-objective optimization problems are those where each evaluation of the objectives costs real resources, so algorithms must find good trade-off solutions with as few evaluations as possible. Existing Pareto set learning methods fit Gaussian process surrogates and then train a hypernetwork to map preference vectors to solutions, but fragmented surrogate landscapes create pseudo-local optima that trap the search. This paper proposes SVH-PSL, which replaces the plain gradient update of the hypernetwork with a Stein Variational Gradient Descent (SVGD) update: particles representing candidate solutions feel kernel-weighted attraction toward the scalarized objective plus a repulsive force that keeps them spread out. A newly designed local kernel weights each objective dimension separately, which the authors show helps on real-world problems with complex Pareto fronts. If the method works as claimed, practitioners can recover a full Pareto front for an expensive black-box problem from roughly the same evaluation budget that older methods need just to find a few points.

What carries the argument

The load-bearing object is the SVGD-style parameter update of Eq. (14), applied to the hypernetwork that maps preference vectors $r$ to solutions $x_r = h(r|\theta)$. In this update, each particle $F_i = \hat{f}(x_{r_i})$ is moved by $-\xi \sum_{i,j} \nabla_\theta g(F_i|r_i) k(F_i,F_j) - \alpha \nabla_\theta k(F_i,F_j)$, where the first term follows the Chebyshev scalarization gradient and the second term is the repulsive force that maintains diversity. The companion local kernel of Eq. (15) computes $k(F_i,F_j) = \sum_{n=1}^m k_n(\hat{f}_n(x_{r_i}), \hat{f}_n(x_{r_j})) \cdot \mathbb{1}_n$, where $\mathbb{1}_n$ indicates the objective dimension that attains the maximum in the Chebyshev scalarization; this per-dimension weighting is what lets particles respond to the most constraining objective and is credited with the stability gains on complex fronts.

What would settle it

Run SVH-PSL on a simple two-objective problem with a known analytic Pareto front while using a deliberately misspecified Gaussian process to create a highly fragmented surrogate, then check whether the particle set still lands on the true front: if the repulsion term pushes particles into demonstrably non-Pareto regions, or if the update fails to improve any measure of convergence to the target distribution on a one-dimensional analogue, the paper's central claim would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that the failure mode of Pareto set learning for expensive black-box problems—getting stuck in pseudo-local optima created by fragmented Gaussian process surrogates—can be mitigated by treating the hypernetwork's outputs as interacting SVGD particles. The update in Eq. (14) modifies ordinary gradient descent on the Chebyshev scalarization loss by adding a kernel-weighted average over particles and a repulsion term $\alpha \nabla_\theta k(F_i, F_j)$, and the local kernel in Eq. (15) computes the kernel per objective dimension, activating only the dimension that achieves the maximum in the scalarization. The paper reports that this collective particle movement smooths the solution space, prevents particles from clustering at false optima, and yields significantly lower Log Hypervolume Difference than the PSL-MOBO baseline and other MOBO methods across ZDT, F2, VLMOP2, and seven real-world benchmarks.

Load-bearing premise

The load-bearing premise is that the kernel-weighted update in Eq. (14) is a valid Stein-type functional gradient: the repulsion between particles must actually guide the whole cloud toward the true Pareto set rather than merely being a heuristic that happened to work on the benchmarks tested.

Editorial extensions

If this is right

  • If correct, the evaluation budget for approximating a full Pareto front in expensive black-box problems can be substantially reduced, since the interacting particles explore diverse regions in parallel rather than collapsing onto false optima.
  • The local kernel makes the method more stable on problems with complex or disconnected Pareto fronts, as shown by lower variance across runs on RE37, RE41, and RE42.
  • Replacing the multi-gradient descent of earlier MOO-SVGD with the scalarization gradient preserves preference conditioning, so the learned mapping still lets users trade off objectives in real time.
  • The method inherits hypernetwork benefits: after training, new preference vectors can be evaluated without additional expensive function calls.
  • Stronger diversity of the learned Pareto set follows directly from the repulsion term, which the ablation ties to the $\alpha$ trade-off parameter.

Reading between the lines

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

  • The paper does not prove that Eq. (14) is a true Stein functional gradient for the hypernetwork parameters; a sympathetic reading treats it as a heuristic that borrows SVGD's repulsion intuition, and the theoretical status of the update is an open question.
  • The local kernel idea could be transplanted to other acquisition functions (EI, UCB) and other surrogate models beyond Gaussian processes; nothing in its formulation is GP-specific.
  • The method's claim of smoothing fragmented landscapes suggests a testable comparison against ensemble or deep surrogate models that produce smoother uncertainty estimates, which might reduce the need for repulsion altogether.
  • If the repulsion term is doing the main work, then a simpler diversity-preserving mechanism (e.g., deterministic particle repulsion without the Stein weighting) might achieve similar gains; comparing against such an ablation would isolate the contribution of the Stein formalism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper proposes SVH-PSL, a method for learning the Pareto set of expensive multi-objective optimization problems by combining Stein Variational Gradient Descent (SVGD) with a hypernetwork that maps preference vectors to candidate solutions. The core update (Eq. 14) modifies ordinary gradient descent on a Chebyshev scalarization by adding kernel-weighted attraction and repulsion terms over a set of sampled particles. A 'local kernel' (Eq. 15) is introduced to focus kernel interactions on the objective that currently dominates each particle's scalarization. The authors report experiments on ZDT, VLMOP2, and real-world benchmarks (RE21, RE32, RE33, RE36, RE37, RE41, RE42), comparing LHD against ten MOBO baselines and presenting ablations on the repulsion coefficient and dimensionality. The central claim is that SVH-PSL significantly improves the quality of the learned Pareto set.

Significance. If the method is made well-defined and its theoretical grounding supplied, it addresses a practically important problem: GP surrogates in expensive MOO can create fragmented landscapes with pseudo-local optima, and a particle-interaction mechanism that provably avoids these traps would be a useful contribution. The paper is also commendable for benchmarking against a wide range of established MOBO methods and for releasing code. However, the current manuscript does not establish that Eq. (14) is a valid SVGD functional-gradient update, and Eq. (15) does not define a symmetric positive-definite kernel, so the central mechanism—the claimed 'collective particle movement'—is not reproducible from the description. The empirical claims would be stronger with statistical significance testing. The testing setup uses external Pareto fronts and standard baselines, so there is no circularity in the evaluation.

major comments (4)
  1. [Section 'Stein Variational Hypernetworks', Eq. (14)] The update rule in Eq. (14) is stated without derivation and does not follow from the SVGD formalism in Eqs. (10)-(12). Standard SVGD updates each particle using a functional gradient in an RKHS that involves the log-density gradient ∇_x log p(x) and a symmetric positive-definite kernel; the optimal perturbation φ* is the maximizer of a KL-divergence rate. Eq. (14), by contrast, updates shared hypernetwork parameters θ with a double sum over particles, replacing ∇_x log p(x) with ∇_θ g(F_i|r_i) and adding a repulsion term. The paper does not define a target distribution p, does not show which KL divergence is minimized, and does not explain how the per-particle Stein gradient translates to a parameter-space update. Because the abstract and introduction attribute the avoidance of pseudo-local optima to the SVGD-style collective movement, this is a load-bearing gap: without a derivation (or an explicit statement that Eq. (14) is a heuristic modification), the claimed theoretical grounding is unsupported.
  2. [Section 'Design of SVGD Local Kernel', Eq. (15)] The local kernel in Eq. (15) is not well-defined for a pair of particles. The indicator id_n is described as 1 when n is the index of the objective that achieves the maximum in formula (9), but for a pair (F_i,F_j) the Chebyshev maximizer is generally different for particle i and for particle j, and the formula does not specify whose maximizer is used. If id_n is taken from particle i, the kernel is asymmetric: k(F_i,F_j) ≠ k(F_j,F_i) in general. Standard SVGD (Eqs. 10-12) requires a symmetric positive-definite kernel in an RKHS; an asymmetric 'kernel' has no associated RKHS, so Eq. (14) cannot be interpreted as a Stein functional gradient. This is load-bearing because the paper's novelty is precisely this local kernel and its integration with SVGD. The authors must specify a symmetric construction (e.g., using the union or average of the active objectives of both particles, or a soft indicator) and confirm that the resulting kernel is positive definite.
  3. [Section 'Design of SVGD Local Kernel', Figure 2] The claim that the local kernel 'facilitates precise adjustments' and prevents particles from getting stuck in pseudo-local optima is not supported by analysis. Even if Eq. (15) is symmetrized, the data-dependent indicator makes the kernel a random, input-dependent object, and no convergence, diversity, or repulsion guarantee is provided. The paper should either prove the relevant kernel properties (symmetry, positive definiteness) and state what distribution the SVGD update targets, or explicitly label the method as heuristic and give empirical diagnostics (e.g., particle distances over training, or failure cases without the local kernel) that support the proposed mechanism.
  4. [Section 'Experimental Results and Analysis', Figure 3] The abstract claims SVH-PSL 'significantly improves the quality of the learned Pareto set', but the experimental section reports only mean LHD curves with shaded standard-deviation regions and provides no statistical significance tests. Several subplots (e.g., RE21, RE32) show overlapping confidence bands, and no table of final values or pairwise comparison tests (e.g., Wilcoxon signed-rank) is given. The word 'significantly' is therefore not currently justified by the evidence. Please add proper statistical analysis or temper the claim.
minor comments (8)
  1. [Equation (2)] The hypervolume formula is garbled: the notation 'Πn i=1 h y(i), ρi i' is not meaningful as printed, and the union of hypercubes does not correspond to the standard definition of the dominated region. Please rewrite this formula in standard notation.
  2. [Equation (3)] The covariance kernel is typed as 'k : X × X → R2'; it should be a scalar map, e.g., 'k : X × X → R' (or 'R+').
  3. [Algorithm 1 and Section 'Experiments'] The pseudocode says preference vectors are sampled from the simplex S_m, while the text says they are sampled from a Dirichlet distribution Dir(α). Please make the sampling procedure consistent and specify the parameters used in the experiments.
  4. [Equation (8)] The constraint 'h(S_m|θ*) = P' is stated without explanation; it is unclear in what sense the image of the entire simplex under the hypernetwork equals the Pareto set, especially given that the model is trained with a finite set of sampled preference vectors.
  5. [Section 'Design of SVGD Local Kernel'] The form of the per-dimension kernel k_n(·) is not specified. The authors state that a Gaussian kernel is used in Eq. (14), but it is not clear whether Eq. (15) uses the same kernel and the same bandwidth c for each dimension. Please state the exact kernel form and bandwidth chosen for the local kernel.
  6. [Figure 1 caption] The caption says the comparison is 'after the first 6 function evaluations', but the experimental setup uses 20 initial evaluations plus batches of 5. Please clarify how the six-evaluation snapshot is obtained.
  7. [Section 'Stein Variational Hypernetworks'] The sentence 'Notably, our study proposes an innovative approach which is different from MOO-SVGD by (Liu, Tong, and Liu 2021)' is not grammatically complete and the contrast with MOO-SVGD is not explained precisely; please rephrase and clarify the difference.
  8. [Throughout] There are several typos and notation inconsistencies, including 'Pareto Set Learningor PSL', 'the operator creating the n-dimensional hypercube', and inconsistent use of hats over objective functions (e.g., F_i sometimes written with and without hat). A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SVH-PSL is an empirical heuristic tested against external Pareto fronts and independent baselines.

full rationale

The paper's derivation chain does not reduce to its own inputs. Equation (14) is presented as a modification of gradient descent with a kernel-weighted repulsion term, and although the paper calls this an SVGD-style update without deriving it from the KL-divergence objective in Eq. (12), that is a correctness and well-definedness concern, not circularity. The local kernel in Eq. (15) is defined using the same Chebyshev scalarization from Eq. (9) that is used for training, but this is a design choice rather than a fitted parameter renamed as a prediction. The central empirical claim, that SVH-PSL improves the learned Pareto set, is evaluated against external ground-truth Pareto fronts and multiple independent baselines such as PSL-MOBO, qEHVI, qNEHVI, and DA-PSL; no fitted quantity is relabeled as a prediction. Citations to the authors' prior works, such as Hoang et al. (2023) and Tuan et al. (2024), supply background hypernetwork methodology, but the load-bearing experimental comparison is self-contained against external benchmarks, so the self-citation is not load-bearing. The skeptical observation that the kernel in Eq. (15) may be asymmetric as written is a legitimate technical critique of the method's specification, but it is not an instance of the paper deriving a result from its own assumptions by construction. Accordingly, no circular step meeting the required evidentiary standard was found.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim sits on three standard domain assumptions (surrogate calibration, Chebyshev coverage, hypernetwork expressiveness) and one ad hoc methodological assumption (Eq. 14 as a valid SVGD update). The free parameters are tuning choices rather than fitted physical constants; none is hidden as a prediction.

free parameters (3)
  • SVGD repulsion coefficient alpha = 0.1
    Controls the divergence/repulsion term in Eq. 14; ablation in Figure 5 shows sensitivity, and the paper fixes it at 0.1 without principled selection.
  • Gaussian kernel bandwidth c = median of pairwise distances
    Computed by median heuristic from samples; it is a data-dependent scale rather than an optimized parameter, but it changes the interaction weights in Eq. 14.
  • Chebyshev ideal point z* = not specified
    Required in Eq. 9 but the paper never states how z* is estimated; typically approximated from data, which is a free choice affecting the scalarization.
assumptions (4)
  • domain assumption Gaussian process surrogates provide calibrated uncertainty for LCB and hypervolume improvement acquisition.
    The entire acquisition loop in Algorithm 1 updates GPs on D and uses their predictive means and variances to define Fi and to select the next batch.
  • domain assumption Chebyshev scalarization with ideal point z* is sufficient to recover the full Pareto set as preference vectors vary.
    Eq. 9 defines the training loss; the local kernel in Eq. 15 uses the maximizing dimension, so the method inherits this coverage assumption.
  • domain assumption The hypernetwork h(r|theta) can represent the true Pareto set mapping with enough capacity.
    Standard PSL assumption; the paper cites hypernetwork PSL works rather than proving representability.
  • ad hoc to paper Equation 14 is a valid SVGD-style functional gradient update for hypernetwork parameters.
    Stated as a modification of Eq. 13 without derivation; the paper does not show it minimizes a KL divergence or preserves convergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks." pith.science (2026). https://pith.science/paper/ZWTNL7SJ

@misc{pith2026241217312,
  author       = {Pith},
  title        = {Pith review of: Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZWTNL7SJ}},
  note         = {Machine review of arXiv:2412.17312}
}
read the original abstract

Expensive multi-objective optimization problems (EMOPs) are common in real-world scenarios where evaluating objective functions is costly and involves extensive computations or physical experiments. Current Pareto set learning methods for such problems often rely on surrogate models like Gaussian processes to approximate the objective functions. These surrogate models can become fragmented, resulting in numerous small uncertain regions between explored solutions. When using acquisition functions such as the Lower Confidence Bound (LCB), these uncertain regions can turn into pseudo-local optima, complicating the search for globally optimal solutions. To address these challenges, we propose a novel approach called SVH-PSL, which integrates Stein Variational Gradient Descent (SVGD) with Hypernetworks for efficient Pareto set learning. Our method addresses the issues of fragmented surrogate models and pseudo-local optima by collectively moving particles in a manner that smooths out the solution space. The particles interact with each other through a kernel function, which helps maintain diversity and encourages the exploration of underexplored regions. This kernel-based interaction prevents particles from clustering around pseudo-local optima and promotes convergence towards globally optimal solutions. Our approach aims to establish robust relationships between trade-off reference vectors and their corresponding true Pareto solutions, overcoming the limitations of existing methods. Through extensive experiments across both synthetic and real-world MOO benchmarks, we demonstrate that SVH-PSL significantly improves the quality of the learned Pareto set, offering a promising solution for expensive multi-objective optimization problems.

Figures

Figures reproduced from arXiv: 2412.17312 by the authors.

Figure 1
Figure 1. Approximate Pareto front comparison after the first [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the impact of neighboring particles [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Mean Log Hypervolume Differences between the truth Pareto Front and the learned Pareto Front with respect to the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of SVH-PSL performance with and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Impact of α on the trade-off between exploration and exploitation in ZDT1 (LHD & Pareto front). 40 80 120 4 2 0 LHD ZDT1 (Dim=30) 40 80 120 2.5 2.0 1.5 1.0 0.5 ZDT2 (Dim=30) 40 80 120 Number of evaluation 1.5 1.0 0.5 LHD ZDT1 (Dim=50) 40 80 120 Number of evaluation 1.5…
Figure 6
Figure 6. Figure 6: Performance comparison of high-dimensional [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 26 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Abdolshah, M.; Shilton, A.; Rana, S.; Gupta, S.; and Venkatesh, S. 2019. Multi-objective Bayesian optimisation with preferences over objectives. arXiv:1902.04228

    work page Pith review arXiv 2019

  4. [4]

    Ai, Q.; Liu, S.; He, L.; and Xu, Z. 2023. Stein variational gradient descent with multiple kernels. Cognitive Computation, 15(2): 672--682

    work page 2023

  5. [5]

    M.; Grover, A.; Jin, N.; Severson, K

    Attia, P. M.; Grover, A.; Jin, N.; Severson, K. A.; Markov, T. M.; Liao, Y.-H.; Chen, M. H.; Cheong, B.; Perkins, N.; Yang, Z.; Herring, P. K.; Aykol, M.; Harris, S. J.; Braatz, R. D.; Ermon, S.; and Chueh, W. C. 2020. Closed-loop optimization of fast-charging protocols for batteries with machine learning. Nature, 578: 397--402

    work page 2020

  6. [6]

    K.; and Doppa, J

    Belakaria, S.; Deshwal, A.; Jayakodi, N. K.; and Doppa, J. R. 2020. Uncertainty-aware search framework for multi-objective Bayesian optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, 10044--10052

    work page 2020

  7. [7]

    M.; and Lapkin, A

    Bradford, E.; Schweidtmann, A. M.; and Lapkin, A. 2018. Efficient multiobjective optimization employing Gaussian processes, spectral sampling and a genetic algorithm. Journal of global optimization, 71(2): 407--438

    work page 2018

  8. [8]

    K.; Zhou, J.; Lu, P.; Molaei, S.; and Clifton, D

    Chauhan, V. K.; Zhou, J.; Lu, P.; Molaei, S.; and Clifton, D. A. 2023. A Brief Review of Hypernetworks in Deep Learning. arXiv preprint arXiv:2306.06955

    arXiv 2023

Show all 39 references
  1. [9]

    Daulton, S.; Balandat, M.; and Bakshy, E. 2020. Differentiable expected hypervolume improvement for parallel multi-objective Bayesian optimization. Advances in Neural Information Processing Systems, 33: 9851--9864

    2020

  2. [10]

    Daulton, S.; Balandat, M.; and Bakshy, E. 2021. Parallel bayesian optimization of multiple noisy objectives with expected hypervolume improvement. Advances in Neural Information Processing Systems, 34: 2187--2200

    2021

  3. [11]

    Deb, K.; and Srinivasan, A. 2006. Innovization: Innovating design principles through optimization. In Proceedings of the 8th annual conference on Genetic and evolutionary computation, 1629--1636

    2006

  4. [12]

    D \'e sid \'e ri, J.-A. 2012. Multiple-gradient descent algorithm (MGDA) for multiobjective optimization. Comptes Rendus Mathematique, 350(5-6): 313--318

    2012

  5. [13]

    He, C.; Zhang, Y.; Gong, D.; and Ji, X. 2023. A review of surrogate-assisted evolutionary algorithms for expensive optimization problems. Expert Systems with Applications, 217: 119495

    2023

  6. [14]

    P.; Le, D

    Hoang, L. P.; Le, D. D.; Tuan, T. A.; and Thang, T. N. 2023. Improving pareto front learning via multi-sample hypernetworks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 7875--7883

    2023

  7. [15]

    D.; Gelman, A.; et al

    Hoffman, M. D.; Gelman, A.; et al. 2014. The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. J. Mach. Learn. Res., 15(1): 1593--1623

    2014

  8. [16]

    Hou, M.; Chen, F.; and Pei, Y. 2024. Optimization of geometric parameters of ejector for fuel cell system based on multi-objective optimization method. International Journal of Green Energy, 21(2): 228--243

    2024

  9. [17]

    Knowles, J. 2006. ParEGO: a hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems. IEEE Transactions on Evolutionary Computation, 10(1): 50--66

    2006

  10. [18]

    Konakovic Lukovic, M.; Tian, Y.; and Matusik, W. 2020. Diversity-guided multi-objective bayesian optimization with batch evaluations. Advances in Neural Information Processing Systems, 33: 17708--17720

    2020

  11. [19]

    Laumanns, M.; and Ocenasek, J. 2002. Bayesian optimization algorithms for multi-objective optimization. In International Conference on Parallel Problem Solving from Nature, 298--307. Springer

    2002

  12. [20]

    D.; and Lauw, H

    Le, D. D.; and Lauw, H. W. 2017. Indexable bayesian personalized ranking for efficient top-k recommendation. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, 1389--1398

    2017

  13. [21]

    H.; Li, Y.; Ke, J.; Yoo, I.; Zhang, H.; Yu, J.; Wang, Q.; Deng, F.; Entis, G.; He, J.; et al

    Lee, S. H.; Li, Y.; Ke, J.; Yoo, I.; Zhang, H.; Yu, J.; Wang, Q.; Deng, F.; Entis, G.; He, J.; et al. 2024. Parrot: Pareto-optimal Multi-Reward Reinforcement Learning Framework for Text-to-Image Generation. arXiv preprint arXiv:2401.05675

    2024 arXiv

  14. [22]

    Lin, X.; Yang, Z.; and Zhang, Q. 2022. Pareto Set Learning for Neural Multi-Objective Combinatorial Optimization. In International Conference on Learning Representations

    2022

  15. [23]

    Lin, X.; Yang, Z.; Zhang, X.; and Zhang, Q. 2022. Pareto set learning for expensive multi-objective optimization. Advances in Neural Information Processing Systems, 35: 19231--19247

    2022

  16. [24]

    Liu, Q.; and Wang, D. 2016. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29

    2016

  17. [25]

    Liu, X.; Tong, X.; and Liu, Q. 2021. Profiling pareto front with multi-objective stein variational gradient descent. Advances in Neural Information Processing Systems, 34: 14721--14733

    2021

  18. [26]

    Lu, Y.; Li, B.; and Zhou, A. 2024. Are You Concerned about Limited Function Evaluations: Data-Augmented Pareto Set Learning for Expensive Multi-Objective Optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 14202--14210

    2024

  19. [27]

    Milojkovic, N.; Antognini, D.; Bergamin, G.; Faltings, B.; and Musat, C. 2019. Multi-gradient descent for multi-objective recommender systems. arXiv preprint arXiv:2001.00846

    2019 arXiv

  20. [28]

    Navon, A.; Shamsian, A.; Fetaya, E.; and Chechik, G. 2021. Learning the Pareto Front with Hypernetworks. In International Conference on Learning Representations

    2021

  21. [29]

    Neal, R. M. 2012. MCMC using Hamiltonian dynamics. arXiv preprint arXiv:1206.1901

    2012 arXiv

  22. [30]

    Paria, B.; Kandasamy, K.; and P \'o czos, B. 2020. A flexible framework for multi-objective bayesian optimization using random scalarizations. In Uncertainty in Artificial Intelligence, 766--776. PMLR

    2020

  23. [31]

    Phan, H.; Tran, N.; Le, T.; Tran, T.; Ho, N.; and Phung, D. 2022. Stochastic multiple target sampling gradient descent. Advances in neural information processing systems, 35: 22643--22655

    2022

  24. [32]

    D.; Bodduluri, M.; Murphy, M

    Schweikard, A.; Glosser, G. D.; Bodduluri, M.; Murphy, M. J.; and Adler, J. R. 2000. Robotic motion compensation for respiratory movement during radiosurgery. Computer aided surgery : official journal of the International Society for Computer Aided Surgery, 5 4: 263--77

    2000

  25. [33]

    Swersky, K.; Snoek, J.; and Adams, R. P. 2013. Multi-task bayesian optimization. Advances in neural information processing systems, 26

    2013

  26. [34]

    Tanabe, R.; and Ishibuchi, H. 2020. An easy-to-use real-world multi-objective optimization problem suite. Applied Soft Computing, 89: 106078

    2020

  27. [35]

    A.; Hoang, L

    Tuan, T. A.; Hoang, L. P.; Le, D. D.; and Thang, T. N. 2024. A framework for controllable pareto front learning with completed scalarization functions and its applications. Neural Networks, 169: 257--273

    2024

  28. [36]

    A.; and Lamont, G

    Van Veldhuizen, D. A.; and Lamont, G. B. 1999. Multiobjective evolutionary algorithm test suites. In Proceedings of the 1999 ACM symposium on Applied computing, 351--357

    1999

  29. [37]

    Zhang, Q.; and Li, H. 2007. MOEA/D: A multiobjective evolutionary algorithm based on decomposition. IEEE Transactions on evolutionary computation, 11(6): 712--731

    2007

  30. [38]

    Zhang, Q.; Liu, W.; Tsang, E.; and Virginas, B. 2009. Expensive multiobjective optimization by MOEA/D with Gaussian process model. IEEE Transactions on Evolutionary Computation, 14(3): 456--474

    2009

  31. [39]

    Zitzler, E.; and Thiele, L. 1999. Multiobjective evolutionary algorithms: a comparative case study and the strength Pareto approach. IEEE transactions on Evolutionary Computation, 3(4): 257--271

    1999

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.