Pith. sign in

REVIEW 4 major objections 5 minor 26 references

VOPy: A Framework for Black-box Vector Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper introduces VOPy, an open-source Python library for black-box vector optimization that orders solutions by convex-cone partial orders instead of componentwise comparison.

desk verdict VOPy fills a real gap for cone-based vector optimization, but the paper describes rather than demonstrates the core comparison solvers. read the letter →

arxiv 2412.06604 v1 pith:YAJH26TW submitted 2024-12-09 cs.LG stat.APstat.ML

classification cs.LGstat.APstat.ML
keywords vectoroptimizationmulti-objectiveBayesianmulti-armedbanditblack-boxconvexconepartialorderconfidenceregioncomparisonopen-sourcesoftware
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

VOPy is an open-source Python library for black-box vector optimization, the setting where several objectives are optimized together under a partial order defined by a convex cone rather than the componentwise order of standard multi-objective optimization. The paper's central claim is that VOPy is the first library built for this task, covering noisy observations, discrete or continuous design spaces, limited budgets, batch observations, and decoupled evaluations. The library organizes around four interfaces—Order, Model, Algorithm, and Problem—so that users can apply existing methods, assemble new algorithms from standard components, or drop in custom orders and models. It also implements a confidence-region comparison mechanism that decides dominance between uncertain solutions by solving convex programs, extending comparisons beyond the hyperrectangles used in ordinary multi-objective tools. A sympathetic reader would care because this fills a missing infrastructure layer for a growing family of vector-optimization algorithms.

What carries the argument

The load-bearing mechanism is the combination of an Order interface and a ConfidenceRegion abstraction. The Order interface encodes a vector partial order $\preceq_C$ on $\mathbb{R}^D$ induced by a closed convex cone $C$, so that $\mu \preceq_C \nu$ iff $\nu - \mu \in C$. The ConfidenceRegion abstraction turns each noisy estimate of an objective vector into a set that contains the true value with high probability, then asks whether one such region is dominated by another under $\preceq_C$. RectangularConfidenceRegion handles the componentwise case, while EllipsoidalConfidenceRegion handles dependent objectives and non-polyhedral cones; both reduce the comparison to a convex program that VOPy solves automatically. This mechanism is what lets the library escape the hyperrectangle comparisons of ordinary multi-objective optimization and support general cone-based preferences.

What would settle it

Run VOPy's confidence-region comparisons on a small grid of noisy objective estimates for a 2D polyhedral cone and verify every dominance verdict against an exact brute-force check; the first mismatch would show the comparison mechanism fails.

Watch

Extended reading notes

Core claim

The paper presents VOPy as the first software framework for developing black-box vector-optimization algorithms. Its key design claim is that any partial order induced by a closed convex cone can be used throughout the stack: the Order interface defines the cone, the Model interface supplies uncertainty estimates, the Algorithm interface schedules evaluations, and the Problem interface connects datasets or real-world systems. To make dominance checks tractable under arbitrary cones, VOPy provides confidence-region classes—rectangular for componentwise orders and ellipsoidal for dependent or non-polyhedral settings—and solves the required comparisons as convex optimization problems. The paper reports that these components are used by implementations of existing vector-optimization and multi-objective algorithms, including PaVeBa, VOGP, naive elimination, and epsilon-PAL, and that decoupled variants allow partial objective evaluations when costs differ.

Load-bearing premise

The library's usefulness rests on the assumption that its built-in confidence-region comparisons, solved as convex programs, return correct dominance verdicts quickly enough for any supported cone; the paper does not provide equations, experiments, or runtime numbers that demonstrate this.

Editorial extensions

If this is right

  • Algorithm developers can build and benchmark new vector-optimization methods against PaVeBa, VOGP, naive elimination, epsilon-PAL, and the decoupled variants without reimplementing infrastructure.
  • Practitioners with noisy, batch, or decoupled objective evaluations can apply the same library to real-world problems such as material design, vehicle safety, and streaming-sorting networks using built-in datasets.
  • Users with non-polyhedral preferences can define a custom Order derivative and reuse existing models and algorithms, extending the library beyond componentwise optimization.
  • The convex-programming confidence-region comparisons replace hyperrectangle comparisons, making dependent-objective models representable exactly instead of approximated by rectangles.
  • The authors report that VOPy's early stages contributed to the development of published vector-optimization algorithms, positioning it as a natural common testbed for future research.

Reading between the lines

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

  • If the confidence-region convex programs scale well, the same comparison machinery could be reused outside vector optimization, such as in safe or constrained Bayesian optimization where feasibility regions need to be compared under general cones.
  • The authors do not report runtime benchmarks for the convex-programming solvers; a natural next step is a systematic comparison of rectangular versus ellipsoidal confidence regions on the same cone-ordered problems, checking both wall-clock time and statistical efficiency.
  • A broader consequence, left implicit, is that a common library with standardized cones and benchmarks could make fairness of comparison across vector-optimization algorithms much easier to audit, since all methods would share the same order logic and uncertainty models.
  • The library's support for decoupled algorithms suggests an untested opportunity: using VOPy to study active-learning strategies that query only the cheapest objectives when costs differ, a direction the included entropy-based decoupled algorithm only begins to explore.
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 / 5 minor

Summary. The paper introduces VOPy, an open-source Python library for black-box vector optimization, where multiple objectives are optimized with respect to a partial order induced by a convex cone. The library is organized around four interfaces (Order, Model, Algorithm, Problem), provides confidence-region abstractions with comparisons solved via CVXPY, and includes implementations of several published algorithms (e.g., PaVeBa, VOGP, epsilon-PAL, and an Auer et al. algorithm) as well as decoupled variants. The manuscript describes the architecture and usage but contains no empirical results, no mathematical formulation of the confidence-region comparison problems, and no runtime or benchmark evaluations; the main supporting evidence is a reported code coverage of at least 95% and the availability of a public repository.

Significance. If the implementation is correct and efficient, VOPy would fill a genuine gap: existing multi-objective optimization libraries, such as PyMOO and BoTorch, are built around the componentwise order and do not directly support general convex-cone-induced partial orders. The modular design, the inclusion of state-of-the-art vector-optimization algorithms, and the support for decoupled evaluations are all valuable contributions for algorithm developers. It is also a strength that the source code is public and that the authors follow standard software-engineering practices (PEP8, Flake8, Black, Bandit). However, the paper does not demonstrate that the central implementation is correct or practical: the convex-programming solutions for confidence-region comparisons are never written down, no correctness proofs are given, and no benchmarks or end-to-end experiments are reported. The significance of the contribution therefore rests on unverified claims about an external repository rather than on evidence contained in the manuscript.

major comments (4)
  1. [Section 2.1] The ConfidenceRegion abstraction is the computational core of the library, but the paper never states the convex program that decides whether one confidence region is preferred over another with respect to a cone C. The text only says that RectangularConfidenceRegion and EllipsoidalConfidenceRegion support these comparisons using CVXPY. Since every algorithm in VOPy must repeatedly invoke this dominance test, the correctness of this encoding is load-bearing. Please provide the explicit optimization formulation for both confidence-region classes, a correctness argument, and a discussion of how the formulation handles polyhedral versus non-polyhedral cones and dependent objectives.
  2. [Section 2.1] The claim that VOPy 'provides significant speed-ups when possible' is supported only by a reference to a 'Performance notebook' that is not included in the paper and has no link or version identifier. No runtime measurements appear anywhere in the manuscript. As a result, the practicality of contribution (2) cannot be assessed. Include the performance notebook or an equivalent benchmark section with wall-clock times, solver statistics, and comparisons against a baseline, or qualify the performance claim so that it is not verifiable only through the repository.
  3. [Section 2.2] Reporting code coverage of at least 95% does not establish the correctness of the convex-programming solvers or the reliability of the implemented algorithms. Coverage measures how many code statements are executed, not whether the returned solutions are mathematically correct, whether confidence regions have the claimed coverage probability, or whether the algorithms recover the true Pareto set. The 'demonstrated reliability' stated in Section 4 is therefore not supported. Add end-to-end experiments on synthetic problems with known Pareto sets, verification of the confidence-region comparisons against brute-force oracles, and, where relevant, results on the provided real-world datasets.
  4. [Section 3] The paper announces a 'novel entropy-based decoupled algorithm' but gives no description, no pseudocode, and no evaluation. If this algorithm is intended as a substantive contribution, it needs a full technical treatment; if it is only an example of the framework's flexibility, that status should be stated explicitly. As written, the contribution list in the introduction and the mention in Section 3 promise more than the manuscript delivers.
minor comments (5)
  1. [Section 3] The text mentions 'polyhedral approximations of 3D ice-cream cones' without quantifying the approximation error or explaining whether the algorithms' correctness guarantees apply to the approximate cone or to the true cone; please clarify.
  2. [Section 2.1] The 'Performance notebook' is referenced without a URL or repository path; if it exists, provide a direct link and a commit hash or version tag for reproducibility.
  3. [Section 2.1] The phrase 'has been checked for vulnerabilities with the tool Bandit and is secure' overstates what a static-analysis tool can establish; rephrase to 'has been scanned with Bandit' or otherwise qualify the security claim.
  4. [Introduction] The claim that VOPy is the 'first library' for black-box vector optimization would be easier to evaluate if the authors stated the search criteria or the specific features that exclude existing MOO libraries beyond the cone-order distinction.
  5. [Abstract] The abstract lists 'batch observations' as a supported setting, but the paper does not explain which built-in algorithms or acquisition functions support batch evaluation; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: VOPy is a software framework paper whose algorithmic content reimplements independently published methods; no prediction or derivation reduces to its own inputs.

full rationale

This is a software/engineering paper, not a derivation paper. The central claims are that VOPy is the first library for black-box vector optimization, that it provides modular interfaces for orders/models/algorithms/problems, and that it solves confidence-region comparison convex programs via CVXPY. None of these claims is obtained by fitting a parameter to data and then predicting that same data; there are no fitted values, no empirical prediction claims, and no equations whose equivalence could be exhibited. The self-citations to the authors' prior work (Ararat and Tekin 2023; Karagozlu et al. 2024; Korkmaz et al. 2024) are used as provenance for the implemented algorithms, which are published, peer-reviewed results with independent support; the paper does not invoke those citations as the sole justification for a uniqueness theorem or as a way to smuggle in an ansatz. The statement in Section 4 that 'Its early stages contributed to the development of algorithms now existing in the literature' is a historical note, not a load-bearing reduction. The main unresolved issue is engineering verification: the confidence-region comparison solvers are described only as 'supporting solutions to these comparisons using the CVXPY library' without mathematical formulations or runtime benchmarks, and a 'Performance notebook' is referenced but not included. That is a correctness or completeness concern, not circularity. Accordingly, the appropriate circularity score is 0.

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

No new physical or mathematical entities are introduced. The central claim rests on software engineering assumptions and the correctness of prior algorithms, not on fitted parameters.

assumptions (4)
  • domain assumption The comparison of confidence regions with respect to a cone-induced partial order can be reduced to convex optimization problems that CVXPY solves correctly.
    Section 2.1 states that confidence region comparisons are solved using CVXPY, but no formulation or certificates are provided.
  • domain assumption The implemented algorithms from prior work (PaVeBa, VOGP, epsilon-PAL) are correct as originally published.
    Section 3 says VOPy implements these algorithms; the paper does not re-derive their guarantees.
  • domain assumption A test coverage above 95 percent implies the library is reliable.
    Section 2.2 claims high coverage but provides no test results or failure cases.
  • standard math A closed convex cone C with C intersect -C = {0} defines a valid partial order.
    Section 1 introduces this standard definition from vector optimization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VOPy: A Framework for Black-box Vector Optimization." pith.science (2026). https://pith.science/paper/YAJH26TW

@misc{pith2026241206604,
  author       = {Pith},
  title        = {Pith review of: VOPy: A Framework for Black-box Vector Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YAJH26TW}},
  note         = {Machine review of arXiv:2412.06604}
}
read the original abstract

We introduce VOPy, an open-source Python library designed to address black-box vector optimization, where multiple objectives must be optimized simultaneously with respect to a partial order induced by a convex cone. VOPy extends beyond traditional multi-objective optimization (MOO) tools by enabling flexible, cone-based ordering of solutions; with an application scope that includes environments with observation noise, discrete or continuous design spaces, limited budgets, and batch observations. VOPy provides a modular architecture, facilitating the integration of existing methods and the development of novel algorithms. We detail VOPy's architecture, usage, and potential to advance research and application in the field of vector optimization. The source code for VOPy is available at https://github.com/Bilkent-CYBORG/VOPy.

Figures

Figures reproduced from arXiv: 2412.06604 by the authors.

Figure 1
Figure 1. Overview of the dependencies, core modules, and built-in algorithms of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 25 canonical work pages

  1. [1]

    Vector optimization with stochastic bandit feedback

    C a g n Ararat and Cem Tekin. Vector optimization with stochastic bandit feedback. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 2165--2190, 2023

  2. [2]

    Pareto front identification from stochastic bandit feedback

    Peter Auer, Chao-Kai Chiang, Ronald Ortner, and Madalina Drugan. Pareto front identification from stochastic bandit feedback. In Proceedings of The 19th International Conference on Artificial Intelligence and Statistics, pages 939--947, 2016

  3. [3]

    Jiang, Samuel Daulton, Benjamin Letham, Andrew Gordon Wilson, and Eytan Bakshy

    Maximilian Balandat, Brian Karrer, Daniel R. Jiang, Samuel Daulton, Benjamin Letham, Andrew Gordon Wilson, and Eytan Bakshy. B o T orch: a framework for efficient M onte- C arlo B ayesian optimization. In Advances in Neural Information Processing Systems 33 (NIPS 2020), pages 21524--21538, 2020

  4. [4]

    Pymoo: Multi-objective optimization in P ython

    Julian Blank and Kalyanmoy Deb. Pymoo: Multi-objective optimization in P ython. I E E E Access , 8: 0 89497--89509, 2020

  5. [5]

    R B F O pt: an open-source library for black-box optimization with costly function evaluations

    Alberto Costa and Giacomo Nannicini. R B F O pt: an open-source library for black-box optimization with costly function evaluations. Mathematical Programming Computation, 10: 0 597--629, 2018

  6. [6]

    A tutorial on multiobjective optimization: fundamentals and evolutionary methods

    Michael TM Emmerich and Andr \'e H Deutz. A tutorial on multiobjective optimization: fundamentals and evolutionary methods. Natural Computing, 17: 0 585--609, 2018

  7. [7]

    Google V izier: a service for black-box optimization

    Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and David Sculley. Google V izier: a service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1487--1495, 2017

  8. [8]

    Bayesian optimization for synthetic gene design

    Javier Gonzalez, Joseph Longworth, David C James, and Neil D Lawrence. Bayesian optimization for synthetic gene design. arXiv preprint 1505.01627, 2015

Show all 26 references
  1. [9]

    Portfolio allocation for B ayesian optimization

    Matthew Hoffman, Eric Brochu, Nando De Freitas, et al. Portfolio allocation for B ayesian optimization. In Proceedings of the 27th Conference on Uncertainty in Artificial Intelligence, pages 327--336, 2011

  2. [10]

    Bias free multiobjective active learning for materials design and discovery

    Kevin Maik Jablonka, Giriprasad Melpatti Jothiappan, Shefang Wang, Berend Smit, and Brian Yoo. Bias free multiobjective active learning for materials design and discovery. Nature Communications, 12 0 (1): 0 2312, 2021

  3. [11]

    Openbox: a P ython toolkit for generalized black-box optimization

    Huaijun Jiang, Yu Shen, Yang Li, Beicheng Xu, Sixian Du, Wentao Zhang, Ce Zhang, and Bin Cui. Openbox: a P ython toolkit for generalized black-box optimization. Journal of Machine Learning Research, 25 0 (120): 0 1--11, 2024

  4. [12]

    Kirthevasan Kandasamy, Willie Neiswanger, Jeff Schneider, Barnab\' a s P\' o czos, and Eric P. Xing. Neural architecture search with B ayesian optimisation and optimal transport. In Advances in Neural Information Processing Systems 31 (NIPS 2018), pages 2020--2029, 2018

  5. [13]

    Learning the P areto set under incomplete preferences: Pure exploration in vector bandits

    Efe Mert Karag \"o zl \"u , Ya s ar Cahit Y ld r m, C a g n Ararat, and Cem Tekin. Learning the P areto set under incomplete preferences: Pure exploration in vector bandits. In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, pages 30...

  6. [14]

    Vector optimization with G aussian process bandits

    İlter Onat Korkmaz, Yaşar Cahit Yıldırım, Çağın Ararat, and Cem Tekin. Vector optimization with G aussian process bandits. arXiv preprint 2412.02484, 2024

  7. [15]

    Multiobjective optimization for crash safety design of vehicles using stepwise regression model

    Xingtao Liao, Qing Li, Xujing Yang, Weigang Zhang, and Wei Li. Multiobjective optimization for crash safety design of vehicles using stepwise regression model. Structural and Multidisciplinary Optimization, 35: 0 561--569, 2008

  8. [16]

    Crystal structure prediction using ab initio evolutionary techniques: principles and applications

    Artem R Oganov and Colin W Glass. Crystal structure prediction using ab initio evolutionary techniques: principles and applications. The Journal of Chemical Physics, 124 0 (24), 2006

  9. [17]

    Prediction of low-thermal-conductivity compounds with first-principles anharmonic lattice-dynamics calculations and B ayesian optimization

    Atsuto Seko, Atsushi Togo, Hiroyuki Hayashi, Koji Tsuda, Laurent Chaput, and Isao Tanaka. Prediction of low-thermal-conductivity compounds with first-principles anharmonic lattice-dynamics calculations and B ayesian optimization. Physical Review Letters, 115 0 (20): 0 205901, 2015

  10. [18]

    Practical B ayesian optimization of machine learning algorithms

    Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical B ayesian optimization of machine learning algorithms. Advances in Neural Information Processing Systems 25 (NIPS 2012), pages 2951--2959, 2012

  11. [19]

    An easy-to-use real-world multi-objective optimization problem suite

    Ryoji Tanabe and Hisao Ishibuchi. An easy-to-use real-world multi-objective optimization problem suite. Applied Soft Computing, 89: 0 106078, 2020

  12. [20]

    Black-box optimization for automated discovery

    Kei Terayama, Masato Sumita, Ryo Tamura, and Koji Tsuda. Black-box optimization for automated discovery. Accounts of Chemical Research, 54 0 (6): 0 1334--1346, 2021

  13. [21]

    Joint entropy search for multi-objective B ayesian optimization

    Ben Tu, Axel Gandy, Nikolas Kantas, and Behrang Shafei. Joint entropy search for multi-objective B ayesian optimization. In Advances in Neural Information Processing Systems 35 (NIPS 2022), pages 9922--9938, 2022

  14. [22]

    C O M B O : An efficient B ayesian optimization library for materials science

    Tsuyoshi Ueno, Trevor David Rhone, Zhufeng Hou, Teruyasu Mizoguchi, and Koji Tsuda. C O M B O : An efficient B ayesian optimization library for materials science. Materials Discovery, 4: 0 18--21, 2016

  15. [23]

    Style guide for Python code

    Guido van Rossum, Barry Warsaw, and Nick Coghlan. Style guide for Python code. PEP 8, Python Software Foundation, 2001. URL https://www.python.org/dev/peps/pep-0008/

  16. [24]

    Hyperparameter optimization for machine learning models based on B ayesian optimization

    Jia Wu, Xiu-Yun Chen, Hao Zhang, Li-Dong Xiong, Hang Lei, and Si-Hao Deng. Hyperparameter optimization for machine learning models based on B ayesian optimization. Journal of Electronic Science and Technology, 17 0 (1): 0 26--40, 2019

  17. [25]

    Computer generation of streaming sorting networks

    Marcela Zuluaga, Peter Milder, and Markus P \"u schel. Computer generation of streaming sorting networks. In Proceedings of the 49th Annual Design Automation Conference, pages 1245--1253, 2012

  18. [26]

    e- P A L : An active learning approach to the multi-objective optimization problem

    Marcela Zuluaga, Andreas Krause, and Markus P\" u schel. e- P A L : An active learning approach to the multi-objective optimization problem. Journal of Machine Learning Research, 17 0 (104): 0 1--32, 2016

Pith tools

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