Pith. sign in

REVIEW 1 major objections 5 minor 50 references

Multi-Objective Covariance Matrix Adaptation MAP-Annealing

T0 review · 1 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A new multi-objective quality-diversity algorithm, MO-CMA-MAE, uses covariance-matrix-adaptation search and per-cell threshold fronts to beat four established baselines on two benchmark domains and match the strongest baseline on two…

desk verdict Solid empirical MOQD paper with a genuine algorithmic novelty, but the threshold-front acceptance rule can admit dominated solutions into the archive, so the stated invariant and Pareto-compliance justification don't hold as written. read the letter →

arxiv 2505.20712 v1 pith:C3NBGM4E submitted 2025-05-27 cs.NE

Shihan Zhao , Stefanos Nikolaidis This is my paper
classification cs.NE
keywords multi-objectivequality-diversityCMA-EShypervolumeimprovementthresholdacceptingMAP-Elitesbehaviordiversitygamelevelgenerationevolutionaryoptimization
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 show that adaptive covariance-matrix search, which transformed single-objective quality-diversity optimization, can also solve multi-objective quality-diversity (MOQD) problems better than the mutation-and-crossover methods previously used. It proposes MO-CMA-MAE, which drives CMA-ES with a per-solution reward: the hypervolume improvement a candidate contributes to its behavior cell's archive front. To keep the search exploring under-visited behaviors, each cell maintains a lagging 'threshold front' built from discounted objectives, so early visitors to a cell earn inflated rewards that fade as the cell matures. On sphere and rastrigin benchmarks the method significantly outperforms MOME, NSGA-II, SMS-EMOA, and COMO-CMA-ES, and on arm and Overcooked map generation it matches MOME while beating the rest. If true, the result means the CMA-ES acceleration that reshaped single-objective QD transfers cleanly to the multi-objective setting.

What carries the argument

The central object is the threshold front T_e per archive cell, maintained alongside the cell's Pareto set P_e. When a candidate x maps to cell e, its reward is the hypervolume improvement Phi_i = HVI_r(f(x_i), T_e) over that threshold front; only candidates with Phi_i > 0 enter the Pareto set. After insertion, a discounted copy d_i f(x_i) goes into T_e, with d_i chosen by bisection so that HVI_r(d_i f(x_i), T_e) approx alpha * HVI_r(f(x_i), T_e), closing a fixed fraction alpha of the gap. CMA-ES emitters rank candidates by Phi_i and adapt their means and covariances; crowding-distance downsizing keeps each cell within a memory budget, and a cycle-restart rule escapes repeated insert-and-drop loops.

What would settle it

Re-run the sphere benchmark and, after 5000 iterations, inspect every occupied archive cell for pairs of mutually dominated solutions within the same cell's Pareto set, or for solutions dominated by the cell's threshold front; if a substantial fraction of cells contain such solutions, the acceptance rule has admitted degraded solutions, contradicting the claim that HVI > 0 preserves Pareto-set geometry.

Watch

Extended reading notes

Core claim

MO-CMA-MAE is a Multi-Objective Quality-Diversity algorithm that uses CMA-ES to maximize, for every archive cell, the hypervolume improvement of newly sampled solutions over a per-cell threshold front. The central move is to compute the CMA-ES reward as HVI_r(f(x), T_e) instead of over the cell's incumbent Pareto front, and to maintain T_e by inserting discounted objectives d_i f(x_i) whose discount factors are found by bisection so that each insertion closes a fixed fraction alpha of the current gap between threshold and real front. This 'threshold accepting' mechanism, inherited from single-objective CMA-MAE, boosts rewards in under-explored cells and anneals them away as cells fill. The result is a search that tries both to discover new behaviors and to expand the local Pareto front of discovered behaviors, and the paper's experiments report that this beats all baselines on two of four domains and matches the strongest baseline on the other two.

Load-bearing premise

The load-bearing premise is that the discounted threshold front T_e stays a faithful, lagging image of each cell's true Pareto front, so that positive hypervolume improvement consistently means this solution genuinely advances the cell, and the boosted rewards in young cells reliably point the search toward under-explored behaviors.

Editorial extensions

If this is right

  • MO-CMA-MAE can be plugged into any MOQD setting that exposes a measure function and multiple objectives, replacing mutation and crossover emitters with covariance-adaptive ones.
  • The threshold-front construction gives a principled scalar reward for multi-objective search, so other indicator-based multi-objective evolutionary algorithms could adopt the same annealing trick.
  • Because hypervolume is Pareto-compliant, using HVI as the CMA-ES reward prevents deteriorative cycles that would otherwise let the search rediscover a dominated front.
  • The Overcooked results suggest the method is practical for generating diverse game levels with multiple design objectives, such as difficulty and construction cost.

Reading between the lines

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

  • (Editorial inference) The bisection-based discount search makes each threshold front track its real front locally per trade-off, but a simpler fixed discount factor per cell might retain most exploration benefits with much lower computational cost.
  • (Editorial inference) Because exact hypervolume computation is the stated bottleneck, the method's scalability to many objectives hinges on approximate hypervolume estimators; if those preserve the order of HVI values, MO-CMA-MAE could extend well beyond the two- and three-objective regimes tested here.
  • (Editorial inference) The same threshold-front idea could be applied to other quality indicators, such as sparsity or the epsilon-indicator, to convert any single-quality QD archive into a multi-objective one without changing the CMA-ES machinery.
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

1 major / 5 minor

Summary. The paper introduces MO-CMA-MAE, a multi-objective Quality-Diversity (MOQD) algorithm that uses CMA-ES to search for solutions with high hypervolume improvement relative to a per-cell threshold front. The threshold front is constructed by inserting discounted objectives so that the hypervolume improvement of a candidate is boosted in under-explored cells. The algorithm is benchmarked on sphere, rastrigin, arm, and overcooked domains against MOME, NSGA-II, SMS-EMOA, and a modified COMO-CMA-ES. The authors report that MO-CMA-MAE achieves significantly higher MOQD-scores than all baselines on sphere and rastrigin, and is on par with MOME while outperforming the other baselines on arm and overcooked.

Significance. If the results hold, this is a valuable contribution to MOQD: it is the first algorithm to bring adaptive covariance-based search (CMA-ES) to multi-objective MAP-Elites, and the threshold-front mechanism is a plausible way to balance exploration and exploitation in measure space. The empirical evaluation uses a reasonable protocol (ANOVA and Tukey HSD tests) and the improvements on two domains are large. The paper does not release code, which limits reproducibility, and the central algorithmic justification is undermined by a correctness issue described below. The idea is novel and likely to interest the GECCO community.

major comments (1)
  1. [Algorithm 1, lines 10-14 and result comment] The insertion rule accepts a candidate whenever Phi_i = HVI_r(f(x_i), T_e) > 0 and then adds it to P_e, removing only solutions dominated by x_i. Because the threshold front T_e lags behind the true local front P_e, a candidate can be non-dominated by T_e while being dominated by an existing member of P_e. In that case line 12 retains the dominating incumbent and also inserts x_i, so P_e contains mutually dominated solutions. This directly contradicts the stated result invariant that every P_e is a Pareto set, and it invalidates the Pareto-compliance argument in Section 4.1, which applies to HVI relative to P_e, not relative to the lagging T_e. The central claim that MO-CMA-MAE optimizes the hypervolume of every Pareto set (Eq. 4) is therefore not established by construction. Please correct the acceptance rule to require that x_i is not dominated by P_e (in addition to having Phi_i > 0), and then re-run the experiments or explicitly justify why the current behavior does not affect the reported MOQD-scores.
minor comments (5)
  1. [Section 4.2] The statement that HVI_r(d f(x), T_e) is 'strictly monotonic' over d in [0,1] is too strong: if d f(x) is dominated by T_e for small d, the function is zero on an interval, so it is non-decreasing but not strictly increasing. The bisection search still works for a positive target because HVI is continuous and goes from 0 to a positive value, but the wording should be corrected.
  2. [Abstract and Section 5] The abstract says the method is tested on 'three MOQD domains, and for generating maps of a co-operative video game,' but Section 5 lists four domains (sphere, rastrigin, arm, overcooked). Please clarify whether the video game is counted as a domain or as a separate application.
  3. [General] No code repository or link is provided, despite the paper being an empirical algorithms paper. Releasing the implementation would greatly aid reproducibility and allow reviewers to verify that the reported behavior matches Algorithm 1.
  4. [Section 5.3] The sensitivity analysis for the learning rate alpha is performed only on the sphere domain (Figure 6). Since alpha=0.1 is used for all other domains, providing at least one additional sensitivity plot (e.g., rastrigin) would strengthen the claim that the algorithm is robust to this hyperparameter across domains.
  5. [Algorithm 1 and Section 4.3] The cyclerestart rule uses the constant '10 times the average number of visits' in the text, but this threshold is not listed among the hyperparameters in Appendix A. Please state it explicitly as a hyperparameter and report its value.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the algorithm is an empirical composition of HVI-based CMA-ES and threshold acceptance; self-citations to CMA-MAE are not load-bearing. The threshold-front acceptance rule creates an internal-validity gap with the stated MOQD-score derivation, but that is not a circular reduction.

full rationale

The paper's central claims are empirical: MO-CMA-MAE maximizes hypervolume improvement relative to per-cell threshold fronts, and Table 1 reports MOQD-scores against MOME, NSGA-II, SMS-EMOA, and COMO-CMA-ES with ANOVA and Tukey HSD tests (Section 5.2, Appendix F). The MOQD-score of Eq. 4 is a fixed evaluation metric applied identically to all algorithms, not a quantity fitted by the algorithm; no benchmark result is derived from assumptions that include that result. The threshold-front mechanism is inspired by the authors' CMA-MAE line [15, 48], but the load-bearing claim that threshold acceptance improves exploration is independently supported within this paper by the alpha ablation in Section 5.3 (alpha=1 disables the threshold and performs poorly, Figure 6). Thus the self-citations are ancestry, not the sole support for the conclusion. I found no step in which a prediction reduces by construction to a fitted parameter or to a definition. The closest concern is an internal-validity gap rather than circularity: Algorithm 1 (lines 10-14) inserts a candidate into the cell's Pareto set whenever HVI_r(f(x_i), T_e) > 0, where T_e is built from discounted objectives d_i f(x_i) with d_i <= 1, so positive HVI over the lagging threshold front does not by Eqs. 3-4 imply an increase in the cell's true hypervolume HV(F_e); the Pareto-compliance guarantee cited from [3, 37] applies to HVI relative to the true front, not T_e. This weakens the derivation in Section 4.1 that HVI-based CMA-ES 'thereby optimizes the MOQD score,' but it does not make the empirical comparison circular, and the dynamic-archive ablation (Appendix E.3) and the alpha sweep provide external checks on the mechanism. Section 6 honestly notes the hypervolume computation bottleneck; this is a scalability limitation, not circularity. Overall score 1 reflects only minor self-citation ancestry.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The algorithm introduces no new physical or conceptual entities; threshold fronts and discounted objectives are algorithmic constructs within the method. The central claim rests on standard definitions plus several domain-specific and unproved mathematical assumptions, most notably the monotonicity of discounted HVI used by the bisection search.

free parameters (5)
  • alpha (learning rate) = 0.1
    Controls the fraction of the HVI gap closed per threshold-front insertion. Selected in Section 5.1, with sensitivity analyzed only on the sphere domain in Section 5.3, and central to the threshold mechanism.
  • epsilon (bisection error tolerance) = 1e-3
    Termination tolerance for the discount-factor bisection search; set in Section 5.1 and used in Algorithm 2.
  • sigma0 (initial step size) = 0.5
    Initial CMA-ES step size for MO-CMA-MAE and COMO-CMA-ES; listed in Appendix A for all domains.
  • reference point r for hypervolume = 0 after min-max normalization to [0,100]
    The hypervolume reference point is set by per-domain normalization in Appendix D; the choice affects all HVI rewards and therefore the search behavior.
  • Archive configuration = 1000 cells, max 10 solutions per cell
    Chosen in Section 5.1 to match the 10000-solution memory budget across baselines; interacts with threshold-front retraction and the cycle restart rule.
assumptions (5)
  • standard math Hypervolume indicator and Pareto dominance as defined in Equations 1-3.
    Used to define Pareto set quality and the improvement signal for CMA-ES; these are standard definitions from prior literature.
  • domain assumption Hypervolume is strictly Pareto-compliant, so optimizing it cannot deteriorate a Pareto front.
    Invoked in Section 4.1 to justify using HVI as the CMA-ES objective, relying on cited results [3,37].
  • ad hoc to paper HVI_r(d f(x), T_e) is strictly monotonic in the discount factor d over [0,1], enabling bisection search.
    Claimed by observation in Section 4.2 and used by Algorithm 2; no proof is provided, and this monotonicity is central to the threshold-front construction.
  • domain assumption CMA-ES can effectively adapt its sampling distribution from rankings of scalar HVI values.
    Assumed throughout Section 4 and Algorithm 1; this is standard practice in CMA-ME and CMA-MAE but an empirical assumption in the multi-objective archive setting.
  • domain assumption Overcooked difficulty is adequately estimated by two fixed agents and the reward model, with measures aggregated over 4 trials.
    Defined in Appendix D; the validity of the difficulty proxy is inherited from prior work and not independently assessed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Objective Covariance Matrix Adaptation MAP-Annealing." pith.science (2026). https://pith.science/paper/C3NBGM4E

@misc{pith2026250520712,
  author       = {Pith},
  title        = {Pith review of: Multi-Objective Covariance Matrix Adaptation MAP-Annealing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C3NBGM4E}},
  note         = {Machine review of arXiv:2505.20712}
}
read the original abstract

Quality-Diversity (QD) optimization is an emerging field that focuses on finding a set of behaviorally diverse and high-quality solutions. While the quality is typically defined w.r.t. a single objective function, recent work on Multi-Objective Quality-Diversity (MOQD) extends QD optimization to simultaneously optimize multiple objective functions. This opens up multi-objective applications for QD, such as generating a diverse set of game maps that maximize difficulty, realism, or other properties. Existing MOQD algorithms use non-adaptive methods such as mutation and crossover to search for non-dominated solutions and construct an archive of Pareto Sets (PS). However, recent work in QD has demonstrated enhanced performance through the use of covariance-based evolution strategies for adaptive solution search. We propose bringing this insight into the MOQD problem, and introduce MO-CMA-MAE, a new MOQD algorithm that leverages Covariance Matrix Adaptation-Evolution Strategies (CMA-ES) to optimize the hypervolume associated with every PS within the archive. We test MO-CMA-MAE on three MOQD domains, and for generating maps of a co-operative video game, showing significant improvements in performance.

Figures

Figures reproduced from arXiv: 2505.20712 by the authors.

Figure 1
Figure 1. Game maps found by MO-CMA-MAE that induce diverse task divisions while achieving good trade-off between the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The gap between 𝑇𝑒 and 𝐹𝑒 boosts HVI for a new solution, and should be narrower at well-explored trade-offs. Discount Factor Search. When searching the discount factor 𝑑𝑖 for a solution 𝒙𝑖 assigned to cell 𝑒, our goal is to ensure that in￾serting 𝑑𝑖f(𝒙𝑖) to 𝑇𝑒 closes its gap to 𝐹𝑒 by a fraction. To quantify this goal, we use the hypervolume improvement of the discounted objectives over the pre-insertion threshold fr… view at source ↗
Figure 3
Figure 3. When inserting a new solution with objectives [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: MOQD-scores and Coverages achieved by tested algorithms on each domain. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Heatmaps representing the passive MOQD archive populated by each algorithm after 5000 iterations of optimization [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Sphere domain archive coverages at different [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Frequency of CMA-ES samples from each cell at [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: An example of how the threshold front approximates the real front in the static archive version of MO-CMA-MAE [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: An example of how the threshold front approximates the real front in the dynamic archive version of MO-CMA-MAE. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 11
Figure 11. Figure 11: MOQD-scores and archive coverages achieved by [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: The incumbent local PFs from all cells found by [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Heatmaps representing the passive MOQD archives for the [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [1]

    Hype: An algorithm for fast hypervolume- based many-objective optimization.Evolutionary computation, 19(1):45–76, 2011

    Johannes Bader and Eckart Zitzler. Hype: An algorithm for fast hypervolume- based many-objective optimization.Evolutionary computation, 19(1):45–76, 2011

    work page 2011

  2. [2]

    Proximal Policy Gradient Arborescence for Quality Diversity Reinforcement Learning

    Sumeet Batra, Bryon Tjanaka, Matthew C Fontaine, Aleksei Petrenko, Stefanos Nikolaidis, and Gaurav Sukhatme. Proximal policy gradient arborescence for quality diversity reinforcement learning.arXiv preprint arXiv:2305.13795, 2023

    work page Pith review arXiv 2023

  3. [3]

    Set-based multi- objective optimization, indicators, and deteriorative cycles

    Rudolf Berghammer, Tobias Friedrich, and Frank Neumann. Set-based multi- objective optimization, indicators, and deteriorative cycles. InProceedings of the 12th annual conference on Genetic and evolutionary computation, pages 495–502, 2010

    work page 2010

  4. [4]

    Sms-emoa: Multiobjective selection based on dominated hypervolume.European Journal of Operational Research, 181(3):1653–1669, 2007

    Nicola Beume, Boris Naujoks, and Michael Emmerich. Sms-emoa: Multiobjective selection based on dominated hypervolume.European Journal of Operational Research, 181(3):1653–1669, 2007

    work page 2007

  5. [5]

    Deep surrogate assisted generation of environments.Advances in Neural Information Processing Systems, 35:37762–37777, 2022

    Varun Bhatt, Bryon Tjanaka, Matthew Fontaine, and Stefanos Nikolaidis. Deep surrogate assisted generation of environments.Advances in Neural Information Processing Systems, 35:37762–37777, 2022

    work page 2022

  6. [6]

    Blank and K

    J. Blank and K. Deb. pymoo: Multi-objective optimization in python.IEEE Access, 8:89497–89509, 2020

    2020

  7. [7]

    Two-dimensional subset selection for hypervolume and epsilon-indicator

    Karl Bringmann, Tobias Friedrich, and Patrick Klitzke. Two-dimensional subset selection for hypervolume and epsilon-indicator. InProceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation, pages 589–596, 2014

    work page 2014

  8. [8]

    On the utility of learning about humans for human-ai coordi- nation.Advances in neural information processing systems, 32, 2019

    Micah Carroll, Rohin Shah, Mark K Ho, Tom Griffiths, Sanjit Seshia, Pieter Abbeel, and Anca Dragan. On the utility of learning about humans for human-ai coordi- nation.Advances in neural information processing systems, 32, 2019

    work page 2019

Show all 50 references
  1. [9]

    Qdax: A library for quality-diversity and population-based algorithms with hardware acceleration, 2023

    Felix Chalumeau, Bryan Lim, Raphael Boige, Maxime Allard, Luca Grillotti, Manon Flageat, Valentin Macé, Arthur Flajolet, Thomas Pierrot, and Antoine Cully. Qdax: A library for quality-diversity and population-based algorithms with hardware acceleration, 2023

    2023

  2. [10]

    Quality and diversity optimization: A unify- ing modular framework.IEEE Transactions on Evolutionary Computation, 22(2): 245–259, 2017

    Antoine Cully and Yiannis Demiris. Quality and diversity optimization: A unify- ing modular framework.IEEE Transactions on Evolutionary Computation, 22(2): 245–259, 2017

    2017

  3. [11]

    Robots that can adapt like animals.Nature, 521(7553):503–507, 2015

    Antoine Cully, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret. Robots that can adapt like animals.Nature, 521(7553):503–507, 2015

    2015

  4. [12]

    A fast and elitist multiobjective genetic algorithm: Nsga-ii.IEEE transactions on evolutionary computation, 6(2):182–197, 2002

    Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii.IEEE transactions on evolutionary computation, 6(2):182–197, 2002

    2002

  5. [13]

    A ranking method based on the r2 indicator for many- objective optimization

    Alan Díaz-Manríquez, Gregorio Toscano-Pulido, Carlos A Coello Coello, and Ricardo Landa-Becerra. A ranking method based on the r2 indicator for many- objective optimization. In2013 IEEE Congress on Evolutionary Computation, pages 1523–1530. IEEE, 2013

    2013

  6. [14]

    Quality diversity through human feedback.arXiv preprint arXiv:2310.12103, 2023

    Li Ding, Jenny Zhang, Jeff Clune, Lee Spector, and Joel Lehman. Quality diversity through human feedback.arXiv preprint arXiv:2310.12103, 2023

    2023 arXiv

  7. [15]

    Covariance matrix adaptation map- annealing

    Matthew Fontaine and Stefanos Nikolaidis. Covariance matrix adaptation map- annealing. InProceedings of the Genetic and Evolutionary Computation Conference, pages 456–465, 2023

    2023

  8. [16]

    Evaluating human–robot interac- tion algorithms in shared autonomy via quality diversity scenario generation

    Matthew C Fontaine and Stefanos Nikolaidis. Evaluating human–robot interac- tion algorithms in shared autonomy via quality diversity scenario generation. ACM Transactions on Human-Robot Interaction (THRI), 11(3):1–30, 2022

    2022

  9. [17]

    Covariance matrix adaptation for the rapid illumination of behavior space

    Matthew C Fontaine, Julian Togelius, Stefanos Nikolaidis, and Amy K Hoover. Covariance matrix adaptation for the rapid illumination of behavior space. In Proceedings of the 2020 genetic and evolutionary computation conference, pages 94–102, 2020

    2020

  10. [18]

    On the importance of environments in human-robot coordination

    Matthew C Fontaine, Ya-Chuan Hsu, Yulun Zhang, Bryon Tjanaka, and Stefanos Nikolaidis. On the importance of environments in human-robot coordination. arXiv preprint arXiv:2106.10853, 2021

    2021 arXiv

  11. [19]

    Brax–a differentiable physics engine for large scale rigid body simulation.arXiv preprint arXiv:2106.13281, 2021

    C Daniel Freeman, Erik Frey, Anton Raichuk, Sertan Girgin, Igor Mordatch, and Olivier Bachem. Brax–a differentiable physics engine for large scale rigid body simulation.arXiv preprint arXiv:2106.13281, 2021

    2021 arXiv

  12. [20]

    Preference-learning emit- ters for mixed-initiative quality-diversity algorithms.IEEE Transactions on Games, 2023

    Roberto Gallotta, Kai Arulkumaran, and Lisa B Soros. Preference-learning emit- ters for mixed-initiative quality-diversity algorithms.IEEE Transactions on Games, 2023

    2023

  13. [21]

    Overcooked, 2016

    GhostTownGames. Overcooked, 2016. URL https://store.steampowered.com/ app/448510/Overcooked/

    2016

  14. [22]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

    2014

  15. [23]

    Unsupervised behavior discovery with quality- diversity optimization.IEEE Transactions on Evolutionary Computation, 26(6): 1539–1552, 2022

    Luca Grillotti and Antoine Cully. Unsupervised behavior discovery with quality- diversity optimization.IEEE Transactions on Evolutionary Computation, 26(6): 1539–1552, 2022

    2022

  16. [24]

    The cma evolution strategy: A tutorial.arXiv preprint arXiv:1604.00772, 2016

    Nikolaus Hansen. The cma evolution strategy: A tutorial.arXiv preprint arXiv:1604.00772, 2016

    2016 arXiv

  17. [25]

    Moarchiving

    Nikolaus Hansen. Moarchiving. https://github.com/CMA-ES/moarchiving/tree/ master, 2020

    2020

  18. [26]

    Coco: A platform for comparing continuous optimizers in a black-box setting.Optimization Methods and Software, 36(1):114–144, 2021

    Nikolaus Hansen, Anne Auger, Raymond Ros, Olaf Mersmann, Tea Tušar, and Dimo Brockhoff. Coco: A platform for comparing continuous optimizers in a black-box setting.Optimization Methods and Software, 36(1):114–144, 2021

    2021

  19. [27]

    Covariance matrix adaptation for multi-objective optimization.Evolutionary computation, 15(1):1–28, 2007

    Christian Igel, Nikolaus Hansen, and Stefan Roth. Covariance matrix adaptation for multi-objective optimization.Evolutionary computation, 15(1):1–28, 2007

    2007

  20. [28]

    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. InEvolutionary Multi-Criterion Optimization: 8th International Conference, EMO 2015, Guimarães, Portugal, March 29–Apr...

    2015

  21. [29]

    Improving the data efficiency of multi-objective quality-diversity through gradient assistance and crowding exploration

    Hannah Janmohamed, Thomas Pierrot, and Antoine Cully. Improving the data efficiency of multi-objective quality-diversity through gradient assistance and crowding exploration. InProceedings of the Genetic and Evolutionary Computation Conference, pages 165–173, 2023

    2023

  22. [30]

    A simple and fast hypervolume indicator-based multiobjective evolutionary algorithm.IEEE Transactions on Cybernetics, 45(10):2202–2213, 2014

    Siwei Jiang, Jie Zhang, Yew-Soon Ong, Allan N Zhang, and Puay Siew Tan. A simple and fast hypervolume indicator-based multiobjective evolutionary algorithm.IEEE Transactions on Cybernetics, 45(10):2202–2213, 2014

    2014

  23. [31]

    Learning policies for partially observable environments: Scaling up

    Michael L Littman, Anthony R Cassandra, and Leslie Pack Kaelbling. Learning policies for partially observable environments: Scaling up. InMachine Learning Proceedings 1995, pages 362–370. Elsevier, 1995

    1995

  24. [32]

    Gd-moea: A new multi- objective evolutionary algorithm based on the generational distance indicator

    Adriana Menchaca-Mendez and Carlos A Coello Coello. Gd-moea: A new multi- objective evolutionary algorithm based on the generational distance indicator. InInternational conference on evolutionary multi-criterion optimization, pages 156–170. Springer, 2015

    2015

  25. [33]

    Illuminating search spaces by mapping elites.arXiv preprint arXiv:1504.04909, 2015

    Jean-Baptiste Mouret and Jeff Clune. Illuminating search spaces by mapping elites.arXiv preprint arXiv:1504.04909, 2015

    2015 arXiv

  26. [34]

    Policy gradient assisted map-elites

    Olle Nilsson and Antoine Cully. Policy gradient assisted map-elites. InProceedings of the Genetic and Evolutionary Computation Conference, pages 866–875, 2021

    2021

  27. [35]

    Multi- objective quality diversity optimization

    Thomas Pierrot, Guillaume Richard, Karim Beguir, and Antoine Cully. Multi- objective quality diversity optimization. InProceedings of the Genetic and Evolu- tionary Computation Conference, pages 139–147, 2022

    2022

  28. [36]

    On the evolutionary optimization of many conflicting objectives.IEEE transactions on evolutionary computation, 11(6): 770–784, 2007

    Robin C Purshouse and Peter J Fleming. On the evolutionary optimization of many conflicting objectives.IEEE transactions on evolutionary computation, 11(6): 770–784, 2007

    2007

  29. [37]

    A survey on the hyper- volume indicator in evolutionary multiobjective optimization.IEEE Transactions on Evolutionary Computation, 25(1):1–20, 2020

    Ke Shang, Hisao Ishibuchi, Linjun He, and Lie Meng Pang. A survey on the hyper- volume indicator in evolutionary multiobjective optimization.IEEE Transactions on Evolutionary Computation, 25(1):1–20, 2020

    2020

  30. [38]

    Hv-net: Hypervolume approximation based on deepsets.IEEE Transactions on Evolutionary Computation, 27(4):1154–1160, 2022

    Ke Shang, Weiyu Chen, Weiduo Liao, and Hisao Ishibuchi. Hv-net: Hypervolume approximation based on deepsets.IEEE Transactions on Evolutionary Computation, 27(4):1154–1160, 2022

    2022

  31. [39]

    An indicator- based multiobjective evolutionary algorithm with reference point adaptation for better versatility.IEEE Transactions on Evolutionary Computation, 22(4):609–622, 2017

    Ye Tian, Ran Cheng, Xingyi Zhang, Fan Cheng, and Yaochu Jin. An indicator- based multiobjective evolutionary algorithm with reference point adaptation for better versatility.IEEE Transactions on Evolutionary Computation, 22(4):609–622, 2017

    2017

  32. [40]

    Approximating gradients for differentiable quality diversity in reinforcement learning

    Bryon Tjanaka, Matthew C Fontaine, Julian Togelius, and Stefanos Nikolaidis. Approximating gradients for differentiable quality diversity in reinforcement learning. InProceedings of the Genetic and Evolutionary Computation Conference, pages 1102–1111, 2022

    2022

  33. [41]

    pyribs: A bare-bones python library for quality diversity optimiza- tion

    Bryon Tjanaka, Matthew C Fontaine, David H Lee, Yulun Zhang, Nivedit Reddy Balam, Nathaniel Dennler, Sujay S Garlanka, Nikitas Dimitri Klapsis, and Stefanos Nikolaidis. pyribs: A bare-bones python library for quality diversity optimiza- tion. InProceedings of the Genetic and E...

    2023

  34. [42]

    Uncrowded hypervolume improvement: Como-cma-es and the sofomore framework

    Cheikh Touré, Nikolaus Hansen, Anne Auger, and Dimo Brockhoff. Uncrowded hypervolume improvement: Como-cma-es and the sofomore framework. In Proceedings of the Genetic and Evolutionary Computation Conference, pages 638– 646, 2019

    2019

  35. [43]

    R2-emoa: Focused mul- tiobjective search using r2-indicator-based selection

    Heike Trautmann, Tobias Wagner, and Dimo Brockhoff. R2-emoa: Focused mul- tiobjective search using r2-indicator-based selection. InLearning and Intelligent Optimization: 7th International Conference, LION 7, Catania, Italy, January 7-11, 2013, Revised Selected Papers 7, pages ...

    2013

  36. [44]

    Discovering the elite hypervol- ume by leveraging interspecies correlation

    Vassiiis Vassiliades and Jean-Baptiste Mouret. Discovering the elite hypervol- ume by leveraging interspecies correlation. InProceedings of the Genetic and Evolutionary Computation Conference, pages 149–156, 2018

    2018

  37. [45]

    Scaling up map-elites using centroidal voronoi tessellations.arXiv preprint arXiv:1610.05729, 2016

    Vassilis Vassiliades, Konstantinos Chatzilygeroudis, and Jean-Baptiste Mouret. Scaling up map-elites using centroidal voronoi tessellations.arXiv preprint arXiv:1610.05729, 2016

    2016 arXiv

  38. [46]

    Recombination for learning strategy parameters in the mo-cma-es

    Thomas Voß, Nikolaus Hansen, and Christian Igel. Recombination for learning strategy parameters in the mo-cma-es. InInternational Conference on Evolutionary GECCO ’25, July 14–18, 2025, Malaga, Spain Shihan Zhao and Stefanos Nikolaidis Multi-Criterion Optimization, pages 155–1...

    2025

  39. [47]

    Improved step size adaptation for the mo-cma-es

    Thomas Voß, Nikolaus Hansen, and Christian Igel. Improved step size adaptation for the mo-cma-es. InProceedings of the 12th annual conference on Genetic and evolutionary computation, pages 487–494, 2010

    2010

  40. [48]

    Covariance matrix adaptation map-annealing: Theory and experiments.ACM Transactions on Evolutionary Learning, 2024

    Shihan Zhao, Bryon Tjanaka, Matthew C Fontaine, and Stefanos Nikolaidis. Covariance matrix adaptation map-annealing: Theory and experiments.ACM Transactions on Evolutionary Learning, 2024

    2024

  41. [49]

    Indicator-based selection in multiobjective search

    Eckart Zitzler and Simon Künzli. Indicator-based selection in multiobjective search. InInternational conference on parallel problem solving from nature, pages 832–842. Springer, 2004

    2004

  42. [50]

    skip over

    Eckart Zitzler and Lothar Thiele. Multiobjective evolutionary algorithms: a comparative case study and the strength pareto approach.IEEE transactions on Evolutionary Computation, 3(4):257–271, 1999. Multi-Objective Covariance Matrix Adaptation MAP-Annealing GECCO ’25, July 14–...

    1999

Pith tools

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