Pith. sign in

REVIEW 4 major objections 6 minor 24 references

Improving population size adapting CMA-ES algorithm on step-size blow-up in weakly-structured multimodal functions

T0 review · 4 major / 6 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 The paper claims that the step-size blow-up of PSA-CMA-ES on weakly structured multimodal functions is caused by its own step-size correction mechanism, and that gating and scaling that correction eliminates the blow-up and improves…

desk verdict Plausible heuristic fix for a real PSA-CMA-ES step-size blow-up, but the analytical theorems are invalid and the evidence is in-sample; treat it as a pilot study. read the letter →

arxiv 2506.00825 v1 pith:MRT4PFZO submitted 2025-06-01 cs.NE

Chandula Fernando , Kushani De Silva This is my paper
classification cs.NE
keywords CMA-ESpopulationsizeadaptationstep-sizeblow-upcorrectionmultimodaloptimizationweakly-structuredfunctionsRastriginfunctionSchaffer
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

Population size adapting CMA-ES (PSA-CMA-ES) uses a dynamic population size to balance exploring new basins and refining solutions within one basin. The paper claims that on weakly structured multimodal problems such as Rastrigin and Schaffer the algorithm's step-size blows up without bound, and that the blow-up is caused by the mechanism designed to correct the step-size after population size changes, not by the step-size adaptation itself. The authors prove that when the expected best candidates improve and the effective selection mass changes by only a small amount, the correction ratio is always at least one, so each generation inherits an amplified step-size. They then show that such amplification happens precisely when the population size change stays below a threshold $L$, reformulate the correction so that it is applied only when the optimizer is converging and scaled by a factor $\kappa=0.5$ when the population change is small, and report better final values and much lower CPU time than the original on two 2D benchmarks. The payoff of the claim is practical: a small, principled change to the correction step restores convergence on hard multimodal landscapes while retaining the exploration benefits of population size adaptation.

What carries the argument

The load-bearing object is the step-size correction ratio $\rho(\lambda_r^{(g+1)})/\rho(\lambda_r^{(g)})$ with $\rho(\lambda_r) = (-\sum_i w_i E[N_{i:\lambda_r}])\, n \mu_w / ( n - 1 + (-\sum_i w_i E[N_{i:\lambda_r}])^2 / \mu_w )$. Theorem 1 proves that this ratio is at least one under the paper's monotonicity assumptions, which turns the correction into a self-amplifying step-size multiplier. The supporting machinery is the gate used by the reformulation: the correction is applied only when the cumulative step-size path $\|p_\sigma\|$ is shorter than its expected length $\mathbb{E}[\|\mathcal{N}(0,I)\|]$, and it is multiplied by $\kappa=0.5$ when the population size change $|\lambda^{(g+1)}-\lambda^{(g)}|$ is below $L$ (experimentally $L=6$) and unscaled otherwise. The threshold $L$ is derived in Theorem 2 from a fitted linear relation $\mu_w = 0.2642 \lambda + 0.5328$.

What would settle it

Record the candidates of one PSA-CMA-ES generation on 2D Rastrigin and compare their ranking by $f$ with their ranking by the raw $\mathcal{N}(m,\sigma^2 C)$ draws near the generation where blow-up begins; if the two rankings differ substantially, Theorem 1's premise $E[N_{i:\lambda_r}] = m + \sigma E[\mathcal{N}(0,I)]$ is violated. As a second check, re-estimate $\mu_w$ as a function of $\lambda$ using the exact weights of Algorithm 2; if the linear fit $\mu_w = 0.2642 \lambda + 0.5328$ is poor at the population sizes where blow-up appears, the $L=6$ threshold does not follow.

Watch

Extended reading notes

Core claim

The central discovery is that the step-size correction formula of PSA-CMA-ES is itself the amplifier that drives step-size blow-up. The correction sets $\sigma_c^{(g+1)} = \sigma^{(g+1)} \rho(\lambda_r^{(g+1)})/\rho(\lambda_r^{(g)})$, where $\rho$ contains the sum of weighted expected normal order statistics and the variance-effective selection mass $\mu_w$. Theorem 1 shows that if the expected normal order statistics and the mean are non-increasing across generations while the selection mass changes by at most a small $\delta$, then $\rho(\lambda_r^{(g+1)})/\rho(\lambda_r^{(g)}) \geq 1$, hence $\sigma_c^{(g+1)}/\sigma^{(g+1)} \geq 1$; since the corrected step-size is the starting step-size of the next generation, the inflation compounds. Theorem 2 identifies when this regime is active: because $\mu_w \approx 0.2642 \lambda + 0.5328$ over the algorithm's weights, a population size change below $L$ keeps the change in $\mu_w$ below $\delta$, and the experiments place that threshold at $L=6$ for 2D Rastrigin. The reformulation therefore gates the correction on the cumulative step-size path, applying it only when $\|p_\sigma\|$ is below its expected length $\mathbb{E}[\|\mathcal{N}(0,I)\|]$ (the adaptation is asking for smaller steps), and inside that gate scales it by $\kappa=0.5$ when the population size change is below $L$ while keeping the original correction for larger changes. On 2D Rastrigin and Schaffer the reformulation reaches final values closer to zero in roughly a third to half the CPU time of the general algorithm, at the price of more function evaluations per generation on Rastrigin.

Load-bearing premise

The proof assumes that sorting candidates by the objective function $f$ produces the same order statistics as the raw Gaussian samples, so the expected value of the $i$-th best candidate is $m + \sigma$ times an expected normal order statistic, and that the effective selection mass is linear in $\lambda$ with fitted coefficients; if either fails, the claimed $\sigma_c/\sigma \geq 1$ and the $L=6$ threshold need not hold.

Editorial extensions

If this is right

  • The step-size correction in PSA-CMA-ES should be applied conditionally rather than uniformly; gating it on the step-size adaptation's convergence signal prevents blow-up without causing premature convergence.
  • When the population size change is below the significance level, the correction is the dangerous regime, so scaling it down (here by a factor of 0.5) is enough to stabilize the algorithm.
  • Removing the correction completely is not a viable fix, since the experiments show the step-size collapses and the algorithm converges prematurely in that setting.
  • On the two 2D benchmarks, the reformulation reaches better final objective values with substantially lower CPU time than general PSA-CMA-ES, although Rastrigin runs consume more function evaluations per generation.

Reading between the lines

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

  • Editorial inference: the same conditional-correction principle could transfer to other evolution strategies that correct step-size via normal order statistics, since the amplification mechanism is tied to the correction formula rather than to this specific algorithm.
  • Editorial inference: the scaling factor 0.5 and the threshold 6 were calibrated on two two-dimensional benchmarks; in higher dimensions or on other landscapes the optimal values would likely change, and an online estimator of the actual effective-selection-mass change could make the gating self-tuning.
  • Editorial inference: the paper's own observation that population size keeps growing after convergence suggests that adding a population-size shrinking rule to the gated correction would reduce the extra function evaluations, a natural next step the authors identify as future work.
  • Editorial inference: because the proof relies on objective-function ranking matching raw Gaussian order statistics, on landscapes whose geometry strongly distorts that ranking the blow-up may be milder or absent, so the failure may be landscape-dependent rather than universal.
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 / 6 minor

Summary. The paper studies step-size blow-up in PSA-CMA-ES on weakly structured multimodal functions. It proposes a reformulated step-size correction (Algorithm 1) that applies the original correction only when the CSA evolution path is shorter than expected (anti-correlated steps), and scales the correction by a factor kappa when the population-size change is below a threshold L. The claimed analytical contribution is Theorem 1, which asserts that the correction ratio rho(lambda^(g+1))/rho(lambda^(g)) >= 1 under a decreasing mean and small change in mu_w, and Theorem 2, which derives L=6 from a linear fit of mu_w against lambda. Experiments on 2D Rastrigin and Schaffer functions report lower CPU time and better final values than general PSA-CMA-ES, with kappa=0.5 selected by grid search.

Significance. If the theoretical claims were sound, the paper would provide a useful diagnosis of a known failure mode of PSA-CMA-ES and a simple, cheap fix; the gating idea is interesting, and the experiments suggest the reformulation may help on the two tested functions. The authors also report detailed per-run tables, which is transparent, and they identify a real limitation of PSA-CMA-ES. However, the central analytical identification is not established, and the threshold and scaling parameters are calibrated on the same benchmarks used for evaluation. The practical heuristic may still be of interest, but as submitted the paper does not substantiate the claimed analytical cause or a transferable significance level.

major comments (4)
  1. [Section 3.1.1, Eqs. (2.13)-(2.14) and Theorem 1] The proof assumes E[N_{i:lambda_r}] = m + sigma E[N(0,I)], i.e., that the i-th best candidate under f has the same distribution as the i-th order statistic of the raw Gaussian sample. This is only valid if ranking by f preserves the sample order of the multivariate normal draws, which is false for the multimodal Rastrigin and Schaffer functions used in the paper (Eqs. (3.8)-(3.9)). Since the ordering can be permuted by the basin structure, the inequalities (3.1)-(3.6) do not follow, and the conclusion rho(lambda^(g+1)) >= rho(lambda^g), hence sigma_c/sigma >= 1, is unsupported. The theorem also never uses the covariance matrix C or the geometry of f, so it cannot distinguish multimodal from unimodal landscapes.
  2. [Proof of Theorem 1, Eqs. (3.2)-(3.3)] The proof is internally inconsistent. Eq. (3.2) is obtained by multiplying the reversed inequality in (3.1) by n*mu_w^(g+1) and comparing with generation g; for the stated direction one needs mu_w^(g+1) >= mu_w^(g). Two lines later the proof asserts 'When lim_{delta->0}, we have mu_w^(g+1) < mu_w^(g)' and uses that to obtain the squared inequality in (3.3). Both cannot hold simultaneously, so even under the paper's own order-statistic assumption the derivation is not valid.
  3. [Section 3.1.2 and Section 3.2.1, Theorem 2 and Figs. 2-3] The existence of the significance level L is not established. The proof of Theorem 2 consists of fitting mu_w = 0.2642 lambda + 0.5328 to unspecified data and then defining q = 0.2642; no fitting range, no error estimates, and no statistical justification are given. The value delta = 1.5852 is then implied by L = 6, which was read off the same Rastrigin experiments. Moreover, the text in Section 3.2.1 equates a step-size difference 'Delta sigma(g) >= 6' with a population-size significance level L = 6, although these are different quantities. Since L is used in Algorithm 1 and Algorithm 2 and the algorithm is evaluated on the very same Rastrigin and Schaffer functions, the procedure is a calibration, not a validation, of L.
  4. [Sections 3.2.2 and 3.4, Fig. 4 and Tables 3-9] The scaling parameter kappa is selected by grid search on the same two benchmark functions (Fig. 4, Tables 6-9), and then the reformulated algorithm with kappa = 0.5 and L = 6 is evaluated on those same functions (Fig. 5, Tables 3-4). There is no independent test set, no comparison with other state-of-the-art algorithms (e.g., IPOP-CMA-ES or standard CMA-ES with a large population), no higher-dimensional results, and no statistical significance testing; the reported averages, such as CPU time 33.1779 versus 116.5880, are not accompanied by variance measures or confidence intervals. The conclusions about general improvement are therefore not supported beyond the specific 2D cases.
minor comments (6)
  1. [Section 3.2.1] The text refers to Eqs. (4.3a) and (4.3b), but the equations appear in Section 3 and should be numbered consistently, likely as (3.10a) and (3.10b).
  2. [Section 3.2.1] The phrase 'when Delta sigma(g) >= 6' is ambiguous: Delta sigma(g) denotes the step-size difference, while L is a threshold on |lambda^(g+1) - lambda^g|; these quantities are used interchangeably without justification.
  3. [Abstract and Section 4] There are several typos and formatting issues, including 'covariance matrix adaption' in the abstract, a missing period before 'enhancing convergence', and the heading 'F uture W orks' in Section 4; 'algrithm' appears in the text.
  4. [Eqs. (2.13)-(2.14)] Eq. (2.13) writes E[N_{i:lambda_r}] = m + sigma E[N(0,I)], but E[N(0,I)] is a vector, while Eq. (2.14) gives a scalar approximation; the notation is not reconciled.
  5. [Algorithm 1 and Theorem 2] Theorem 2 states lambda^(g+1) - lambda^g <= L, while Algorithm 1 uses |lambda^(g+1) - lambda^g| < L; the strictness and the absolute value should be made consistent.
  6. [Algorithm 2 and Table 2] The paper does not provide code or detailed initialization and termination specifications; for reproducibility, the exact random number generation and stopping criteria should be stated precisely.

Circularity Check

2 steps flagged · score 7.0 of 10

The claimed 'significance level' L and the scaling parameter κ are fitted on the very benchmarks used for evaluation; Theorem 2 defines L through δ=qL after L was read off from the same experiments, so the analytical derivation of the safe threshold reduces to the empirical observation.

  1. fitted input called prediction [Section 3.1.2, Theorem 2; Section 3.2.1, Experiments 1 and 2]
    "Due to the complexity of its analytical form in the algorithm, g (·) was numerically estimated using curve fitting. Accordingly, the best fit estimation yields the following linear function, µw = 0.2642 λ + 0.5328. ... λ(g+1) − λ(g) ≤ L, such that δ = qL. ... Based on these results for 2D Rastrigin function, ∆ σ(g) starts blowing up after generation 6 when ∆ σ(g) ≥ 6. ... Therefore it can be stated that 6 is a significant level ... This further clarifies that step-size blow up when the change in µw falls below δ = 1.5852 according to Theorem 1."

    The threshold L is first read off from the same controlled experiments ('6 is a significant level'), and δ is then set to qL = 0.2642 × 6 = 1.5852. Since δ is the unspecified small constant from Theorem 1 and q is the fitted slope, the equation δ = qL is an identity rather than a constraint. Algorithm 1's condition |λ(g+1) − λ(g)| < L therefore uses the empirically observed cutoff relabeled as an analytically derived significance level; the derivation does not independently produce L.

  2. fitted input called prediction [Section 3.2.2 and Section 3.4]
    "The optimal value of κ was determined using experimentally, specifically for 2D Rastrigin and Schaffer functions... For both functions the optimum κ was found to be 0.5. However, it should be noted that this value may change for other functions."

    Algorithm 1 and Algorithm 2 take κ as an input, and κ is selected by grid search on the two benchmark functions that are later used in Section 3.4 to demonstrate the reformulation's superior performance. The reported CPU-time and convergence improvements are thus re-evaluations on the tuning set, not independent predictions; the comparison does not separate the effect of the reformulation from the effect of choosing κ (and L) on those two functions.

full rationale

The paper's central analytical claim is that the step-size correction mechanism itself causes blow-up (Theorem 1) and that there is a significance level L for population-size change (Theorem 2). The L result reduces by construction to the empirical observation: L=6 is identified from Experiment 1 on 2D Rastrigin, then 'derived' by fitting µw(λ) as a line and choosing δ=qL=1.5852; no independent bound on δ is given. The reformulation's second input, κ, is grid-searched on the same two functions on which the algorithm is later evaluated, so the performance comparison is not an out-of-sample test. Separately, Theorem 1's proof assumes E[N_{i:λ_r}] = m + σ E[N(0,I)] and that ranking by f yields these raw order statistics; this is not verified for multimodal Rastrigin/Schaffer, and the proof's inequality chain (3.2)-(3.5) has sign inconsistencies (e.g., Eq. (3.2) needs µw^(g+1) ≥ µw^(g) while the δ→0 limit uses µw^(g+1) < µw^(g)). Those are validity concerns rather than circularity per se, but they add to the weight: the analytical identification of the blow-up mechanism is not independently established. Because the central threshold and the evaluation share their fitted inputs, a score of 7 reflects partial-to-substantial circularity without claiming the entire reformulation is vacuous.

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

The central claims rest on several fitted constants: kappa and L are tuned on the evaluation benchmarks, delta is back-calculated to match L, and mu_w is replaced by a linear fit. The proof of Theorem 1 additionally assumes a specific order-statistics model that ignores the objective function. No code or data is provided, so the fitted values cannot be independently checked.

free parameters (4)
  • kappa (step-size correction scaling factor) = 0.5
    Selected by grid search on 2D Rastrigin and Schaffer functions (Section 3.2.2) and used in the final evaluation on the same functions.
  • L (significance level for population size change) = 6
    Determined from controlled experiments on Rastrigin (Section 3.2.1) and used as a fixed threshold for both Rastrigin and Schaffer in the reformulation.
  • delta (bound on mu_w change) = 1.5852
    Back-calculated as q times L with q = 0.2642; not derived independently in the paper.
  • q and intercept of the mu_w vs lambda linear fit = q = 0.2642, intercept = 0.5328
    Empirical curve fit to numerically estimated mu_w values in Theorem 2; used to derive L.
assumptions (4)
  • domain assumption E[Ni:lambda] = m + sigma E[N(0,I)] for the expected normal order statistic of ranked candidates
    Invoked in Section 2.2 and in the proof of Theorem 1; conflates ordering by objective-function values with order statistics of raw Gaussian samples, ignoring f and C.
  • domain assumption The mean vector m is non-increasing across generations when minimizing
    Assumed in Theorem 1; not guaranteed for multimodal functions where the algorithm may move between basins.
  • ad hoc to paper mu_w is a linear function of lambda with slope 0.2642 over the operating range
    Empirical curve fit in Theorem 2; no error bars or range justification is provided.
  • ad hoc to paper The direction of step-size adaptation (norm of p_sigma vs E[||N(0,I)||]) correctly identifies when to apply the correction
    Core design choice of Algorithm 1; not justified beyond intuition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving population size adapting CMA-ES algorithm on step-size blow-up in weakly-structured multimodal functions." pith.science (2026). https://pith.science/paper/MRT4PFZO

@misc{pith2026250600825,
  author       = {Pith},
  title        = {Pith review of: Improving population size adapting CMA-ES algorithm on step-size blow-up in weakly-structured multimodal functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRT4PFZO}},
  note         = {Machine review of arXiv:2506.00825}
}
read the original abstract

Multimodal optimization requires both exploration and exploitation. Exploration identifies promising attraction basins, while exploitation finds the best solutions within these basins. The balance between exploration and exploitation can be maintained by adjusting parameter settings. The population size adaptation covariance matrix adaption evolutionary strategy algorithm (PSA-CMA-ES) achieves this balance by dynamically adjusting population size. PSA-CMA-ES performs well on well-structured multimodal benchmark problems. In weakly structured multimodal problems, however, the algorithm struggles to effectively manage step-size increases, resulting in uncontrolled step-size blow-ups that impede convergence near the global optimum. In this study, we reformulated the step-size correction strategy to overcome this limitation. We analytically identified the cause of the step-size blow-up and demonstrate the existence of a significance level for population size change guiding a safe passage to step-size correction. These insights were incorporated to form the reformulation. Through computer experiments on two weakly structured multimodal benchmark problems, we evaluated the performance of the new approach and compared the results with the state-of-the-art algorithm. The improved algorithm successfully mitigates step-size blow-up, enabling a better balance between exploration and exploitation near the global optimum enhancing convergence.

Figures

Figures reproduced from arXiv: 2506.00825 by the authors.

Figure 1
Figure 1. Visualization of the benchmark test functions; Rastrigin function is depicted in the upper panel [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Summary of the results across experiments 1 and 2 for the 2D Rastrigin function. Panels (a) [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Summary of the results across experiments 1 and 2 for the 2D Rastrigin function. Panels (a) [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The values of observed performance metrics for each [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison of the Reformulation against General PSA-CMA-ES on the 2D Rastrigin [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

  1. [1]

    Akimoto, Y., Auger, A., and Hansen, N. (2017). Quality gain analysis of the weighted recombination evolution strategy on general convex quadratic functions. In Proceedings of the 14th ACM/SIGEVO Conference on Foundations of Genetic Algorithms, pages 111–126

    work page 2017

  2. [2]

    and Hansen, N

    Auger, A. and Hansen, N. (2005a). Performance evaluation of an advanced local search evolutionary algorithm. In 2005 IEEE congress on evolutionary computation, volume 2, pages 1777–1784. IEEE

    work page 2005

  3. [3]

    and Hansen, N

    Auger, A. and Hansen, N. (2005b). A restart cma evolution strategy with increasing population size. In 2005 IEEE congress on evolutionary computation, volume 2, pages 1769–1776. IEEE

    work page 2005

  4. [4]

    Back, T. (1996). Evolutionary algorithms in theory and practice: evolution strategies, evolutionary programming, genetic algorithms. Oxford university press. 20

    work page 1996

  5. [5]

    and Sendhoff, B

    Beyer, H.-G. and Sendhoff, B. (2008). Covariance matrix adaptation revisited–the cmsa evolution strategy–. In International Conference on Parallel Problem Solving from Nature, pages 123–132. Springer

    work page 2008

  6. [6]

    Fujii, G., Akimoto, Y., and Takahashi, M. (2018). Exploring optimal topology of thermal cloaks by cma-es. Applied Physics Letters, 112(6)

    work page 2018

  7. [7]

    and Schmidhuber, J

    Ha, D. and Schmidhuber, J. (2018). World models. arXiv preprint arXiv:1803.10122

    arXiv 2018

  8. [8]

    Hansen, N. (2006). The cma evolution strategy: a comparing review. Towards a new evolutionary computation: Advances in the estimation of distribution algorithms, pages 75–102

    work page 2006

Show all 24 references
  1. [9]

    Hansen, N. (2009). Benchmarking a bi-population cma-es on the bbob-2009 func- tion testbed. In Proceedings of the 11th annual conference companion on genetic and evolutionary computation conference: late breaking papers, pages 2389–2396

    2009

  2. [10]

    Hansen, N. (2016). The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772

    2016 arXiv

  3. [11]

    and Kern, S

    Hansen, N. and Kern, S. (2004). Evaluating the cma evolution strategy on multimodal test functions. In International conference on parallel problem solving from nature, pages 282–291. Springer

    2004

  4. [12]

    D., and Koumoutsakos, P

    Hansen, N., M¨ uller, S. D., and Koumoutsakos, P. (2003). Reducing the time complex- ity of the derandomized evolution strategy with covariance matrix adaptation (cma-es). Evolutionary computation, 11(1):1–18

    2003

  5. [13]

    and Ostermeier, A

    Hansen, N. and Ostermeier, A. (1996). Adapting arbitrary normal mutation distri- butions in evolution strategies: The covariance matrix adaptation. In Proceedings of IEEE international conference on evolutionary computation, pages 312–317. IEEE

    1996

  6. [14]

    and Ostermeier, A

    Hansen, N. and Ostermeier, A. (2001). Completely derandomized self-adaptation in evolution strategies. Evolutionary computation, 9(2):159–195

    2001

  7. [15]

    and Ros, R

    Hansen, N. and Ros, R. (2010). Benchmarking a weighted negative covariance matrix update on the bbob-2010 noiseless testbed. In Proceedings of the 12th annual conference companion on Genetic and evolutionary computation, pages 1673–1680

    2010

  8. [16]

    Maki, A., Sakamoto, N., Akimoto, Y., Nishikawa, H., and Umeda, N. (2020). Appli- cation of optimal control theory based on the evolution strategy (cma-es) to automatic berthing. Journal of Marine Science and Technology, 25:221–233. 21

    2020

  9. [17]

    and Akimoto, Y

    Nishida, K. and Akimoto, Y. (2016). Population size adaptation for the cma-es based on the estimation accuracy of the natural gradient. In Proceedings of the Genetic and Evolutionary Computation Conference 2016, pages 237–244

    2016

  10. [18]

    and Akimoto, Y

    Nishida, K. and Akimoto, Y. (2018a). Benchmarking the psa-cma-es on the bbob noiseless testbed. In Proceedings of the Genetic and Evolutionary Computation Confer- ence Companion, pages 1529–1536

    2018

  11. [19]

    and Akimoto, Y

    Nishida, K. and Akimoto, Y. (2018b). Psa-cma-es: Cma-es with population size adaptation. In Proceedings of the Genetic and Evolutionary Computation Conference, pages 865–872

    2018

  12. [20]

    Piergiovanni, A., Angelova, A., and Ryoo, M. S. (2020). Evolving losses for unsu- pervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 133–142

    2020

  13. [21]

    Royston, J. (1982). Algorithm as 177: Expected normal order statistics (exact and approximate). Journal of the royal statistical society. Series C (Applied statistics), 31(2):161–165

    1982

  14. [22]

    Ryan, P. I. C. et al. (2007). References to cma-es applications. Strategies, 4527(467)

    2007

  15. [23]

    M., Smith, A., and Risi, S

    Volz, V., Schrum, J., Liu, J., Lucas, S. M., Smith, A., and Risi, S. (2018). Evolving mario levels in the latent space of a deep convolutional generative adversarial network. In Proceedings of the genetic and evolutionary computation conference, pages 221–228. 22

    2018

  16. [24]

    Appendix Algorithm 2:PSA-CMA-ES with Reformulated Step-size Correction Input: m(0) ∈ Rn, σ(0) ∈ R+, κ, L, a, b, gmax, ϵ Set: cm = 1, Pµ i=1 wi = 1, α= 1.4, β= 0.4, λmin = λdef, λmax = 512 λdef Initialize: C(0) = I, p(0) c = 0, p(0) σ = 0, p(0) θ = 0, γ(0) c = 0, γ(0) σ = 0, γ(...

    1906

Pith tools

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