pith. sign in

arxiv: 2606.19861 · v1 · pith:ZQCTZBSFnew · submitted 2026-06-18 · 💻 cs.NE

Weight Adaptation for Improving Parallel Performance of Adaptive Stochastic Natural Gradient

Yutaro Yamada , Kento Uchida , Shinichi Shirakawa This is my paper

Pith reviewed 2026-06-26 15:25 UTC · model grok-4.3

classification 💻 cs.NE
keywords weight adaptationadaptive stochastic natural gradientblack-box optimizationevolutionary algorithmsnatural gradientbinary optimizationpopulation sizeparallel evaluation
0
0 comments X

The pith

WA-ASNG adapts weights by maximizing estimated signal from natural gradient accumulations to improve binary optimization over ASNG and PBIL.

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

The paper introduces WA-ASNG as an extension of adaptive stochastic natural gradient methods for black-box optimization. It computes an estimated signal from accumulations of the natural gradient and uses gradient ascent to adapt weight parameters in order to maximize that signal. This mechanism is presented as a way to increase the degree of improvement in expected objective value, complementing the learning rate adaptation that already satisfies a condition for monotonic improvement. Experiments on binary problems with population sizes from 25 to 100 show WA-ASNG outperforming both PBIL and the original ASNG, including under strong noise.

Core claim

WA-ASNG calculates the estimated signal of the update direction from the accumulations of the natural gradient. Then, to maximize the signal, WA-ASNG adaptively updates its weight parameters by a gradient ascent over the optimization. While the learning rate adaptation plays a role in satisfying a sufficient condition for monotonic improvement of the expected objective value, the mechanism of weight adaptation is intended to maximize this improvement.

What carries the argument

Weight adaptation mechanism that performs gradient ascent on the estimated signal computed from accumulated natural gradients.

If this is right

  • WA-ASNG outperforms PBIL and ASNG across population sizes 25 to 100 on binary optimization problems.
  • WA-ASNG performs efficiently in the presence of strong noise.
  • Weight adaptation maximizes the improvement that learning rate adaptation already guarantees to be monotonic.
  • The method addresses the need for suitable weights when using larger populations for parallel evaluation of time-consuming tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The signal-maximization idea could be tested as a general addition to other probabilistic model-based evolutionary algorithms.
  • Weight adaptation may reduce manual tuning effort when scaling population size for parallel black-box optimization.
  • The same estimated-signal approach might be examined for continuous rather than binary search spaces.

Load-bearing premise

The estimated signal computed from accumulations of the natural gradient is a reliable proxy that weight updates can maximize to produce genuine improvement rather than amplifying noise or bias.

What would settle it

An experiment showing that WA-ASNG with weight adaptation produces equal or worse results than fixed-weight ASNG on the same binary problems with population sizes 25-100 would falsify the claimed benefit.

Figures

Figures reproduced from arXiv: 2606.19861 by Kento Uchida, Shinichi Shirakawa, Yutaro Yamada.

Figure 1
Figure 1. Figure 1: The performance comparison between ASNG, PBIL and WA-ASNG (pro￾posed). Each figure plots the medians and interquartile ranges of the number of eval￾uations over 31 trials. The success rate is shown near the marker if there exist one or more failed trials. is conditional on the values of all preceding variables, that fluctuate during optimization. The optimal solution for these functions is x ∗ = (1, · · · … view at source ↗
Figure 2
Figure 2. Figure 2: The transitions of the mean of the elements in the probability vector in ASNG, PBIL, and WA-ASNG. We plot the medians and interquartile ranges over 31 trials. 0 1 2 3 4 # evaluations ×10 4 2.5 5.0 7.5 10.0 12.5 15.0 BinVal, N = 500, = 50 0 1 2 3 # evaluations ×10 5 5 10 15 20 LeadingOnes, N = 500, = 50 ASNG WA-ASNG [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The transitions of the learning rates δθ in ASNG and WA-ASNG. We plot the result of typical trials on BinVal and LeadingOnes for N = 500 and λθ = 100. 4.2 Comparison with ASNG and PBIL First, we validated the search performance of WA-ASNG by comparing it with ASNG and PBIL [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The transitions of the elements in the probability vector in typical trials on LeadingOnes for N = 500 with λθ = 50 (left) and λθ = 25 (right). quentially, from the first to the last. To investigate this behavior, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The performance comparison between the variants of ASNG with different weight settings and WA-ASNG (proposed). Each figure plots SP1 value divided by the number of dimensions. The success rate is shown near the marker if there exist one or more failed trials. There is no marker when all trials failed. in successful trials, divided by the success rate. The results with the various fixed weights show that th… view at source ↗
Figure 6
Figure 6. Figure 6: The transitions of the learning rates δθ in the ASNG with different weight pa￾rameters and in WA-ASNG (proposed). We plot the results of typical trials on OneMax for N = 500 and λθ = 25. 0.0 0.5 1.0 1.5 2.0 2.5 3.0 # evaluations ×10 4 0.02 0.00 0.02 0.04 0.06 weight WA-ASNG 0.0 0.5 1.0 1.5 2.0 2.5 3.0 # evaluations ×10 4 0.02 0.00 0.02 0.04 0.06 ASNG (CMA-ES normalized) w1 w11 w21 w31 w41 w51 w61 w71 w81 w… view at source ↗
Figure 7
Figure 7. Figure 7: The transitions of the weight parameters in WA-ASNG (proposed). We compare them with the “CMA-ES normalized" weight. We plot the weight value for every 10 ranks in a typical trial on BinVal with N = 500 and λθ = 100. We plot the normalized weight so that the sum of their absolute values equals to that of “CMA-ES normalized”. of WA-ASNG worked on BinVal with N = 500 and λθ = 100. As the opti￾mization procee… view at source ↗
Figure 8
Figure 8. Figure 8: The performance comparison between the variants of ASNG with different types for weight parameters and WA-ASNG (proposed). Each figure plots SP1 value divided by the noise variance on OneMax with Gaussian noise. The success rate is shown near the marker if there exist one or more failed trials. No marker indicates all trials failed. (CMA-ES normalized) deteriorated as the noise variance increased, despite … view at source ↗
read the original abstract

Probabilistic model-based evolutionary algorithms are promising for black-box optimization. Specifically, the adaptive stochastic natural gradient (ASNG) adaptively updates its learning rate, a typical hyperparameter in probabilistic model-based evolutionary algorithms, thereby realizing efficient and robust optimization. Although weight parameters are common hyperparameters, with the increasing demand for parallel evaluation of time-consuming tasks, it remains unclear how to set suitable weights for larger population sizes. In this paper, we propose Weight Adaptation ASNG (WA-ASNG), which incorporates a weight adaptation mechanism into ASNG. We calculated the estimated signal of the update direction from the accumulations of the natural gradient. Then, to maximize the signal, WA-ASNG adaptively updates its weight parameters by a gradient ascent over the optimization. While the learning rate adaptation plays a role in satisfying a sufficient condition for monotonic improvement of the expected objective value, the mechanism of weight adaptation is intended to maximize this improvement. The experimental results demonstrate that WA-ASNG outperforms PBIL and ASNG across various settings with population sizes ranging from 25 to 100 for binary optimization problems. Furthermore, WA-ASNG can perform efficiently in the presence of strong noise. Our code is available at https://github.com/shiralab/WA-ASNG .

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes Weight Adaptation ASNG (WA-ASNG) as an extension of adaptive stochastic natural gradient (ASNG) for probabilistic model-based evolutionary algorithms. It adds a weight adaptation mechanism that computes an estimated signal from accumulations of the natural gradient and performs gradient ascent on the weight parameters to maximize this signal. The authors state that learning-rate adaptation satisfies a sufficient condition for monotonic improvement of the expected objective value, while weight adaptation is intended to maximize this improvement. Experiments claim that WA-ASNG outperforms PBIL and ASNG on binary optimization problems for population sizes 25-100 and remains efficient under strong noise. The code is made available at a public GitHub repository.

Significance. If the weight adaptation mechanism is shown to produce reliable gains beyond the learning-rate adaptation and the reported performance improvements hold under scrutiny, the work could aid parallel scalability of evolutionary algorithms on noisy black-box problems. The public code release is a clear strength that supports reproducibility.

major comments (2)
  1. [Abstract] Abstract: the claim that weight adaptation 'maximizes this improvement' lacks any derivation, bound, or monotonicity argument showing that gradient ascent on the estimated signal from natural-gradient accumulations necessarily increases the true expected objective improvement (in contrast to the sufficient condition stated for learning-rate adaptation). This is load-bearing for the central claim that WA-ASNG outperforms ASNG.
  2. [Abstract] Abstract: the performance claim that WA-ASNG 'outperforms PBIL and ASNG across various settings' with population sizes 25-100 is presented without reference to statistical tests, error bars, number of runs, or specific benchmark instances, making it impossible to assess whether the reported gains are robust or could be explained by noise amplification in the weight-update proxy.
minor comments (1)
  1. The abstract refers to 'various settings' and 'strong noise' without naming the concrete binary problems or noise models used; adding these details would improve clarity even if they appear later in the manuscript.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. Below we respond point-by-point to the major comments and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that weight adaptation 'maximizes this improvement' lacks any derivation, bound, or monotonicity argument showing that gradient ascent on the estimated signal from natural-gradient accumulations necessarily increases the true expected objective improvement (in contrast to the sufficient condition stated for learning-rate adaptation). This is load-bearing for the central claim that WA-ASNG outperforms ASNG.

    Authors: We agree that the abstract phrasing requires clarification. The manuscript states that learning-rate adaptation satisfies a sufficient condition for monotonic improvement of the expected objective value, while weight adaptation is described as intended to maximize the estimated update signal obtained from natural-gradient accumulations. No formal derivation, bound, or monotonicity guarantee is provided for the weight-adaptation step with respect to the true objective (it remains a heuristic based on the proxy signal). We will revise the abstract to explicitly distinguish the two mechanisms, replacing any implication of guaranteed maximization with the wording 'intended to maximize the estimated signal' and noting that performance gains are supported empirically rather than by a theoretical guarantee equivalent to that for the learning rate. revision: yes

  2. Referee: [Abstract] Abstract: the performance claim that WA-ASNG 'outperforms PBIL and ASNG across various settings' with population sizes 25-100 is presented without reference to statistical tests, error bars, number of runs, or specific benchmark instances, making it impossible to assess whether the reported gains are robust or could be explained by noise amplification in the weight-update proxy.

    Authors: The abstract is a concise summary; the full manuscript reports experiments on binary optimization problems with population sizes 25-100, using multiple independent runs, error bars on performance plots, and direct comparisons against PBIL and ASNG, including under noise. To make the abstract self-contained and address the concern, we will add a short clause referencing the experimental protocol (e.g., 'based on 10 independent runs with error bars and consistent outperformance across standard binary benchmarks'). This will allow readers to evaluate robustness without altering the reported outcomes. revision: yes

Circularity Check

0 steps flagged

No significant circularity; weight adaptation is a heuristic justified by experiments

full rationale

The paper describes WA-ASNG as computing an estimated signal from natural-gradient accumulations and performing gradient ascent on weights to maximize that signal, while noting that learning-rate adaptation satisfies a monotonicity condition but weight adaptation is only 'intended to maximize this improvement.' No equations are supplied that reduce the claimed performance gain to a fitted quantity by construction, nor does any step equate a prediction to its own inputs. The central claim of outperformance rests on experimental comparisons with PBIL and ASNG across population sizes and noise levels rather than on a self-referential derivation. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the provided text. This is the normal case of an empirical algorithmic proposal whose justification is external to the method definition itself.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The method implicitly assumes that an estimated signal derived from natural-gradient accumulations is a valid objective for weight adaptation and that gradient ascent on this signal improves the underlying optimization.

pith-pipeline@v0.9.1-grok · 5751 in / 1191 out tokens · 10525 ms · 2026-06-26T15:25:45.129430+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages

  1. [1]

    Theoretical Computer Science832, 42–67 (2020)

    Akimoto, Y., Auger, A., Hansen, N.: Quality gain analysis of the weighted re- combination evolution strategy on general convex quadratic functions. Theoretical Computer Science832, 42–67 (2020). https://doi.org/10.1016/j.tcs.2018.05.015

    work page doi:10.1016/j.tcs.2018.05.015 2020

  2. [2]

    In: International Conference on Machine Learning (ICML)

    Akimoto, Y., Shirakawa, S., Yoshinari, N., Uchida, K., Saito, S., Nishida, K.: Adap- tive stochastic natural gradient method for one-shot neural architecture search. In: International Conference on Machine Learning (ICML). pp. 171–180 (2019)

    2019

  3. [3]

    In: 8th International Workshop on Foundations of Genetic Algorithms (FOGA) 2005

    Arnold, D.V.: Optimal weighted recombination. In: 8th International Workshop on Foundations of Genetic Algorithms (FOGA) 2005. pp. 215–237. Springer (2005). https://doi.org/10.1007/11513575_12

    work page doi:10.1007/11513575_12 2005

  4. [4]

    In: Machine Learning Proceedings 1995, pp

    Baluja, S., Caruana, R.: Removing the genetics from the standard genetic algo- rithm. In: Machine Learning Proceedings 1995, pp. 38–46. Morgan Kaufmann, San Francisco (CA) (1995). https://doi.org/10.1016/B978-1-55860-377-6.50014-1

    work page doi:10.1016/b978-1-55860-377-6.50014-1 1995

  5. [5]

    Algorithmica81(2), 668–702 (Feb 2019)

    Dang, D.C., Lehre, P.K., Nguyen, P.T.: Level-based analysis of the univari- ate marginal distribution algorithm. Algorithmica81(2), 668–702 (Feb 2019). https://doi.org/10.1007/s00453-018-0507-5, https://doi.org/10.1007/s00453-018- 0507-5

    work page doi:10.1007/s00453-018-0507-5 2019

  6. [6]

    Meier, J

    Evans, M., Swartz, T.: Approximating Integrals via Monte Carlo and Deterministic Methods. Oxford University Press (2000). https://doi.org/10.1093/oso/9780198502784.001.0001

    work page doi:10.1093/oso/9780198502784.001.0001 2000

  7. [7]

    Exponential Natural Evolution Strategies,

    Glasmachers, T., Schaul, T., Yi, S., Wierstra, D., Schmidhuber, J.: Exponential natural evolution strategies. In: Proceedings of the 12th Annual Conference on Genetic and Evolutionary Computation. pp. 393–400. Association for Computing Machinery, New York, NY, USA (2010). https://doi.org/10.1145/1830483.1830557

    work page doi:10.1145/1830483.1830557 2010

  8. [8]

    Computers & Chemical Engineering170, 108110 (2023)

    González, L.D., Zavala, V.M.: New paradigms for exploiting parallel experiments in bayesian optimization. Computers & Chemical Engineering170, 108110 (2023). https://doi.org/10.1016/j.compchemeng.2022.108110

    work page doi:10.1016/j.compchemeng.2022.108110 2023

  9. [9]

    In: Proceedings of IEEE International Conference on Evolutionary Computation

    Hansen, N., Ostermeier, A.: Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation. In: Proceedings of IEEE International Conference on Evolutionary Computation. pp. 312–317 (1996). https://doi.org/10.1109/ICEC.1996.542381

    work page doi:10.1109/icec.1996.542381 1996

  10. [10]

    Hansen, N., Auger, A.: Principled Design of Continuous Stochastic Search: From Theory to Practice, pp. 145–180. Springer Berlin Heidelberg, Berlin, Heidelberg (2014). https://doi.org/10.1007/978-3-642-33206-7_8

    work page doi:10.1007/978-3-642-33206-7_8 2014

  11. [11]

    IEEE Transactions on Evolutionary Computation3, 287–297 (1999)

    Harik, G.R., Lobo, F.G., Goldberg, D.E.: The compact genetic algo- rithm. IEEE Transactions on Evolutionary Computation3, 287–297 (1999). https://doi.org/10.1109/4235.797971

    work page doi:10.1109/4235.797971 1999

  12. [12]

    In: 2006 IEEE International Conference on Evolutionary Computation

    Jastrebski, G., Arnold, D.: Improving evolution strategies through active covari- ance matrix adaptation. In: 2006 IEEE International Conference on Evolutionary Computation. pp. 2814–2821 (2006). https://doi.org/10.1109/CEC.2006.1688662

    work page doi:10.1109/cec.2006.1688662 2006

  13. [13]

    Algorithmica83(4), 1096–1137 (2021)

    Lengler, J., Sudholt, D., Witt, C.: The complex parameter landscape of the compact genetic algorithm. Algorithmica83(4), 1096–1137 (2021). https://doi.org/10.1007/s00453-020-00778-4

    work page doi:10.1007/s00453-020-00778-4 2021

  14. [14]

    In: GECCO

    Nishida, K., Akimoto, Y.: PSA-CMA-ES: CMA-ES with population size adaptation. In: Proceedings of the Genetic and Evolutionary Computa- tion Conference. pp. 865–872. Association for Computing Machinery (2018). https://doi.org/10.1145/3205455.3205467 16 Y. Yamada et al

    work page doi:10.1145/3205455.3205467 2018

  15. [15]

    ACM Transactions on Evolutionary Learning and Optimization (2024)

    Nomura, M., Akimoto, Y., Ono, I.: CMA-ES with learning rate adapta- tion. ACM Transactions on Evolutionary Learning and Optimization (2024). https://doi.org/10.1145/3698203

    work page doi:10.1145/3698203 2024

  16. [16]

    Journal of Machine Learning Research18(1), 564–628 (2017)

    Ollivier, Y., Arnold, L., Auger, A., Hansen, N.: Information-geometric optimiza- tion algorithms: A unifying picture via invariance principles. Journal of Machine Learning Research18(1), 564–628 (2017)

    2017

  17. [17]

    Salimans, T., Ho, J., Chen, X., Sidor, S., Sutskever, I.: Evolution strategies as a scalable alternative to reinforcement learning (2017), https://arxiv.org/abs/1703.03864

    Pith/arXiv arXiv 2017

  18. [18]

    In: Bäck, T., Preuss, M., Deutz, A., Wang, H., Doerr, C., Emmerich, M., Trautmann, H

    Yamaguchi, T., Uchida, K., Shirakawa, S.: Adaptive stochastic natural gradient method for optimizing functions with low effective dimensionality. In: Bäck, T., Preuss, M., Deutz, A., Wang, H., Doerr, C., Emmerich, M., Trautmann, H. (eds.) Parallel Problem Solving from Nature – PPSN XVI. pp. 719–731. Springer Inter- national Publishing, Cham (2020)

    2020

  19. [19]

    In: 2009 IEEE Congress on Evolutionary Computation

    Yang, S., Richter, H.: Hyper-learning for population-based incremental learning in dynamic environments. In: 2009 IEEE Congress on Evolutionary Computation. pp. 682–689 (2009). https://doi.org/10.1109/CEC.2009.4983011

    work page doi:10.1109/cec.2009.4983011 2009