Convergence Analysis of Evolution Strategies for Mixed-Integer Optimization
Pith reviewed 2026-05-21 01:50 UTC · model grok-4.3
The pith
The (1+1)-LUB-ES achieves linear convergence on mixed-integer benchmarks after integers are optimized, unlike the LB-ES which risks premature convergence with many integers.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
By applying drift analysis to the phase after integer variables are fixed, the paper shows that on a benchmark function for mixed-integer domains, the (1+1)-LB-ES variant experiences premature convergence for large numbers of integer variables, while the (1+1)-LUB-ES variant, which uses both lower and upper bounds on the standard deviation, attains linear convergence when parameters are chosen appropriately.
What carries the argument
Drift analysis of the (1+1) evolution strategy variants with lower-bound and lower-upper-bound constraints on the standard deviation for the integer coordinates.
If this is right
- If correct, then (1+1)-LB-ES will show decreasing convergence speed as integer count rises due to the fixed lower bound affecting continuous steps.
- (1+1)-LUB-ES will deliver consistent linear convergence rates for continuous variables regardless of integer count under suitable bounds.
- The choice of bounds directly impacts the tradeoff between preventing premature integer convergence and maintaining continuous progress.
- These results apply specifically to the post-integer-optimization phase on the given benchmark function.
Where Pith is reading between the lines
- Similar bounding techniques could be investigated for simultaneous optimization of integers and continuous variables from the initial population.
- Applying the same analysis to other evolution strategy variants or population sizes might uncover broader patterns in mixed-integer convergence.
- Real-world applications with high-dimensional integers could test whether the linear convergence translates to practical speedups.
Load-bearing premise
The convergence analysis holds only in the phase after the integer variables have already been optimized and is based on results from one specific benchmark function for mixed-integer domains.
What would settle it
Conducting numerical experiments with the (1+1)-LB-ES on the benchmark function while varying the number of integer variables upward and observing if the continuous variable convergence rate deteriorates as expected.
Figures
read the original abstract
Mixed-integer extensions of evolution strategies (ES) that discretize selected coordinates of sampled continuous vectors often impose a lower bound on the standard deviation of integer variables to prevent premature convergence. While these methods show promising empirical results, this handling can slow the convergence of continuous variables, and its impact has lacked a clear theoretical account. In this paper, we provide a convergence analysis of evolution strategies for mixed-integer optimization, inspired by the drift analysis of the (1+1)-ES in the continuous domain. Specifically, we consider two (1+1)-ES variants for mixed-integer domains: (1+1)-LB-ES, which introduces a lower bound on the standard deviation for integer variables, and (1+1)-LUB-ES, which combines both lower and upper bounds to enhance the convergence of the continuous variables. Focusing on the optimization phase after the integer variables have been optimized, we rigorously analyze their convergence behavior on a benchmark function designed for mixed-integer domains. Our results show that (1+1)-LB-ES can suffer from premature convergence when the number of integer variables is large, while (1+1)-LUB-ES achieves linear convergence under suitable parameter settings. These findings provide theoretical insights into the impact of integer handling on convergence performance and guidance for the design of mixed-integer ES.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript analyzes the convergence of two (1+1) evolution strategy variants for mixed-integer optimization: (1+1)-LB-ES with a lower bound on the standard deviation for integer variables and (1+1)-LUB-ES with both lower and upper bounds. Using drift analysis inspired by continuous-domain results, it focuses on the phase after integer variables are optimized and evaluates performance on a purpose-built benchmark function for mixed-integer domains. The results indicate that (1+1)-LB-ES may experience premature convergence as the number of integer variables increases, whereas (1+1)-LUB-ES attains linear convergence with appropriate parameter choices.
Significance. If the analysis holds, this work offers valuable theoretical insights into how bounding mechanisms for integer variable standard deviations affect the convergence speed of continuous variables in mixed-integer evolution strategies. The application of drift analysis to this setting is a positive step toward rigorous understanding, and the identification of conditions for linear convergence in LUB-ES could inform practical algorithm design. However, the restriction to the post-optimization phase for integers limits the generality of the findings for the full mixed-integer problem.
major comments (2)
- Abstract: The analysis is explicitly restricted to the optimization phase after the integer variables have already been optimized. This scoping decouples the integer and continuous search, so the observed premature convergence in LB-ES with large numbers of integer variables and the linear convergence in LUB-ES are demonstrated only in this simplified regime, not in the joint optimization setting that motivates the lower-bound handling on integer standard deviations.
- The manuscript states it provides a 'rigorous' drift analysis but the provided text omits full proof details; without these, it is unclear whether the linear convergence claim for LUB-ES under suitable parameters is free of hidden assumptions or relies on post-hoc choices for the lower/upper bounds on integer standard deviations.
minor comments (1)
- Ensure that the benchmark function is fully specified with all parameters in a dedicated section or appendix to allow reproducibility of the drift analysis results.
Simulated Author's Rebuttal
Thank you for the opportunity to respond to the referee's report. We address the major comments point by point below and have made revisions to the manuscript to incorporate the feedback where possible.
read point-by-point responses
-
Referee: Abstract: The analysis is explicitly restricted to the optimization phase after the integer variables have already been optimized. This scoping decouples the integer and continuous search, so the observed premature convergence in LB-ES with large numbers of integer variables and the linear convergence in LUB-ES are demonstrated only in this simplified regime, not in the joint optimization setting that motivates the lower-bound handling on integer standard deviations.
Authors: We thank the referee for highlighting this important scoping issue. Our analysis deliberately focuses on the post-integer-optimization phase because this is where the lower bound on integer standard deviations directly affects the convergence rate of the continuous variables. The lower bound is introduced to avoid premature convergence in the integer variables during the joint search, but once the integers are fixed, this bound can cause the step sizes for continuous variables to remain too large, leading to slower convergence. The benchmark function and drift analysis are designed to isolate and study this effect. We agree that extending the analysis to the full joint optimization would provide a more complete picture, but this would require a significantly different approach and is left for future work. In the revised manuscript, we have updated the abstract and introduction to more explicitly state the scope and motivation. revision: yes
-
Referee: The manuscript states it provides a 'rigorous' drift analysis but the provided text omits full proof details; without these, it is unclear whether the linear convergence claim for LUB-ES under suitable parameters is free of hidden assumptions or relies on post-hoc choices for the lower/upper bounds on integer standard deviations.
Authors: We apologize for the omission of full proof details in the submitted manuscript, which was due to length limitations. The drift analysis is rigorous and follows the methodology established in the continuous domain literature, adapted to account for the mixed-integer structure and the bounding mechanisms. The lower and upper bounds on the standard deviations for integer variables are derived from the conditions necessary to satisfy the drift inequalities for linear convergence; they are not chosen post-hoc but are part of the parameter settings that guarantee the desired behavior. In the revised version, we will include the complete proofs in an appendix to allow readers to verify all assumptions and steps. revision: yes
Circularity Check
Standard drift analysis on scoped post-integer benchmark without reduction to fitted inputs or self-citations
full rationale
The paper applies established drift-analysis techniques from continuous (1+1)-ES to the restricted regime after integer variables are fixed, on a purpose-built mixed-integer benchmark function. The claimed premature convergence for LB-ES (with growing integer count) and linear convergence for LUB-ES under suitable parameters follow directly from the drift bounds and parameter choices in that simplified setting. No equations reduce the linear rate to a quantity defined by fitting the same data, no self-definitional loops appear, and no load-bearing self-citations or uniqueness theorems imported from the authors' prior work are invoked to force the result. The derivation remains self-contained against the external benchmark of classical drift analysis.
Axiom & Free-Parameter Ledger
free parameters (2)
- lower bound on standard deviation for integer variables
- upper bound on standard deviation for integer variables
axioms (2)
- domain assumption Analysis is performed only after integer variables have reached their optimal values.
- domain assumption The benchmark function is representative of mixed-integer domains for the purpose of drift analysis.
Reference graph
Works this paper leans on
-
[1]
Vahab Akbarzadeh, Albert Hung-Ren Ko, Christian Gagné, and Marc Parizeau
-
[2]
In Parallel Problem Solving from Nature - PPSN XI, Vol
Topography-Aware Sensor Deployment Optimization with CMA-ES. In Parallel Problem Solving from Nature - PPSN XI, Vol. 6239. 141–150. doi:10.1007/ 978-3-642-15871-1_15 GECCO ’26, July 13–17, 2026, San Jose, Costa Rica R. Hamano et al
work page 2026
-
[3]
Youhei Akimoto, Anne Auger, and Tobias Glasmachers. 2018. Drift theory in continuous search spaces: expected hitting time of the (1+1)-ES with 1/5 success rule. InProceedings of the Genetic and Evolutionary Computation Conference. 801–808. doi:10.1145/3205455.3205606
-
[4]
Hans-Georg Beyer and Hans-Paul Schwefel. 2002. Evolution strategies - A comprehensive introduction.Natural Computing1 (03 2002), 3–52. doi:10.1023/A: 1015059928466
work page doi:10.1023/a: 2002
-
[5]
Thomas Bäck. 1996.Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary Programming, Genetic Algorithms. Oxford University Press. doi:10.1093/oso/9780195099713.001.0001
-
[6]
2013.Contemporary Evolution Strategies
Thomas Bäck, Christophe Foussette, and Peter Krause. 2013.Contemporary Evolution Strategies. Vol. 47. doi:10.1007/978-3-642-40137-4
-
[7]
Diego Dominici. 2003. The inverse of the cumulative standard normal probability function. arXiv:math/0306264 [math.CA] https://arxiv.org/abs/math/0306264
work page internal anchor Pith review Pith/arXiv arXiv 2003
-
[8]
Yinpeng Dong, Hang Su, Baoyuan Wu, Zhifeng Li, Wei Liu, Tong Zhang, and Jun Zhu. 2019. Efficient Decision-Based Black-Box Adversarial Attacks on Face Recog- nition. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7706–7714. doi:10.1109/CVPR.2019.00790
-
[9]
Michael Emmerich, Monika Grötzner, Bernd Groß, and Martin Schütz. 2000. Mixed-Integer Evolution Strategy for Chemical Plant Optimization with Sim- ulators. InEvolutionary Design and Manufacture, I. C. Parmee (Ed.). Springer London, London, 55–67
work page 2000
-
[10]
Garuda Fujii, Masayuki Takahashi, and Youhei Akimoto. 2018. CMA-ES- based structural topology optimization using a level set boundary expression– Application to optical and carpet cloaks.Computer Methods in Applied Mechanics and Engineering332 (2018), 624–643. doi:10.1016/j.cma.2018.01.008
-
[11]
Robert D. Gordon. 1941. Values of Mills’ Ratio of Area to Bounding Ordinate and of the Normal Probability Integral for Large Values of the Argument.The Annals of Mathematical Statistics12, 3 (1941), 364–366
work page 1941
-
[12]
Ryoki Hamano, Masahiro Nomura, Shota Saito, Kento Uchida, and Shinichi Shirakawa. 2025. CatCMA with Margin: Stochastic Optimization for Continuous, Integer, and Categorical Variables. InProceedings of the Genetic and Evolutionary Computation Conference (GECCO ’25). 737–745. doi:10.1145/3712256.3726471
-
[13]
Ryoki Hamano, Shota Saito, Masahiro Nomura, and Shinichi Shirakawa. 2022. CMA-ES with Margin: Lower-Bounding Marginal Probability for Mixed-Integer Black-Box Optimization. InProceedings of the Genetic and Evolutionary Computa- tion Conference. 639–647. doi:10.1145/3512290.3528827
-
[14]
Ryoki Hamano, Shota Saito, Masahiro Nomura, and Shinichi Shirakawa. 2024. Marginal Probability-Based Integer Handling for CMA-ES Tackling Single- and Multi-Objective Mixed-Integer Black-Box Optimization.ACM Transactions on Evolutionary Learning and Optimization4, 2, Article 10 (June 2024), 26 pages. doi:10.1145/3632962
-
[15]
Ryoki Hamano, Shota Saito, Masahiro Nomura, Kento Uchida, and Shinichi Shi- rakawa. 2024. CatCMA: Stochastic Optimization for Mixed-Category Problems. InProceedings of the Genetic and Evolutionary Computation Conference. 656–664. doi:10.1145/3638529.3654198
-
[16]
2011.A CMA-ES for Mixed-Integer Nonlinear Optimization
Nikolaus Hansen. 2011.A CMA-ES for Mixed-Integer Nonlinear Optimization. Research Report. INRIA
work page 2011
-
[17]
Diestel,Graph Theory, 5th ed., Springer, Berlin, 2017
Nikolaus Hansen, Dirk V. Arnold, and Anne Auger. 2015.Evolution Strategies. Springer Berlin Heidelberg, Berlin, Heidelberg, 871–898. doi:10.1007/978-3-662- 43505-2_44
-
[18]
N. Hansen and A. Ostermeier. 1996. Adapting Arbitrary Normal Mutation Dis- tributions in Evolution Strategies: The Covariance Matrix Adaptation. InPro- ceedings of IEEE International Conference on Evolutionary Computation. 312–317. doi:10.1109/ICEC.1996.542381
-
[19]
Koki Ikeda and Isao Ono. 2023. Natural Evolution Strategy for Mixed-Integer Black-Box Optimization. InProceedings of the Genetic and Evolutionary Computa- tion Conference. 831–838. doi:10.1145/3583131.3590518
-
[20]
Timo Kötzing and Aishwarya Radhakrishnan. 2025. Mixed-Binary Problems Op- timized with Fast Discrete Solver. InEvolutionary Computation in Combinatorial Optimization. 167–183
work page 2025
- [21]
-
[22]
Emmerich, Jeroen Eggermont, Thomas Bäck, M
Rui Li, Michael T.M. Emmerich, Jeroen Eggermont, Thomas Bäck, M. Schütz, J. Dijkstra, and J.H.C. Reiber. 2013. Mixed Integer Evolution Strategies for Parameter Optimization.Evolutionary Computation21, 1 (03 2013), 29–64. arXiv:https://direct.mit.edu/evco/article-pdf/21/1/29/1496431/evco_a_00059.pdf doi:10.1162/EVCO_a_00059
-
[23]
Ilya Loshchilov and Frank Hutter. 2016. CMA-ES for Hyperparameter Optimiza- tion of Deep Neural Networks.arXiv:1604.07269(2016). https://arxiv.org/abs/ 1604.07269
work page internal anchor Pith review Pith/arXiv arXiv 2016
-
[24]
2010.Fractional Calculus and Waves in Linear Viscoelasticity
Francesco Mainardi. 2010.Fractional Calculus and Waves in Linear Viscoelasticity. IMPERIAL COLLEGE PRESS. arXiv:https://www.worldscientific.com/doi/pdf/10.1142/p614 doi:10.1142/p614
-
[25]
Tristan Marty, Nikolaus Hansen, Anne Auger, Yann Semet, and Sébastien Héron
-
[26]
InParallel Problem Solving from Nature – PPSN XVIII
LB+IC-CMA-ES: Two Simple Modifications of CMA-ES to Handle Mixed- Integer Problems. InParallel Problem Solving from Nature – PPSN XVIII. 284–299
-
[27]
Daiki Morinaga and Youhei Akimoto. 2019. Generalized Drift Analysis in Con- tinuous Domain: Linear Convergence of (1+1)-ES on Strongly Convex Functions with Lipschitz Continuous Gradients. InProceedings of the 15th ACM/SIGEVO Con- ference on Foundations of Genetic Algorithms. 13–24. doi:10.1145/3299904.3340303
-
[28]
Daiki Morinaga, Kazuto Fukuchi, Jun Sakuma, and Youhei Akimoto. 2021. Conver- gence rate of the (1+1)-evolution strategy with success-based step-size adaptation on convex quadratic functions. InProceedings of the Genetic and Evolutionary Computation Conference. 1169–1177. doi:10.1145/3449639.3459289
-
[29]
Daiki Morinaga, Kazuto Fukuchi, Jun Sakuma, and Youhei Akimoto. 2024. Con- vergence Rate of the (1+1)-ES on Locally Strongly Convex and Lipschitz Smooth Functions.IEEE Transactions on Evolutionary Computation28, 2 (2024), 501–515. doi:10.1109/TEVC.2023.3266955
-
[30]
I. Rechenberg. 1973.Evolutionsstrategie : Optimierung technischer Systeme nach Prinzipien der biologischen Evolution. Number 15 in Problemata
work page 1973
-
[31]
Günter Rudolph. 1994. An evolutionary algorithm for integer programming. In Parallel Problem Solving from Nature — PPSN III, Yuval Davidor, Hans-Paul Schwe- fel, and Reinhard Männer (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 139–148
work page 1994
-
[32]
Günter Rudolph. 2012.Evolutionary Strategies. Springer Berlin Heidelberg, Berlin, Heidelberg, 673–698. doi:10.1007/978-3-540-92910-9_22
-
[33]
Tim Salimans, Jonathan Ho, Xi Chen, Szymon Sidor, and Ilya Sutskever. 2017. Evolution Strategies as a Scalable Alternative to Reinforcement Learning. arXiv:1703.03864(2017). https://arxiv.org/abs/1703.03864
work page internal anchor Pith review Pith/arXiv arXiv 2017
- [34]
-
[35]
Mikhail A. Semenov and Dmitri A. Terkel. 2003. Analysis of Convergence of an Evolutionary Algorithm with Self-Adaptation using a Stochastic Lya- punov function.Evolutionary Computation11, 4 (12 2003), 363–379. doi:10. 1162/106365603322519279
work page 2003
-
[36]
2000.Coupling, Stationarity, and Regeneration
Hermann Thorisson. 2000.Coupling, Stationarity, and Regeneration
work page 2000
-
[37]
2007.Some Bounds for the Logarithmic Function
Flemming Topsøe. 2007.Some Bounds for the Logarithmic Function. Vol. 4. 137– 151
work page 2007
-
[38]
Tea Tušar, Dimo Brockhoff, and Nikolaus Hansen. 2019. Mixed-integer benchmark problems for single- and bi-objective optimization. InProceedings of the Genetic and Evolutionary Computation Conference. 718–726. doi:10.1145/3321707.3321868
-
[39]
𝜏∑︁ 𝑘=0 𝑠−1∑︁ 𝑗=0 log ∥𝒎𝑠𝑘+𝑗+1 ∥ ∥𝒎𝑠𝑘+𝑗 ∥ # =E
Yohei Watanabe, Kento Uchida, Ryoki Hamano, Shota Saito, Masahiro Nomura, and Shinichi Shirakawa. 2023. (1+1)-CMA-ES with Margin for Discrete and Mixed- Integer Problems. InProceedings of the Genetic and Evolutionary Computation Conference. 882–890. doi:10.1145/3583131.3590516 A Existing Theorems and Propositions Theorem 3 (Proposition 3.1 in [ 33], Theor...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.