EvoGrad: Metaheuristics in a Differentiable Wonderland

Andrea Tangherloni; Beatrice F.R. Citterio

REVIEW 4 major objections 5 minor 34 references

EvoGrad: Metaheuristics in a Differentiable Wonderland

T0 review · 4 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 Differentiable population optimizers beat gradient-agnostic baselines on most benchmarks.

desk verdict EvoGrad is a useful integration of known differentiable reparameterizations for metaheuristics, but its headline claim is undermined by a confounded evaluation that cannot separate gradient access from differentiable operators. read the letter →

arxiv 2506.06320 v1 pith:BNSNSEJU submitted 2025-05-28 cs.NE

Beatrice F.R. Citterio , Andrea Tangherloni This is my paper

classification cs.NE

keywords differentiableprogrammingevolutionarycomputationswarmintelligencereparameterizationtrickGumbel-SoftmaxCMA-EShyperparameterlearningbenchmarkoptimization

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

EvoGrad asks whether evolutionary and swarm optimizers can be turned into fully differentiable computation graphs, and claims that they can — with gains over their classical counterparts. The paper rewrites the stochastic operators of CMA-ES, PSO, GA, and DE (sampling, mutation, crossover, parent selection) as reparameterized, differentiable surrogates, so a loss gradient can flow through the entire optimisation loop. On continuous benchmarks, the differentiable versions consistently match or beat the classical algorithms, with differentiable CMA-ES achieving best fitness values of -93.40 to -103.97 on the 500-dimensional Michalewicz function versus -28.82 for classical CMAES. The practical point is that metaheuristics become trainable modules: their hyperparameters and population positions can be learned end-to-end, which matters for composing them with neural networks and meta-learning pipelines.

What carries the argument

The load-bearing mechanism is the reparameterization, or pathwise, trick applied to every stochastic operator: a random draw that depends on parameters is rewritten as a deterministic transformation of parameter-free noise. Gaussian sampling becomes a learnable Cholesky transform, Bernoulli decisions (mutation masks, crossover events) become Binary-Concrete/Gumbel–Sigmoid variables with straight-through estimators, and parent selection becomes Gumbel–Softmax. With the population and all algorithm hyperparameters registered as learnable parameters of one computation graph, a single outer gradient step can simultaneously move individuals and adapt the algorithm's own coefficients (inertia, step size, covariance, crossover rates).

What would settle it

Re-run the wine-quality regression with Adam tuned (e.g., learning rate 0.01 with a schedule or restarts) and add a gradient-descent-with-restarts baseline to the four benchmark functions; if the differentiable versions no longer win by a wide margin, the claim that differentiability itself drives the gains is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that converting population-based metaheuristics into differentiable operators through the reparameterization trick yields optimizers that consistently outperform traditional, gradient-agnostic algorithms in most tested scenarios. In the strongest numerical evidence, differentiable CMAES on the 500-dimensional Michalewicz function reaches best fitness values of -93.40 to -103.97 versus -28.82 for classical CMAES, and on a small neural-network wine regression task it reaches an average best final loss of 5.66 ± 0.39 versus 71.90 ± 15.34 for Adam under default settings. These results are presented as evidence that fully differentiable evolutionary and swarm optimisation can match or exceed classical forms while adding adaptive hyperparameter learning and compatibility with backpropagation-based pipelines.

Load-bearing premise

The empirical comparisons assume the classical baselines are fair controls, so the central claim collapses if the reported gaps come mainly from unequal hyperparameter settings (for example, Adam's default learning rate versus EvoGrad's higher rate and schedule) or from missing restart-based gradient baselines.

Editorial extensions

If this is right

Differentiable CMA-ES and PSO become the strongest performers, suggesting that learning step-size, covariance, and velocity coefficients provides the largest benefit.
The unified autodiff loop allows metaheuristics to be composed with neural networks in a single training pass, enabling end-to-end hybrid optimisation.
The same reparameterization recipe can be applied to more advanced variants (e.g., SHADE, BIPOP-CMAES) and to other population-based methods, as the paper sketches in its future work.
Because all stochastic operators are relaxed, the optimizers become compatible with gradient-based meta-learning, allowing update rules to be learned rather than hand-coded.

Reading between the lines

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

The wine-regression comparison is not clearly a like-for-like test: Adam ran with its default learning rate of 0.001 while EvoGrad's internal optimizer used 0.01 with a ReduceLROnPlateau schedule, so part of the reported gap could be a hyperparameter artifact.
No gradient-descent-with-restarts baseline appears on the benchmarks, so the demonstrated advantage cannot separate the benefit of differentiability from the benefit of simply adding gradient access to a search procedure.
The paper itself notes that unfurled backpropagation through many generations is memory-hungry; extending to long-horizon tasks will likely require truncated backpropagation or learned meta-optimisers.

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.

Desk Editor's Note

read the letter

First thing you should know: this is a cleanly written, honest paper that builds something useful — a unified differentiable framework for PSO, GA, DE, and CMAES with learnable hyperparameters. The novelty is the integration, not the ingredients: pathwise reparameterization, Gumbel-Softmax selection, and straight-through estimators are all established. What's new is putting them all in one shared autodiff loop with the population itself as nn.Parameters. That is a genuinely useful contribution for people who want metaheuristics as differentiable modules in larger pipelines.

The paper does several things well. The description of the differentiable operators is clear and precise. The authors explicitly state their goal is to match or exceed classical counterparts, not beat state-of-the-art — that is the right scope. They also include a limitation paragraph about memory cost of unrolled optimization, which is more than many papers do. The empirical results are internally consistent: differentiable CMAES is strong, and differentiable PSO follows.

The soft spot is load-bearing. The unified loop in Section 2.2.1 (Listing 1) does loss.backward() and optimizer.step() on the population and hyperparameters before update_state(). That means every EvoGrad generation is really a hybrid that adds a first-order gradient step on the current population to the stochastic operators. The baselines are pure pymoo black-box methods with no gradient access and no restarts. So the reported advantage — e.g., differentiable CMAES at -93 to -104 vs -28.82 on 500D Michalewicz — could come mostly from that extra Adam step, not from making mutation or crossover differentiable. The wine regression comparison makes it worse: Adam with default PyTorch settings (lr=0.001) against differentiable CMAES with Adam lr=0.01 plus ReduceLROnPlateau. That is not a fair control. A standard GD-with-restarts baseline is missing. Without isolating gradients from the differentiable operators, the abstract's claim that differentiable versions 'consistently outperform' gradient-agnostic algorithms is not established.

Minor issues: no code or data released, no bias analysis for the gradient estimators, and the '3x faster on GPU vs CPU' runtime claim conflates hardware with method. The 'new standard' phrasing in the abstract exceeds the paper's own stated scope.

Bottom line: this paper deserves serious referee time — the framework is useful and clearly explained. But it needs fair baselines (ideally a gradient-descent-only baseline with the same optimizer settings, and restarts), code release, and a more measured claim. I'd encourage you to engage with it; it is a solid integration paper with an over-claimed evaluation.

Referee Report

4 major / 5 minor

Summary. The paper introduces EvoGrad, a unified differentiable framework that reformulates four population-based metaheuristics (PSO, GA, DE, CMAES) as differentiable programs. The method uses pathwise reparameterization, Gumbel-Softmax and Binary-Concrete relaxations, and stores the population and hyperparameters as learnable parameters. A single optimization loop (Section 2.2.1, Listing 1) interleaves the stochastic evolutionary operators with an outer Adam step that moves the population and hyperparameters along the gradient of the best fitness. Experiments on continuous benchmark functions (Ackley, Michalewicz, Rosenbrock, Griewank) and a wine-quality regression task claim that the differentiable variants consistently outperform their gradient-agnostic counterparts, with particularly strong results for differentiable CMAES. The paper concludes that the framework enables adaptive hyperparameter learning and opens the way to hybrid gradient-based and gradient-free optimization.

Significance. The idea of making entire metaheuristics differentiable is timely and could bridge evolutionary computation and deep learning, with potential applications in meta-learning and neural architecture search. The paper's formalization of a unified gradient-driven metaheuristic loop is clear, and the use of established reparameterization techniques is reasonable. However, the empirical evidence is currently confounded: the comparison does not isolate the effect of differentiable operators from the effect of the additional gradient descent step, so the central claim of consistent outperformance is not yet established. If the authors address this with controlled ablations, the framework could be a useful contribution.

major comments (4)

[Section 2.2.1, Listing 1] The unified loop applies an Adam step to the population matrix and all learnable parameters after loss.backward(), before update_state() commits the evolutionary changes. Consequently, every generation performs a direct gradient-descent step on the current individuals with respect to the objective, in addition to the stochastic evolutionary operators. The baselines in Section 3 are pure black-box metaheuristics with no gradient access and no restarts, so the reported improvements (e.g., the 500D Michalewicz results in Figure 2) could be caused entirely by this extra gradient component rather than by the differentiable reformulation of the operators. The manuscript lacks an ablation that uses the same outer Adam loop without evolutionary operators, or with non-differentiable operators, and therefore the abstract's central claim ('differentiable versions consistently outperform traditional, gradient-agnostic algorithms') is not supported by the experiments as they stand.
[Section 3, wine regression experiment] The comparison between Adam and differentiable CMAES is not controlled: Adam is run with default PyTorch settings (learning rate 0.001), while EvoGrad's CMAES uses Adam with learning rate 0.01 and a ReduceLROnPlateau schedule. In addition, '3000 epochs' for Adam is not a comparable budget to '3000 fitness evaluations' for CMAES, since an epoch involves many gradient updates. The reported ABF gap (5.66 +/- 0.39 vs 71.90 +/- 15.34) may therefore reflect optimizer tuning and budget differences rather than the differentiable metaheuristic formulation. The authors should match the learning-rate schedules and define a common evaluation budget (e.g., number of forward-backward passes or function evaluations) for both methods.
[Section 2.2] The paper states that the pathwise reparameterization 'produces an unbiased gradient estimate', but this claim does not cover the straight-through estimators used for the Gumbel-Sigmoid masks and Gumbel-Softmax selection in Sections 2.2.3 and 2.2.4, which are biased. Since these biased gradient estimates drive the learnable hyperparameters and population updates that are central to the method, the paper should either provide a bias analysis or explicitly acknowledge the biased nature of the discrete relaxations and discuss its consequences.
[Section 3, runtime comparison] The timing comparisons are not hardware-controlled: the differentiable variants run on an NVIDIA RTX 4090 GPU, while the pymoo baselines run on CPU (two Intel Xeon Gold 6152). The claim that EvoGrad is 'approximately 3x faster' and the 500D Michalewicz timings (1 hour vs 16 minutes for 30 runs) therefore conflate hardware differences with algorithmic efficiency. To support the scalability claim, the authors should compare both implementations on the same hardware, or report a hardware-independent measure such as function evaluations per second on a common device.

minor comments (5)

[Section 4] There is a typo in the conclusion: 'reparameterisedd' should be 'reparameterised'.
[Section 2.2.4] The differentiable DE uses Gumbel-Softmax with soft selection in the forward pass, which effectively replaces discrete parent sampling with a weighted average of all individuals; this is a fundamental change in algorithm semantics, not merely a relaxation, and should be discussed explicitly as a design choice with implications for the comparison.
[Figure 1] The boxplots show performance distributions but no statistical significance tests (e.g., Wilcoxon signed-rank) are reported; adding these would strengthen the claim of consistent improvement.
[Section 2.2.1] The paper claims 'all hyperparameters are fully learnable', but the Adam learning rate and ReduceLROnPlateau scheduler in Section 3 are fixed hyperparameters; the statement should be qualified.
[Section 3] The pymoo baseline configurations are described only as 'closely match our own implementations'; for reproducibility, the exact settings (population size, mutation/crossover rates, etc.) should be listed in a table or appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: EvoGrad's performance claims are empirical comparisons against external benchmarks, not consequences of definitions or self-citations.

full rationale

The paper does not present a derivation chain in which a predicted quantity is defined from the fitted data or from a load-bearing self-citation. EvoGrad is defined in Section 2 as a hybrid loop in which `loss.backward()` and `optimizer.step()` move the population along objective gradients before committing evolutionary state changes (Listing 1, Section 2.2.1). This makes its 'differentiable' variants hybrid gradient-plus-population methods, and the comparison against gradient-agnostic pymoo baselines is therefore not isolating the effect of differentiable operators alone. That is an experimental confound, not a circularity: the headline claim is an empirical outcome, not an equation that reduces to its inputs. The benchmarks (Ackley, Michalewicz, Rosenbrock, Griewank, Wine Quality) are external, the baselines are separate implementations from pymoo, and no parameter is fitted to a subset of data and then renamed as a prediction. The only author self-citation ([32]) appears in future work and is not load-bearing. Thus there is no self-definitional, fitted-input, or self-citation circularity; the central claim retains independent empirical content, even if the comparison fairness is questionable.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard reparameterization tricks from the deep learning literature and on the fairness of the experimental setup. The method introduces no new physical entities, but it relies on several hand-chosen settings and untested gradient estimator assumptions.

free parameters (4)

Initial learning rate = 0.01
Adam optimizer learning rate for all differentiable algorithms; hand-chosen, not tuned per task.
Learning rate scheduler = ReduceLROnPlateau, factor 0.5, patience 100
Hand-chosen schedule; reduces learning rate during run, affecting all results.
Population size = 100 (benchmarks), 30 (wine)
Chosen by authors; not swept.
Gumbel-Softmax / Binary-Concrete temperature = not reported
Temperature tau controls relaxation sharpness but is never specified, harming reproducibility.

assumptions (3)

standard math The reparameterization trick, Gumbel-Softmax, and straight-through estimators provide unbiased or useful gradients for the stochastic operators
Invoked in Section 2.2 as the basis for differentiability, with no proof of unbiasedness in this context.
domain assumption The pymoo baselines with default/canonical settings are adequate controls for the classical algorithms
The comparisons in Section 3 assume that the classical implementations represent standard performance; this is questionable because hyperparameters of baselines are fixed while EvoGrad adapts and uses a different optimizer.
ad hoc to paper The scalar loss (best fitness or softmax-weighted fitness) provides a valid gradient signal for updating the population
For PSO, GA, and DE the loss is the minimum fitness in the population; gradients through this non-smooth selection are not analyzed for bias.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EvoGrad: Metaheuristics in a Differentiable Wonderland." pith.science (2026). https://pith.science/paper/BNSNSEJU

@misc{pith2026250606320,
  author       = {Pith},
  title        = {Pith review of: EvoGrad: Metaheuristics in a Differentiable Wonderland},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNSNSEJU}},
  note         = {Machine review of arXiv:2506.06320}
}

read the original abstract

Differentiable programming has revolutionised optimisation by enabling efficient gradient-based training of complex models, such as Deep Neural Networks (NNs) with billions and trillions of parameters. However, traditional Evolutionary Computation (EC) and Swarm Intelligence (SI) algorithms, widely successful in discrete or complex search spaces, typically do not leverage local gradient information, limiting their optimisation efficiency. In this paper, we introduce EvoGrad, a unified differentiable framework that integrates EC and SI with gradient-based optimisation through backpropagation. EvoGrad converts conventional evolutionary and swarm operators (e.g., selection, mutation, crossover, and particle updates) into differentiable operators, facilitating end-to-end gradient optimisation. Extensive experiments on benchmark optimisation functions and training of small NN regressors reveal that our differentiable versions of EC and SI metaheuristics consistently outperform traditional, gradient-agnostic algorithms in most scenarios. Our results show the substantial benefits of fully differentiable evolutionary and swarm optimisation, setting a new standard for hybrid optimisation frameworks.

Figures

Figures reproduced from arXiv: 2506.06320 by the authors.

**Figure 2.** Convergence comparison between classical CMAES and our differentiable CMAES on the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

[1]

Transformer semantic genetic programming for symbolic regression, 2025

Philipp Anthes, Dominik Sobania, and Franz Rothlauf. Transformer semantic genetic programming for symbolic regression, 2025

work page 2025
[2]

Tutorial cma-es: evolution strategies and covariance matrix adaptation

Anne Auger and Nikolaus Hansen. Tutorial cma-es: evolution strategies and covariance matrix adaptation. In Proc. of the Genetic and Evolutionary Computation Conference Companion, pages 827–848, 2012

work page 2012
[3]

Handbook of evolutionary computation

Thomas Bäck, David B Fogel, and Zbigniew Michalewicz. Handbook of evolutionary computation. Release, 97(1):B1, 1997

work page 1997
[4]

Automatic differentiation in machine learning: a survey

Atilim Gunes Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: a survey. Journal of machine learning research, 18(153):1– 43, 2018

work page 2018
[5]

The elements of differentiable programming, 2024

Mathieu Blondel and Vincent Roulet. The elements of differentiable programming, 2024

work page 2024
[6]

Modeling wine preferences by data mining from physicochemical properties

Paulo Cortez, António Cerdeira, Fernando Almeida, Telmo Matos, and José Reis. Modeling wine preferences by data mining from physicochemical properties. Decision support systems, 47(4):547–553, 2009

work page 2009
[7]

Simulated binary crossover for continuous search space

Kalyanmoy Deb, Ram Bhushan Agrawal, et al. Simulated binary crossover for continuous search space. Complex systems, 9(2):115–148, 1995

work page 1995
[8]

Analysing mutation schemes for real-parameter genetic algorithms

Kalyanmoy Deb and Debayan Deb. Analysing mutation schemes for real-parameter genetic algorithms. International Journal of Artificial Intelligence and Soft Computing, 4(1):1–28, 2014

work page 2014

Show all 34 references

[9]

Swarm intelligence

Russell C Eberhart, Yuhui Shi, and James Kennedy. Swarm intelligence. Elsevier, 2001

2001
[10]

Gradient descent-particle swarm optimization based deep neural network predictive control of pressurized water reactor power

Derjew Ayele Ejigu and Xiaojing Liu. Gradient descent-particle swarm optimization based deep neural network predictive control of pressurized water reactor power. Progress in Nuclear Energy, 145:104108, 2022

2022
[11]

Improving evolutionary strategies with generative neural networks, 2019

Louis Faury, Clement Calauzenes, Olivier Fercoq, and Syrine Krichen. Improving evolutionary strategies with generative neural networks, 2019

2019
[12]

Continuous cartesian genetic programming based representation for multi-objective neural architecture search, 2023

Cosijopii Garcia-Garcia, Alicia Morales-Reyes, and Hugo Jair Escalante. Continuous cartesian genetic programming based representation for multi-objective neural architecture search, 2023

2023
[13]

Spidernet: Hybrid differentiable-evolutionary architecture search via train-free metrics, 2022

Rob Geada and Andrew Stephen McGough. Spidernet: Hybrid differentiable-evolutionary architecture search via train-free metrics, 2022

2022
[14]

How does selective mechanism improve self-attention networks? arXiv preprint arXiv:2005.00979, 2020

Xinwei Geng, Longyue Wang, Xing Wang, Bing Qin, Ting Liu, and Zhaopeng Tu. How does selective mechanism improve self-attention networks? arXiv preprint arXiv:2005.00979, 2020

2005 arXiv
[15]

Benchmarking a bi-population cma-es on the bbob-2009 function testbed

Nikolaus Hansen. Benchmarking a bi-population cma-es on the bbob-2009 function testbed. In Proc. of the Genetic and Evolutionary Computation Conference Companion, pages 2389–2396, 2009

2009
[16]

Completely derandomized self-adaptation in evolution strategies

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

2001
[17]

Regenerating soft robots through neural cellular automata, 2021

Kazuya Horibe, Kathryn Walker, and Sebastian Risi. Regenerating soft robots through neural cellular automata, 2021

2021
[18]

Differentiable genetic programming, 2016

Dario Izzo, Francesco Biscani, and Alessio Mereta. Differentiable genetic programming, 2016

2016
[19]

Categorical reparameterization with gumbel-softmax, 2017

Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax, 2017

2017
[20]

Particle swarm optimization

James Kennedy and Russell Eberhart. Particle swarm optimization. In Proceedings of ICNN’95- international conference on neural networks, volume 4, pages 1942–1948. ieee, 1995

1942
[21]

Kingma, Tim Salimans, and Max Welling

Diederik P. Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick, 2015

2015
[22]

Genetic programming as a means for programming computers by natural selection

John R Koza. Genetic programming as a means for programming computers by natural selection. Statistics and computing, 4:87–112, 1994

1994
[23]

Guiding evolutionary strategies by differentiable robot simulators, 2021

Vladislav Kurenkov and Bulat Maksudov. Guiding evolutionary strategies by differentiable robot simulators, 2021

2021
[24]

Evolution transformer: In-context evolutionary optimization

Robert Lange, Yingtao Tian, and Yujin Tang. Evolution transformer: In-context evolutionary optimization. In Proc. of the Genetic and Evolutionary Computation Conference Companion, pages 575–578, 2024

2024
[25]

Neural network architecture search with differentiable cartesian genetic programming for regression

Marcus Märtens and Dario Izzo. Neural network architecture search with differentiable cartesian genetic programming for regression. InProc. of the Genetic and Evolutionary Computation Conference Companion, pages 181–182, 2019

2019
[26]

An introduction to genetic algorithms

Melanie Mitchell. An introduction to genetic algorithms. MIT press, 1998

1998
[27]

d (tree)-by-dx: Automatic and exact differentiation of genetic programming trees

Peter Rockett, Yuri Kaszubowski Lopes, Tiantian Dou, and Elizabeth A Hathway. d (tree)-by-dx: Automatic and exact differentiation of genetic programming trees. In Hybrid Artificial Intelligent Systems, pages 133–144. Springer, 2019. 10

2019
[28]

Alice’s adventures in a differentiable wonderland – volume i, a tour of the land, 2024

Simone Scardapane. Alice’s adventures in a differentiable wonderland – volume i, a tour of the land, 2024

2024
[29]

Differential evolution–a simple and efficient heuristic for global optimiza- tion over continuous spaces

Rainer Storn and Kenneth Price. Differential evolution–a simple and efficient heuristic for global optimiza- tion over continuous spaces. Journal of global optimization, 11:341–359, 1997

1997
[30]

Evolution and learning in differentiable robots, 2024

Luke Strgar, David Matthews, Tyler Hummer, and Sam Kriegman. Evolution and learning in differentiable robots, 2024

2024
[31]

Success-history based parameter adaptation for differential evolution

Ryoji Tanabe and Alex Fukunaga. Success-history based parameter adaptation for differential evolution. In Proc. of IEEE Congress on Evolutionary Computation, pages 71–78. IEEE, 2013

2013
[32]

A modified eacop im- plementation for real-parameter single objective optimization problems

Andrea Tangherloni, Vasco Coelho, Francesca M Buffa, and Paolo Cazzaniga. A modified eacop im- plementation for real-parameter single objective optimization problems. In Proc. of IEEE Congress on Evolutionary Computation, pages 01–08. IEEE, 2024

2024
[33]

Differentiable genetic programming for high-dimensional symbolic regression, 2023

Peng Zeng, Xiaotian Song, Andrew Lensen, Yuwei Ou, Yanan Sun, Mengjie Zhang, and Jiancheng Lv. Differentiable genetic programming for high-dimensional symbolic regression, 2023

2023
[34]

An efficient hybrid particle swarm and gradient descent method for the estimation of the hosting capacity of photovoltaics by distribution networks

Esau Zulu, Ryoichi Hara, and Hiroyuki Kita. An efficient hybrid particle swarm and gradient descent method for the estimation of the hosting capacity of photovoltaics by distribution networks. Energies, 16(13):5207, 2023. 11

2023

Pith tools

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

[1] [1]

Transformer semantic genetic programming for symbolic regression, 2025

Philipp Anthes, Dominik Sobania, and Franz Rothlauf. Transformer semantic genetic programming for symbolic regression, 2025

work page 2025

[2] [2]

Tutorial cma-es: evolution strategies and covariance matrix adaptation

Anne Auger and Nikolaus Hansen. Tutorial cma-es: evolution strategies and covariance matrix adaptation. In Proc. of the Genetic and Evolutionary Computation Conference Companion, pages 827–848, 2012

work page 2012

[3] [3]

Handbook of evolutionary computation

Thomas Bäck, David B Fogel, and Zbigniew Michalewicz. Handbook of evolutionary computation. Release, 97(1):B1, 1997

work page 1997

[4] [4]

Automatic differentiation in machine learning: a survey

Atilim Gunes Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: a survey. Journal of machine learning research, 18(153):1– 43, 2018

work page 2018

[5] [5]

The elements of differentiable programming, 2024

Mathieu Blondel and Vincent Roulet. The elements of differentiable programming, 2024

work page 2024

[6] [6]

Modeling wine preferences by data mining from physicochemical properties

Paulo Cortez, António Cerdeira, Fernando Almeida, Telmo Matos, and José Reis. Modeling wine preferences by data mining from physicochemical properties. Decision support systems, 47(4):547–553, 2009

work page 2009

[7] [7]

Simulated binary crossover for continuous search space

Kalyanmoy Deb, Ram Bhushan Agrawal, et al. Simulated binary crossover for continuous search space. Complex systems, 9(2):115–148, 1995

work page 1995

[8] [8]

Analysing mutation schemes for real-parameter genetic algorithms

Kalyanmoy Deb and Debayan Deb. Analysing mutation schemes for real-parameter genetic algorithms. International Journal of Artificial Intelligence and Soft Computing, 4(1):1–28, 2014

work page 2014