Pith. sign in

REVIEW 4 major objections 6 minor 45 references

Transformer Semantic Genetic Programming for Symbolic Regression

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A transformer trained once on synthetic semantically similar function pairs can replace genetic programming's variation operators and match or beat four baselines on the tested black-box regression problems.

desk verdict TSGP is a genuinely new semantic GP operator using a pretrained transformer, with mostly supported empirical claims; its main weakness is the narrow training distribution and missing code/data, both addressable. read the letter →

arxiv 2501.18479 v1 pith:KBOQDM4L submitted 2025-01-30 cs.NE

Philipp Anthes , Dominik Sobania , Franz Rothlauf This is my paper · ORCID
classification cs.NE
keywords geneticprogrammingsymbolicregressionsemanticsearchoperatortransformersimilaritysolutionsizeblack-box
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces TSGP, a genetic programming variant that replaces the usual syntactic crossover and mutation operators with a transformer trained to produce offspring that are semantically similar to their parents. The transformer is trained once on pairs of functions taken from standard GP runs on synthetic regression problems, where semantic similarity is measured as the Euclidean distance between the functions' output vectors on standardized inputs. On five black-box symbolic regression benchmarks, TSGP reports prediction quality comparable to or significantly better than standard GP, SLIM_GSGP, DSR, and DAE-GP, while producing statistically smaller solutions than SLIM_GSGP. The central claim is that a single pretrained transformer can serve as a transferable semantic search operator, and that the semantic control it provides explains the improved search behavior.

What carries the argument

The central object is a conditional transformer used as a learned semantic variation operator. It follows the standard encoder–decoder architecture, with the encoder reading a parent function's token sequence and the decoder sampling an offspring token by token; the desired semantic distance is given as an additional input to both sides, so the step size of the search can be controlled. The training set is built by running standard GP on synthetic four-dimensional linear regression problems, evaluating each function on random standardized inputs, and pairing every function with its k-nearest neighbors (k=3) in Euclidean distance between output vectors. Syntax control keeps sampled token sequences valid expression trees. This training pipeline is what teaches the transformer to map a parent to a semantically similar offspring, and it is the mechanism that distinguishes TSGP from variation operators based on linear combinations.

What would settle it

Run TSGP on symbolic regression benchmarks whose target functions contain structures absent from the synthetic training set, such as trigonometric or piecewise terms, and measure both the median parent–offspring semantic distance and the test RMSE; if the offspring are no more semantically similar than stdGP's and the prediction quality no longer improves, the transferability claim is not supported.

Watch

Extended reading notes

Core claim

Train a transformer once on synthetic parent–offspring pairs with small semantic distance, then use it as the variation operator of a GP run: this is TSGP's claim. The parent is encoded as a prefix-order token sequence, the desired semantic distance is supplied as an extra input, and the decoder autoregressively samples a structurally valid offspring. Because the training pairs come from nearest neighbors in output space, the model learns to preserve behavior while allowing flexible structural changes. On ERA, ESL, and LEV, TSGP achieves significantly lower test RMSE than stdGP, SLIM_GSGP, DSR, and DAE-GP; on Galaxy and Pollen it is comparable. TSGP also produces significantly smaller solutions than SLIM_GSGP on all five datasets, which the authors attribute to the transformer generating semantic similarity through many different structures rather than through the linear combinations that inflate geometric-semantic trees.

Load-bearing premise

The method assumes that semantic similarity learned from pairs of functions generated by standard GP on synthetic linear-regression problems transfers to unseen datasets with different data-generating structures; if that learned notion of similarity does not carry over, the TSGP advantage disappears.

Editorial extensions

If this is right

  • On ERA, ESL, and LEV, TSGP reaches high-quality solutions within roughly 10 generations, faster than stdGP and SLIM_GSGP, so semantic pre-training can accelerate GP convergence.
  • TSGP produces statistically smaller solutions than SLIM_GSGP on all tested datasets, with reductions up to about 42% on Pollen, so a learned semantic operator can avoid geometric-semantic bloat.
  • Since the transformer is trained once on synthetic problems and then applied to unseen datasets with no additional training, the approach implies that a single pre-trained model can transfer a semantic variation policy across symbolic regression problems.
  • The variation analysis shows TSGP offspring are semantically more similar to their parents than stdGP offspring on most datasets, implying that controlling semantic distance reduces erratic behavior-changing mutations.

Reading between the lines

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

  • Beyond the paper, the same transformer-plus-semantic-distance recipe could in principle be applied to other program synthesis tasks, provided a semantic distance and a synthetic training distribution are defined for that domain.
  • The synthetic training distribution is limited to four-feature linear models; broadening it to trigonometric, exponential, or nested functions might improve transfer on datasets like Galaxy and Pollen, where TSGP is only comparable to the baselines.
  • The Euclidean distance on random standardized inputs is one possible semantic proxy; a task-aware distance weighted by the actual training samples could change which parent–offspring pairs are treated as similar and might strengthen or alter the search step-size control.
  • The comparison with DSR and DAE-GP uses a fixed 50-generation budget, and those methods adapt their networks during search; a longer budget could narrow the gap, since TSGP's model is fixed after training.
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 introduces Transformer Semantic Genetic Programming (TSGP), a genetic programming variant in which a transformer is trained once on synthetic function pairs and then used as the sole variation operator. Training pairs are generated by running stdGP on 50 synthetic four-feature linear regression problems, computing semantics as output vectors on standardized random inputs, and using k-NN (k=3) to identify semantically similar functions; the transformer is conditioned on the desired semantic distance. In experiments on five PMLB black-box regression datasets (ERA, ESL, Galaxy, LEV, Pollen), TSGP is compared with stdGP, SLIM_GSGP, DSR, and DAE-GP over 30 runs with a fixed budget of 50 generations. The paper reports that TSGP achieves significantly better test RMSE on ERA, ESL, and LEV, is not significantly different from the best method on Galaxy and Pollen, produces significantly smaller solutions than SLIM_GSGP, and generates offspring with lower semantic distance than stdGP and SLIM_GSGP on most datasets.

Significance. If the transferability claim is accepted, the paper makes a useful contribution: it shows that a single pretrained transformer can act as a semantic variation operator that is competitive with or better than established GP and deep-learning baselines while controlling solution size. The experimental design has strengths: 30 independent runs per method, standardized datasets, held-out test splits, and Wilcoxon tests. The separation between synthetic training data and target problems avoids the circularity that would arise if test data were used in training. However, the evidence base is narrow (five datasets, all four-feature), the baselines DSR and DAE-GP are not configured in their recommended regimes, and key hyperparameters such as SD_d are not varied. These gaps currently limit the strength of the central generalization claim.

major comments (4)
  1. [§3.1, §4.1, §4.4] The central transferability claim is not adequately supported. The transformer is trained on functions evolved by stdGP on 50 synthetic four-feature linear-regression problems (§4.1), and all five test problems also have exactly four features. No experiment uses a training distribution that includes nonlinear target functions or test problems with different dimensionality, so the claim in the abstract that TSGP can be used for 'unseen and unknown problems' rests on an untested alignment between the synthetic training distribution and the target problems. The paper's own results show no significant advantage on Galaxy and Pollen, and §4.4 attributes this to structures the transformer 'can generate only semantically similar structures to a limited extent.' I recommend either evaluating on datasets with varying dimensionality and stronger nonlinearity/correlation, or providing a direct analysis of the distribution shift between synthetic training functions and target functions.
  2. [§4.1, Table 2] The comparison to DSR and DAE-GP is not conducted in a way that supports the strong claim that TSGP significantly outperforms them. DSR is run with the default configuration, with batch size set to 100 and 50 iterations, while DAE-GP is run with only 5% edit percentage and no reported tuning; both methods are designed to be used with more extensive training (see refs. [17, 26]). The resulting uniformly poor performance of these baselines in Table 2 may be an artifact of an unreasonably small budget. The authors should use recommended settings or perform a small hyperparameter search, and then either update Table 2 or restrict the claim to the specific fixed-budget comparison.
  3. [§3.2, §4.1] The desired semantic distance SD_d is the main control parameter of the TSGP operator (§3.2), but it is fixed at 0.1 with no sensitivity analysis. Because the entire method is built around generating offspring at a target semantic distance, the results could depend strongly on this value and on the training-data construction parameters (k=3, SD<100 filter). At minimum, the authors should vary SD_d on at least one or two datasets and show that the qualitative ranking and significance pattern are stable.
  4. [§4.3, Table 3] Section 4.3 and Table 3 are internally inconsistent. The text states that 'no statistical differences in solution size were observed between TSGP and stdGP,' but in the ESL row of Table 3 the stdGP value is marked with superscript 'a,' which by the table's stated convention indicates a significant difference from TSGP. Either the superscript convention needs to be defined precisely (e.g., one-sided tests for 'significantly smaller than'), or the claim in Section 4.3 must be qualified by dataset.
minor comments (6)
  1. [Abstract] The abstract says 'trained on synthetic test problems' but this should read 'synthetic problems' or 'synthetic training problems.'
  2. [Table 2 and Table 3 captions] The superscript notation in Tables 2 and 3 is not fully defined in the captions; please specify the test direction (one-sided vs. two-sided) and whether any multiple-comparison correction was applied.
  3. [§4.2] The sentence 'Our results confirm the findings of Vanneschi [30]' should cite the original GSGP paper [23] rather than the SLIM_GSGP paper, which is the reference currently listed as [30].
  4. [Figure 4 caption] The caption should state explicitly that the Euclidean distance is computed between parent and offspring semantics on the test set of the target problem.
  5. [§4.1] The transformer is trained for only 8 epochs on 5 million pairs; please clarify whether this is 8 passes over the full dataset and report the training time and hardware used.
  6. [Reproducibility] The paper does not state whether the source code and the trained transformer will be made available; please include a reproducibility statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TSGP's transformer is trained on synthetic problems and tested on held-out PMLB data; the main RMSE comparison is independent of the training signal.

full rationale

The central derivation is self-contained. The transformer is trained once on 5 million semantically similar function pairs produced by stdGP on 50 synthetic linear-regression problems (Section 3.1 and Section 4.1), before any PMLB test problem is seen. The reported test RMSE values are computed on held-out 50/50 splits (Section 4.1), and no test-set output is used to fit the transformer or to set SD_d=0.1, which is chosen a priori as a step-size control (Section 3.2). The main claim—that TSGP achieves comparable or better prediction quality than the baselines—therefore is an externally falsifiable empirical result rather than a reduction to the training data. The variation-behavior analysis in Section 4.4 measures parent-offspring Euclidean semantic distance on the target test sets; while this is aligned with the training objective, it is a generalization check, not the source of the RMSE comparisons, and the paper reports that it does not hold on Pollen. Self-citations (e.g., references [2], [9], and [36]-[39]) support baselines, significance-test conventions, and implementation details; none is load-bearing for the TSGP transfer claim, and there is no imported uniqueness theorem or ansatz. The only caveats are external-validity limitations acknowledged by the authors, such as the training dimensionality matching the five four-feature datasets and the absence of significant advantage on Galaxy and Pollen; these are correctness risks, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The ledger shows that TSGP introduces no new physical or mathematical entities. The method depends on several hand-chosen hyperparameters and domain assumptions about semantic-distance proxies and training-data coverage. The most important free parameters are the desired semantic distance SD_d and the training-data construction settings, because they directly shape the variation operator and are not supported by sensitivity analysis.

free parameters (6)
  • desired semantic distance SD_d = 0.1
    Conditioning input used at sampling to control offspring semantic similarity; chosen by hand with no sensitivity analysis; it directly controls the variation step size and thus the search behavior.
  • k in k-NN semantic similarity search = 3
    Number of nearest semantic neighbors used to build training pairs; chosen by hand in Section 3.1.
  • semantic distance threshold = SD < 100
    Filter for training pairs in Section 4.1; chosen by hand to exclude very dissimilar pairs, affecting the training distribution.
  • number of synthetic training problems = 50
    stdGP runs on 50 random synthetic problems to generate training functions; chosen by hand without a stated justification.
  • synthetic dataset noise = Gaussian noise (level not specified)
    Noise applied to synthetic datasets to increase diversity; amount not given, which affects the training distribution and is not reproducible from the text.
  • transformer architecture and training = 2 layers, 8 heads, hidden 128, max len 100, 8 epochs, lr 1e-3
    Model capacity and training regimen chosen without systematic tuning (stated in Section 4.1); these affect what semantic similarities the model can learn.
assumptions (5)
  • domain assumption Semantic similarity measured by Euclidean distance between output vectors on standardized random inputs is a reliable proxy for behavioral similarity on target problems.
    The training-signal construction (Section 3.1) and the search-dynamic analysis (Section 4.4) rely entirely on this proxy. If the proxy misaligns with target distributions, the method's advantage may disappear.
  • domain assumption The stdGP-generated functions on 50 synthetic linear-regression problems cover the search space relevant to the real-world test problems.
    The transformer can only learn to vary functions from the distribution produced by stdGP on synthetic problems; if this distribution is too narrow, transfer to PMLB problems will fail. Stated in Section 3.1.
  • domain assumption A transformer can learn semantic similarity from token sequences of parse trees and generalize to unseen functions.
    This is the core hypothesis of the paper; it is tested empirically on five datasets but not proven or bounded by any theory.
  • domain assumption Protected division (returning 1 for division by zero) preserves semantics well enough for training and search.
    The primitive set includes protected division (Section 4.1), which can distort semantics near division-by-zero cases and may affect the semantic distance computations.
  • standard math Standard transformer architecture, prefix-order tokenization, and syntax control work as in prior work.
    The method relies on the transformer architecture of Vaswani et al. and the syntax control of Wittenberg et al.; these are accepted as background from the cited literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transformer Semantic Genetic Programming for Symbolic Regression." pith.science (2026). https://pith.science/paper/KBOQDM4L

@misc{pith2026250118479,
  author       = {Pith},
  title        = {Pith review of: Transformer Semantic Genetic Programming for Symbolic Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KBOQDM4L}},
  note         = {Machine review of arXiv:2501.18479}
}
read the original abstract

In standard genetic programming (stdGP), solutions are varied by modifying their syntax, with uncertain effects on their semantics. Geometric-semantic genetic programming (GSGP), a popular variant of GP, effectively searches the semantic solution space using variation operations based on linear combinations, although it results in significantly larger solutions. This paper presents Transformer Semantic Genetic Programming (TSGP), a novel and flexible semantic approach that uses a generative transformer model as search operator. The transformer is trained on synthetic test problems and learns semantic similarities between solutions. Once the model is trained, it can be used to create offspring solutions with high semantic similarity also for unseen and unknown problems. Experiments on several symbolic regression problems show that TSGP generates solutions with comparable or even significantly better prediction quality than stdGP, SLIM_GSGP, DSR, and DAE-GP. Like SLIM_GSGP, TSGP is able to create new solutions that are semantically similar without creating solutions of large size. An analysis of the search dynamic reveals that the solutions generated by TSGP are semantically more similar than the solutions generated by the benchmark approaches allowing a better exploration of the semantic solution space.

Figures

Figures reproduced from arXiv: 2501.18479 by the authors.

Figure 1
Figure 1. Model Building of TSGP. First, solutions are generated and their semantics are determined. Subsequently, solutions [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Median training RMSE of the best solution of TSGP, stdGP, SLIM_GSGP over generations for all analyzed black-box [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Median solution size of the best solution of TSGP, stdGP, SLIM_GSGP over generations for all analyzed black-box data [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Median Euclidean Distance between the semantics [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 28 canonical work pages

  1. [1]

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Gi- ambattista Parascandolo. 2021. Neural Symbolic Regression that Scales. http: //arxiv.org/abs/2106.06427 arXiv:2106.06427 [cs]

    arXiv 2021

  2. [2]

    Martin Briesch, Dominik Sobania, and Franz Rothlauf. 2024. Large Language Models Suffer From Their Own Output: An Analysis of the Self-Consuming Training Loop. arXiv:2311.16822 [cs.LG] https://arxiv.org/abs/2311.16822

    arXiv 2024

  3. [3]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    arXiv 2020

  4. [4]

    Mauro Castelli, Leonardo Vanneschi, and Sara Silva. 2013. Prediction of high performance concrete strength using Genetic Programming with geometric semantic genetic operators. Expert Systems with Applications 40, 17 (2013), 6856–

    work page 2013

  5. [5]

    François Chollet et al. 2015. Keras. https://keras.io

    2015

  6. [6]

    Grant Dick, Caitlin A Owen, and Peter A Whigham. 2020. Feature standardisation and coefficient optimisation for effective symbolic regression. In Proceedings of the 2020 Genetic and Evolutionary Computation Conference . 306–314

    work page 2020

  7. [7]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. arXiv:2401.08281 [cs.LG] https://arxiv.org/abs/2401.08281

    arXiv 2024

  8. [8]

    Félix-Antoine Fortin, François-Michel De Rainville, Marc-André Gardner Gard- ner, Marc Parizeau, and Christian Gagné. 2012. DEAP: Evolutionary algorithms made easy. The Journal of Machine Learning Research 13, 1 (2012), 2171–2175

    work page 2012

Show all 45 references
  1. [9]

    Alina Geiger, Dominik Sobania, and Franz Rothlauf. 2023. Down-sampled epsilon- lexicase selection for real-world symbolic regression problems. In Proceedings of the Genetic and Evolutionary Computation Conference . 1109–1117

    2023

  2. [10]

    Pierre-Alexandre Kamienny, Stéphane d’Ascoli, Guillaume Lample, and François Charton. 2022. End-to-end symbolic regression with transformers. Advances in Neural Information Processing Systems 35 (2022), 10269–10281

    2022

  3. [11]

    Pierre-Alexandre Kamienny, Guillaume Lample, Sylvain Lamprier, and Marco Virgolin. 2023. Deep generative symbolic regression with Monte-Carlo-tree- search. In International Conference on Machine Learning . PMLR, 15655–15668

    2023

  4. [12]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Opti- mization. arXiv:1412.6980 [cs.LG] https://arxiv.org/abs/1412.6980

    2017 arXiv

  5. [13]

    JRGP Koza. 1992. On the programming of computers by means of natural selection. Genetic programming (1992)

    1992

  6. [14]

    Krzysztof Krawiec and Pawel Lichocki. 2009. Approximating geometric crossover in semantic space. In Proceedings of the 11th Annual conference on Genetic and evolutionary computation. 987–994

    2009

  7. [15]

    Krzysztof Krawiec and Tomasz Pawlak. 2013. Approximating geometric crossover by semantic backpropagation. GECCO 2013 - Proceedings of the 2013 Genetic and Evolutionary Computation Conference, 941–948. https://doi.org/10.1145/2463372. 2463483

    2013 doi

  8. [16]

    William La Cava, Bogdan Burlacu, Marco Virgolin, Michael Kommenda, Patryk Orzechowski, Fabrício Olivetti de França, Ying Jin, and Jason H Moore. 2021. Contemporary symbolic regression methods and their relative performance. Advances in neural information processing systems 202...

    2021

  9. [17]

    Mikel Landajuela, Chak Shing Lee, Jiachen Yang, Ruben Glatt, Claudio P San- tiago, Ignacio Aravena, Terrell Mundhenk, Garrett Mulcahy, and Brenden K Petersen. 2022. A Unified Framework for Deep Symbolic Regression. In Advances in Neural Information Processing Systems , S. Koye...

    2022

  10. [18]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs.LG] https://arxiv.org/abs/1711.05101

    2019 arXiv

  11. [19]

    Sean Luke and Liviu Panait. 2002. Fighting Bloat with Nonparametric Parsimony Pressure. In Parallel Problem Solving from Nature — PPSN VII , Juan Julián Merelo Guervós, Panagiotis Adamidis, Hans-Georg Beyer, Hans-Paul Schwefel, and José- Luis Fernández-Villacañas (Eds.). Sprin...

    2002

  12. [20]

    Joao Francisco B. S. Martins, Luiz Otavio V. B. Oliveira, Luis F. Miranda, Felipe Casadei, and Gisele L. Pappa. 2018. Solving the exponential growth of symbolic regression trees in geometric semantic genetic programming. In Proceedings of the Genetic and Evolutionary Computati...

    2018

  13. [21]

    James McDermott, Alexandros Agapitos, Anthony Brabazon, and Michael O’Neill

  14. [22]

    Nicholas Freitag McPhee, Brian Ohs, and Tyler Hutchison. 2008. Semantic Building Blocks in Genetic Programming. In Genetic Programming , Michael O’Neill, Leonardo Vanneschi, Steven Gustafson, Anna Isabel Esparcia Alcázar, Ivanoe De Falco, Antonio Della Cioppa, and Ernesto Tara...

    2008

  15. [23]

    Alberto Moraglio, Krzysztof Krawiec, and Colin G. Johnson. 2012. Geometric Semantic Genetic Programming. In Parallel Problem Solving from Nature - PPSN XII, Carlos A. Coello Coello, Vincenzo Cutello, Kalyanmoy Deb, Stephanie Forrest, Giuseppe Nicosia, and Mario Pavone (Eds.). ...

    2012 doi

  16. [24]

    Quang Uy Nguyen, Nguyen Hoai, Michael O’Neill, Robert McKay, and Edgar Galván-López. 2011. Semantically-based crossover in genetic programming: Ap- plication to real-valued symbolic regression. Genetic Programming and Evolvable Machines 12 (06 2011), 91–119. https://doi.org/10...

    2011 doi

  17. [25]

    Caitlin A Owen, Grant Dick, and Peter A Whigham. 2018. Feature standardisation in symbolic regression. In Australasian Joint Conference on Artificial Intelligence . Springer, 565–576

    2018

  18. [26]

    Petersen, Mikel Landajuela, T

    Brenden K. Petersen, Mikel Landajuela, T. Nathan Mundhenk, Claudio P. Santiago, Soo K. Kim, and Joanne T. Kim. 2021. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients. http: //arxiv.org/abs/1912.04871 arXiv:1912.04871 [cs, stat]

    2021 arXiv

  19. [27]

    Joseph D Romano, Trang T Le, William La Cava, John T Gregg, Daniel J Gold- berg, Praneel Chakraborty, Natasha L Ray, Daniel Himmelstein, Weixuan Fu, and Jason H Moore. 2021. PMLB v1.0: an open source dataset collection for benchmarking machine learning methods. arXiv preprint ...

    2021 arXiv

  20. [28]

    Parshin Shojaee, Kazem Meidani, Amir Barati Farimani, and Chandan Reddy

  21. [29]

    Mojtaba Valipour, Bowen You, Maysum Panju, and Ali Ghodsi. 2021. Sym- bolicGPT: A Generative Transformer Model for Symbolic Regression. http: //arxiv.org/abs/2106.14131 arXiv:2106.14131 [cs]

    2021 arXiv

  22. [30]

    Leonardo Vanneschi. 2024. SLIM_GSGP: The Non-bloating Geometric Semantic Genetic Programming. In Genetic Programming , Mario Giacobini, Bing Xue, and Luca Manzoni (Eds.). Springer Nature Switzerland, Cham, 125–141. https: //doi.org/10.1007/978-3-031-56957-9_8

    2024 doi

  23. [31]

    Leonardo Vanneschi, Mauro Castelli, Luca Manzoni, and Sara Silva. 2013. A new implementation of geometric semantic GP and its application to problems in pharmacokinetics. 205–216 pages

    2013

  24. [32]

    Leonardo Vanneschi, Mauro Castelli, and Sara Silva. 2014. A survey of semantic methods in genetic programming. Genetic Programming and Evolvable Machines 15 (2014), 195–214

    2014

  25. [33]

    Leonardo Vanneschi, Sara Silva, Mauro Castelli, and Luca Manzoni. 2014. Geomet- ric semantic genetic programming for real life applications.Genetic programming theory and practice xi (2014), 191–209

    2014

  26. [34]

    Martin Vastl, Jonáš Kulhánek, Jiří Kubalík, Erik Derner, and Robert Babuška

  27. [35]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL] https://arxiv.org/abs/1706.03762

    2023 arXiv

  28. [36]

    David Wittenberg. 2022. Using Denoising Autoencoder Genetic Programming to Control Exploration and Exploitation in Search. In EuroGP 2022: Proceedings of the 25th European Conference on Genetic Programming (LNCS, Vol. 13223) , Eric Medvet, Gisele Pappa, and Bing Xue (Eds.). Sp...

    2022

  29. [37]

    David Wittenberg and Franz Rothlauf. 2023. Small Solutions for Real-World Symbolic Regression Using Denoising Autoencoder Genetic Programming. In Genetic Programming (Lecture Notes in Computer Science) , Gisele Pappa, Mario Giacobini, and Zdenek Vasicek (Eds.). Springer Nature...

    2023

  30. [38]

    David Wittenberg, Franz Rothlauf, and Christian Gagné. 2023. Denoising autoen- coder genetic programming: strategies to control exploration and exploitation in search. Genetic Programming and Evolvable Machines 24, 2 (Nov. 2023), 17. https://doi.org/10.1007/s10710-023-09462-2

    2023 doi

  31. [39]

    David Wittenberg, Franz Rothlauf, and Dirk Schweim. 2020. DAE-GP: de- noising autoencoder LSTM networks as probabilistic models in estimation of distribution genetic programming. In Proceedings of the 2020 Genetic and Evo- lutionary Computation Conference . ACM, Cancún Mexico,...

    2020

  32. [116]

    https://doi.org/10.1007/978-3-031-29573-7_7

    doi

  33. [117]

    https://doi.org/doi:10.1007/978-3-031-02056-8_7 Best paper nomination

    doi

  34. [2014]

    InApplications of Evolutionary Computation: 17th European Conference, EvoApplications 2014, Granada, Spain, April 23-25, 2014, Revised Selected Papers 17

    Geometric semantic genetic programming for financial data. InApplications of Evolutionary Computation: 17th European Conference, EvoApplications 2014, Granada, Spain, April 23-25, 2014, Revised Selected Papers 17 . Springer, 215–226

    2014

  35. [2023]

    Advances in Neural Information Processing Systems 36 (2023), 45907–45919

    Transformer-based planning for symbolic regression. Advances in Neural Information Processing Systems 36 (2023), 45907–45919

    2023

  36. [2024]

    IEEE Access 12 (2024), 37840–37849

    SymFormer: End-to-End Symbolic Regression Using Transformer-Based Architecture. IEEE Access 12 (2024), 37840–37849. https://doi.org/10.1109/ ACCESS.2024.3374649

    2024

  37. [6862]

    https://doi.org/10.1016/j.eswa.2013.06.037

    2013 doi

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.