Pith. sign in

REVIEW 4 major objections 5 minor 52 references

On the Generalisation of Koopman Representations for Chaotic System Control

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a Koopman embedding of a chaotic system, learned once for next-state prediction, can be reused for a downstream safety-control task with no loss of accuracy when the pretrained transformer is frozen.

desk verdict Task design sinks the transfer claim: the safety function depends only on q, so the frozen vs fine-tuned result shows the transformer context is irrelevant, not that the representation is reusable. read the letter →

arxiv 2508.18954 v1 pith:Z4VLV6OR submitted 2025-08-26 cs.LG

Kyriakos Hjikakou (1) , Juan Diego Cardenas Cartagena (1) , Matthia Sabatelli (1) ((1) University of Groningen , Department of Artificial Intelligence , Groningen , Netherlands) This is my paper
classification cs.LG
keywords KoopmanoperatorchaoticdynamicalsystemsLorenzsystemtransferlearningtransformersafetyfunctionrepresentationphysics-informedmachine
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

This paper argues that a Koopman embedding of a chaotic system, learned once from next-state prediction, carries enough dynamical structure to support a different downstream task—estimating a global safety function for control—without retraining the sequence model. The authors test this on the Lorenz system through a three-stage pipeline: a Koopman autoencoder creates a latent space where dynamics are approximately linear; a transformer is pretrained to predict next states in that space; and only a small safety head is trained on top of the frozen transformer. They report that this frozen configuration matches the fine-tuned one, and that both beat PCA-based embeddings on safety-function accuracy. The point of the claim is that physics-informed representations can be amortized across tasks, much like pretrained language models are reused in NLP.

What carries the argument

The load-bearing object is the learned Koopman operator K, a 32-by-32 matrix decomposed into a diagonal part D (growth and decay) plus a banded skew-symmetric part S_band (rotation), so that the encoder maps a state s_t to a latent y_t and the next latent is K y_t. This linearising structure is what makes the transformer's next-state pretraining capture durable dynamical content; at transfer time the encoder and transformer are frozen and only a multilayer safety head takes the transformer's final hidden state concatenated with the query state to predict U(q).

What would settle it

Give the frozen and fine-tuned backbones a much larger evaluation set (thousands of test trajectories, multiple seeds) and compute a confidence interval for the difference in safety-function MSE. If the interval excludes zero or crosses a pre-specified equivalence bound, the no-degradation claim fails. A second check: replace the Koopman encoder with an unconstrained autoencoder of the same capacity; if it transfers equally well, the Koopman structure is not the cause.

Watch

Extended reading notes

Core claim

On the Lorenz system, embeddings produced by a structured Koopman autoencoder transfer from short-horizon next-state prediction to a long-horizon safety function that quantifies minimum control effort to stay in a safe region. The paper's central quantitative result is that a transformer frozen after pretraining and a transformer fine-tuned on the safety task perform indistinguishably (Wilcoxon p = 0.335), while both significantly outperform standard PCA and physics-informed PCA baselines. The authors interpret this as evidence that the Koopman latent space encodes reusable dynamical structure—the attractor's geometry and long-term behaviour—rather than task-specific patterns learned for pre

Load-bearing premise

The claim that freezing the transformer costs nothing rests on treating a non-significant Wilcoxon p-value (0.335) as evidence of equivalence; with 252 trajectories and error variances comparable to the observed differences, the test may simply lack the power to detect a practically meaningful drop.

Editorial extensions

If this is right

  • If true, a single pretrained Koopman backbone can serve multiple downstream tasks on the same system, so the cost of learning a physics-informed representation is paid once.
  • Frozen-backbone transfer means safety heads can be trained cheaply and with less data; the authors report a 4-layer frozen transformer uses less than half the peak GPU power of the 11-layer PCA baseline during fine-tuning.
  • Koopman embeddings should be preferred over PCA-derived features when downstream tasks require global, long-horizon understanding rather than local reconstruction.
  • The result supports treating next-state prediction as a viable pretext task for physical systems, analogous to next-token prediction in language models.
  • The comparison validates the structured Koopman operator design (diagonal plus skew-symmetric bands) as the source of the advantage, rather than the transformer architecture alone.

Reading between the lines

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

  • The equal-frozen claim rests on accepting a null result; a higher-powered test with more trajectories or repeated seeds and pre-specified equivalence bounds could reveal a real degradation, since the paper does not report such a test.
  • Because the safety region Q covers only the right lobe of the attractor, the measured transfer may be specific to that region; testing the left lobe or the full attractor would show whether the embedding's geometry generalises globally.
  • The comparison to PCA may be generous to Koopman: the physics-informed PCA baseline includes time derivatives, which the paper itself notes can produce phantom oscillations, so a stronger baseline would be an unconstrained nonlinear autoencoder without the Koopman constraint.
  • An untested extension is whether the same frozen embedding transfers across different chaotic systems or across parameter regimes; if it does, the amortisation argument becomes much stronger.
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 / 5 minor

Summary. The paper proposes a three-stage transfer-learning pipeline for chaotic systems: learn a Koopman autoencoder on Lorenz trajectories, pre-train a transformer for next-state prediction in the Koopman embedding space, then fine-tune (or freeze) the transformer for a safety-function regression task. The authors report that their Koopman-based representations outperform PCA-based baselines and that freezing the pre-trained transformer causes no significant degradation, which they interpret as evidence that the learned representations capture reusable dynamical structure. The paper includes detailed hyperparameter and statistical appendices, paired Wilcoxon tests with Bonferroni correction, and a public project page.

Significance. If the central claim were established, the paper would be a useful step toward foundation-model-style transfer in physics-informed ML, showing that a self-supervised Koopman representation can support a downstream control-oriented task without retraining the backbone. Strengths of the manuscript include the physically structured Koopman operator (diagonal plus banded skew-symmetric), the clear three-stage methodology, and the careful paired statistical testing. However, as detailed in the major comments, the downstream task as designed does not actually test representation transfer: the safety function depends only on the raw query state, which is provided directly to the head. Consequently the headline claims about reusable representations, and the comparisons against PCA baselines, are not currently supported. The issues are identifiable and, in principle, fixable with additional control experiments, so the contribution is conditional rather than conclusive.

major comments (4)
  1. [Sec. 3.3, Fig. 3, Eq. (3)] The downstream task does not test representation transfer. U(q) in Eq. (3) depends only on the query state q, and the Lorenz dynamics are Markovian. In Stage 3, the head receives the transformer's final hidden state (encoding past states) concatenated with the raw query q, so the target is computable from q alone without using any information from the transformer context. The frozen-vs-fine-tuned comparison in Table 2 (p=0.335) therefore only shows that an irrelevant frozen context does not hurt; it cannot establish that Koopman representations are reusable. To make the claim testable, add a q-only head (no transformer context), or feed q through the learned encoder phi_e(q) rather than as raw input, or use a downstream task that genuinely depends on history.
  2. [Sec. 3.4, Table B.1] The standard PCA baseline failed to converge on Task A (Table B.1, 'Converged? No') yet is retained as a Task B comparator. Its poor downstream performance could be due to an untrained transformer backbone or smaller input/head capacity, not to embedding quality. The head input dimensions differ across models (35 for Koopman, 12 for PCA-PI, 6 for PCA), and the safety-head widths also differ, so the observed performance gaps in Table 1 are confounded by capacity. Please match head capacities across models, include a PCA representation that successfully trains on Task A, or explicitly analyze these confounds.
  3. [Table 2, Appendix B.2] The conclusion of 'no degradation' treats a non-significant Wilcoxon p-value as evidence of equivalence. With 252 test trajectories and per-model MSE standard deviations that are large relative to the mean differences (Koopman (F): 3.08±6.55; Koopman (U): 5.59±17.14), the test is likely underpowered. Please report a confidence interval for the difference and an equivalence test (e.g., two one-sided tests or a bootstrap CI) with a pre-specified equivalence bound, or weaken the wording to 'no statistically significant difference detected.'
  4. [Table A.3] The Koopman (F) and Koopman (U) fine-tuning settings are not matched. They differ in learning rate (6.83e-3 vs 1.04e-3), number of epochs (80 vs 50), optimizer (Adam vs AdamW), and safety-head width (128 vs 112). Thus the frozen-vs-fine-tuned comparison is confounded by multiple hyperparameter changes, not just the freeze flag. Use identical training configurations except for the freeze flag, or explicitly analyze the confounding effects.
minor comments (5)
  1. [Sec. 3.3 vs Appendix A.6] The Stage 2 sequence length is inconsistent: Section 3.3 says '64-step non-overlapping sequences' while Appendix A.6 says 'non-overlapping 256-step sequences'; Table A.2 lists context length 64. This ambiguity affects reproducibility.
  2. [Figure 4 caption] The caption mentions 'Region R Boundary (x = 0)' but the text uses region Q throughout; unify the notation.
  3. [Section 4.2] The word 'dissaptivity' appears to be a typo for 'dissipativity'.
  4. [References [11,12]] The author names are spelled inconsistently (Capeáns vs Capeans, Sanjuán vs Sanjuan); please standardize.
  5. [Section 2.3 footnote] The footnote clarifies that 'control' is used by convention, but the paper should state explicitly that no control inputs are applied in the experiments; the safety function is a preparatory component.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the frozen-vs-fine-tuned claim is weakened by a confound, but the pipeline is not circular.

full rationale

The Koopman embedding and transformer are trained on next-state prediction (Task A), while the safety targets (Task B) are computed by an independent recursive sculpting algorithm (Eq. 3, Appendix A.4) that uses the true Lorenz dynamics and a discretized grid. No equation in the paper defines U(q) in terms of the fitted encoder, the Koopman operator, or the transformer, so this is not a case where the prediction is equivalent to the fitted input by construction. The PCA baselines are likewise independent embeddings rather than re-statements of the Koopman target. The only self-citation is [31] (Sabatelli), which is used as a generic reference for transfer learning and is not load-bearing; the Koopman architecture and loss follow Geneva and Zabaras [4], who are not authors of this paper. The strongest claim, that freezing the transformer causes no degradation, is statistically fragile: a non-significant Wilcoxon p-value (0.335, Table B.2) is treated as evidence of equivalence, and the safety head receives the raw query state q even though U(q) is Markovian and depends only on q, so the transformer context could be irrelevant to the task. This is a real validity threat to the transfer interpretation, but it is a confound rather than a circularity, because the ground-truth safety function is not constructed from the model being evaluated. The derivation chain is therefore self-contained and not circular.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a set of domain assumptions about the Lorenz system, the learnability of a finite Koopman embedding, and the accuracy of the safety-function discretization. The free parameters are the usual hyperparameters of the ML pipeline: latent dimension, loss weights, safety region bounds, grid size, and per-embedding transformer depth. No new physical entities are introduced.

free parameters (5)
  • Koopman embedding dimension = 32
    Chosen by hand as the latent size for the encoder and transformer; central to the architecture but not fitted to a specific target.
  • Loss weighting terms = lambda0=1e4, lambda1=1.0, lambda2=0.1
    Hyperparameters balancing reconstruction, dynamics, and regularization in the Koopman autoencoder loss (Eq. 6).
  • Safety region bounds = x in [0,50], y in [-50,50], z in [-50,50]
    Manually chosen region Q that contains the right-wing Lorenz attractor; the safety function and evaluation depend on it.
  • Discretization grid size = 27,000 points
    Number of grid points used by the sculpting algorithm to compute ground-truth safety values; affects approximation accuracy.
  • Transformer depth = 4 layers for Koopman, 11 for PCA (PI), 3 for PCA
    Model capacities tuned separately per embedding, making compute comparisons between methods confounded by depth.
assumptions (3)
  • domain assumption Lorenz equations with sigma=10, rho=28, beta=8/3 produce a representative chaotic system
    The testbed is taken as a canonical stand-in for chaotic dynamics; no argument is given for why results transfer to other systems.
  • domain assumption A finite-dimensional linear Koopman embedding of the Lorenz dynamics exists and is learnable by the autoencoder
    The method assumes the 32D latent space can capture the relevant observables; this is induced by training, not proven.
  • domain assumption The sculpting algorithm (Eq. 3) converges to the true safety function U_infty
    Ground-truth safety values rely on this recursive method over a discretized grid; its approximation quality is not validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Generalisation of Koopman Representations for Chaotic System Control." pith.science (2026). https://pith.science/paper/Z4VLV6OR

@misc{pith2026250818954,
  author       = {Pith},
  title        = {Pith review of: On the Generalisation of Koopman Representations for Chaotic System Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z4VLV6OR}},
  note         = {Machine review of arXiv:2508.18954}
}
read the original abstract

This paper investigates the generalisability of Koopman-based representations for chaotic dynamical systems, focusing on their transferability across prediction and control tasks. Using the Lorenz system as a testbed, we propose a three-stage methodology: learning Koopman embeddings through autoencoding, pre-training a transformer on next-state prediction, and fine-tuning for safety-critical control. Our results show that Koopman embeddings outperform both standard and physics-informed PCA baselines, achieving accurate and data-efficient performance. Notably, fixing the pre-trained transformer weights during fine-tuning leads to no performance degradation, indicating that the learned representations capture reusable dynamical structure rather than task-specific patterns. These findings support the use of Koopman embeddings as a foundation for multi-task learning in physics-informed machine learning. A project page is available at https://kikisprdx.github.io/.

Figures

Figures reproduced from arXiv: 2508.18954 by the authors.

Figure 1
Figure 1. Conceptual visualisation illustrating the local, short-horizon nature of next-state prediction (Task A) with the global, long-horizon calculation of the safety function (Task B). Panel (a) illustrates Task A, where a model predicts the single ”Next Prediction point” (yellow) using only a finite history of preceding ”Context States” (blue). In contrast, panel (b) shows that determining the safety value U(q) for the s… view at source ↗
Figure 2
Figure 2. Example double-banded skew-symmetric Koopman operator structure as defined by Equation 5 (shown here as 8×8 - actual implementation uses 32×32). chaotic dynamics of the Lorenz system in a learned latent space. The model architecture consists of three components: an encoder ϕe : R 3 → R 32 that maps Lorenz states to a 32-dimensional embedding space, a decoder ϕd : R 32 → R 3 that reconstructs the original states from… view at source ↗
Figure 3
Figure 3. An illustration of the three-stage training methodology, which is detailed in Section 3.3. Stage 1 (Representation Learning): A Koopman autoencoder is trained to learn a latent representation where the system’s dynamics evolve linearly. An input state st is mapped by an encoder ϕe to a latent vector yt. This latent state is then propagated forward by the Koopman operator matrix K, and the decoder ϕd produces the nex… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparative accumulated L1 error across the test dataset. The L1 error is defined as L1 = |ytrue−ypred|. The figure displays the X–Z projection for four models. Light regions indicate low error; darker regions represent higher error concentrations. Contextual markers d…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages

  1. [1]

    Learning chaotic dynamics by neural networks

    H. Navone and H. Ceccatto. “Learning chaotic dynamics by neural networks”. In:Chaos, Soli- tons & Fractals 6 (1995), pp. 383–387

    work page 1995

  2. [2]

    Hamiltonian Systems and Transformation in Hilbert Space

    B. O. Koopman. “Hamiltonian Systems and Transformation in Hilbert Space”. In: Proceed- ings of the National Academy of Sciences 17.5 (1931), pp. 315–318

    work page 1931

  3. [3]

    Applied koopmanism

    M. Budiˇ si´ c, R. Mohr, and I. Mezi´ c. “Applied koopmanism”. In: Chaos: An Interdisciplinary Journal of Nonlinear Science 22.4 (2012)

    work page 2012

  4. [4]

    Transformers for modeling physical systems

    N. Geneva and N. Zabaras. “Transformers for modeling physical systems”. In: Neural Net- works 146 (2022), pp. 272–289

    work page 2022

  5. [5]

    M. Lazar. From Product Hilbert Spaces to the Generalized Koopman Operator and the Non- linear Fundamental Lemma. 2025. arXiv: 2508. 07494 [math.OC]. url: https://arxiv.org/ abs/2508.07494

    arXiv 2025

  6. [6]

    Attention is all you need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszko- reit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. “Attention is all you need”. In: Advances in neural information processing sys- tems 30 (2017)

    work page 2017

  7. [7]

    Universal Language Model Fine-tuning for Text Classification

    J. Howard and S. Ruder. “Universal Language Model Fine-tuning for Text Classification”. In: (July 2018). Ed. by I. Gurevych and Y. Miyao, pp. 328–339. doi: 10.18653/v1/P18-

    doi:10.18653/v1/p18- 2018

  8. [8]

    Improving language un- derstanding by generative pre-training

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al. “Improving language un- derstanding by generative pre-training”. In: (2018)

    work page 2018

Show all 52 references
  1. [9]

    Robot learning with sensorimotor pre-training

    I. Radosavovic, B. Shi, L. Fu, K. Goldberg, T. Darrell, and J. Malik. “Robot learning with sensorimotor pre-training”. In: Conference on Robot Learning. PMLR. 2023, pp. 683–693

    2023

  2. [10]

    Dynamics of partial control

    J. Sabuco, M. A. Sanju´ an, and J. A. Yorke. “Dynamics of partial control”. In: Chaos: An Interdisciplinary Journal of Nonlinear Science 22.4 (2012)

    2012

  3. [11]

    Partially controlling transient chaos in the Lorenz equations

    R. Cape´ ans, J. Sabuco, M. A. Sanju´ an, and J. A. Yorke. “Partially controlling transient chaos in the Lorenz equations”. In: Philosophi- cal Transactions of the Royal Society A: Math- ematical, Physical and Engineering Sciences 375.2088 (2017), p. 20160211

  4. [12]

    AI-driven control of chaos: A transformer-based approach for dynami- cal systems

    D. Valle, R. Capeans, A. Wagemakers, and M. A. Sanju´ an. “AI-driven control of chaos: A transformer-based approach for dynami- cal systems”. In: Communications in Nonlin- ear Science and Numerical Simulation (2025), p. 109085

    2025

  5. [13]

    Deterministic Nonperiodic Flow 1

    E. N. Lorenz. “Deterministic Nonperiodic Flow 1”. In: Universality in Chaos, 2nd edition. Routledge, 2017, pp. 367–378

    2017

  6. [14]

    Modeling the dynamics of PDE systems with physics- constrained deep auto-regressive networks

    N. Geneva and N. Zabaras. “Modeling the dynamics of PDE systems with physics- constrained deep auto-regressive networks”. In: Journal of Computational Physics 403 (2020), p. 109056

    2020

  7. [15]

    On the concept of attractor

    J. Milnor. “On the concept of attractor”. In: Communications in Mathematical Physics 99.2 (1985), pp. 177–195

    1985

  8. [16]

    A data–driven approximation of the koopman operator: Extending dynamic mode decomposition

    M. O. Williams, I. G. Kevrekidis, and C. W. Rowley. “A data–driven approximation of the koopman operator: Extending dynamic mode decomposition”. In: Journal of Nonlinear Sci- ence 25 (2015), pp. 1307–1346

    2015

  9. [17]

    Relational inductive bi- ases, deep learning, and graph networks

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Mali- nowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, et al. “Relational inductive bi- ases, deep learning, and graph networks”. In: arXiv preprint arXiv:1806.01261 (2018)

    2018 arXiv

  10. [18]

    Physics-Informed Koopman Net- work for time-series prediction of dynami- cal systems

    Y. Liu, A. Sholokhov, H. Mansour, and S. Nabi. “Physics-Informed Koopman Net- work for time-series prediction of dynami- cal systems”. In: ICLR 2024 Workshop on AI4DifferentialEquations In Science . 2024

    2024

  11. [19]

    Curriculum Learning

    Y. Bengio, J. Louradour, R. Collobert, and J. Weston. “Curriculum Learning”. In: Proceed- ings of the 26th Annual International Confer- ence on Machine Learning. ACM, 2009, pp. 41–

    2009

  12. [20]

    Transfer learning for nonlinear dynamics and its application to fluid turbulence

    M. Inubushi and S. Goto. “Transfer learning for nonlinear dynamics and its application to fluid turbulence”. In: Physical Review E 102.4 (2020), p. 043301

    2020

  13. [21]

    Three-dimensional ESRGAN for super-resolution reconstruction of turbulent flows with tricubic interpolation- based transfer learning

    L. Yu, M. Z. Yousif, M. Zhang, S. Hoyas, R. Vinuesa, and H.-C. Lim. “Three-dimensional ESRGAN for super-resolution reconstruction of turbulent flows with tricubic interpolation- based transfer learning”. In: Physics of Fluids 34.12 (2022)

    2022

  14. [22]

    Language mod- els are unsupervised multitask learners

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al. “Language mod- els are unsupervised multitask learners”. In: OpenAI blog 1.8 (2019), p. 9

    2019

  15. [23]

    On layer normalization in the trans- former architecture

    R. Xiong, Y. Yang, D. He, K. Zheng, S. Zheng, C. Xing, H. Zhang, Y. Lan, L. Wang, and T. Liu. “On layer normalization in the trans- former architecture”. In: International con- ference on machine learning . PMLR. 2020, pp. 10524–10533. 9

    2020

  16. [24]

    A learning al- gorithm for continually running fully recurrent neural networks

    R. J. Williams and D. Zipser. “A learning al- gorithm for continually running fully recurrent neural networks”. In: Neural computation 1.2 (1989), pp. 270–280

    1989

  17. [25]

    Phantom oscillations in principal component analysis

    M. Shinn. “Phantom oscillations in principal component analysis”. In: Proceedings of the National Academy of Sciences 120.48 (2023), e2311420120

    2023

  18. [26]

    C. M. Bishop and N. M. Nasrabadi. Pattern recognition and machine learning . Vol. 4. 4. Springer, 2006

    2006

  19. [27]

    Individual comparisons by rank- ing methods

    F. Wilcoxon. “Individual comparisons by rank- ing methods”. In: Breakthroughs in statistics: Methodology and distribution . Springer, 1992, pp. 196–202

    1992

  20. [28]

    Multiple Comparisons Among Means

    O. J. Dunn. “Multiple Comparisons Among Means”. In: Journal of the American Statis- tical Association 56.293 (1961), pp. 52–64. issn: 01621459, 1537274X. url: http : / / www . jstor . org / stable / 2282330(visited on 07/31/2025)

    1961

  21. [29]

    Stable and unstable regions of the Lorenz system

    B. L. Shen, M. Wang, P. Yan, H. Yu, J. Song, and C. J. Da. “Stable and unstable regions of the Lorenz system”. In: Scientific Reports 8.1 (2018), p. 14982

    2018

  22. [30]

    Learn- ing chaotic dynamics with embedded dissipa- tivity

    S. Tang, T. Sapsis, and N. Azizan. “Learn- ing chaotic dynamics with embedded dissipa- tivity”. In: arXiv preprint arXiv:2410.00976 (2024)

    2024

  23. [31]

    Sabatelli

    M. Sabatelli. Contributions to deep trans- fer learning: from supervised to reinforcement learning. Universite de Liege (Belgium), 2022

    2022

  24. [32]

    Con- trol barrier functions: Theory and applica- tions

    A. D. Ames, S. Coogan, M. Egerstedt, G. No- tomista, K. Sreenath, and P. Tabuada. “Con- trol barrier functions: Theory and applica- tions”. In: 2019 18th European control con- ference (ECC). Ieee. 2019, pp. 3420–3431

    2019

  25. [33]

    SciPy 1.0: Fundamental Algorithms for Scientific Com- puting in Python

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, ˙I. Polat, Y. Feng...

    2020

  26. [34]

    Fehlberg

    E. Fehlberg. Low-order classical Runge-Kutta formulas with stepsize control and their appli- cation to some heat transfer problems . Tech. rep. NASA-TR-R-315. Washington, D.C.: Na- tional Aeronautics and Space Administration, 1969

    1969

  27. [35]

    Rectified Lin- ear Units Improve Restricted Boltzmann Ma- chines

    V. Nair and G. E. Hinton. “Rectified Lin- ear Units Improve Restricted Boltzmann Ma- chines”. In: Proceedings of the 27th Inter- national Conference on Machine Learning (ICML-10). 2010, pp. 807–814

    2010

  28. [36]

    Gaussian Error Linear Units (GELUs)

    D. Hendrycks and K. Gimpel. “Gaussian Error Linear Units (GELUs)”. In: arXiv preprint arXiv:1606.08415 (2016)

    2016 arXiv

  29. [37]

    Optuna: A Next-Generation Hy- perparameter Optimization Framework

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama. “Optuna: A Next-Generation Hy- perparameter Optimization Framework”. In: Proceedings of the 25th ACM SIGKDD Inter- national Conference on Knowledge Discovery & Data Mining . New York, NY, USA: Associa- tion for Computing Ma...

    2019

  30. [38]

    Algorithms for Hyper-Parameter Opti- mization

    J. S. Bergstra, R. Bardenet, Y. Bengio, and B. K´ egl. “Algorithms for Hyper-Parameter Opti- mization”. In: Advances in Neural Information Processing Systems 24. Ed. by J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett, F. Pereira, and K. Q. Weinberger. Curran Associates, Inc., 201...

    2011

  31. [39]

    Adam: A method for stochastic optimization

    D. P. Kingma and J. Ba. “Adam: A method for stochastic optimization”. In: arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  32. [40]

    Decoupled weight decay regularization

    I. Loshchilov and F. Hutter. “Decoupled weight decay regularization”. In: arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  33. [41]

    PyTorch: An Imperative Style, High-Performance Deep Learning Li- brary

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. “PyTorch: An Imperative Style, High-Pe...

    2019

  34. [42]

    Zaharia et al

    M. Zaharia et al. MLflow: A Machine Learning Lifecycle Platform. Version 2.14.1. 2024. url: https://mlflow.org/. 10 A Methodology Details A.1 Lorenz System The following system of differential equations defines the Lorenz system: dx dt = σ(y − x); (10) dy dt = x(ρ − z) − y; (1...

    2024

  35. [46]

    They are obtained by applying a Principal Component Analysis (PCA) transformation to the original state vector s

    PCA Coordinates: The components z1, z2, z3 represent the coordinates of the system in a new, decorrelated basis. They are obtained by applying a Principal Component Analysis (PCA) transformation to the original state vector s

  36. [47]

    Transformed Derivatives: The time derivatives of the system’s state, denoted ˙x, ˙y, and ˙z, are calculated using the Lorenz differential equations [13]. While the table shows this calculation in the original state space, the resulting velocity vectors are subsequently project...

  37. [48]

    doi: 10.1145/1553374.1553380

  38. [49]

    The sine and cosine of the phase an- gle

    Engineered Features: a. The sine and cosine of the phase an- gle. This angle is calculated using the two-argument arctangent function, atan2(z2, z1), which captures the angular position of the trajectory in the primary PCA plane. b. The radial distance from the origin in the p...

  39. [50]

    This feature captures the magnitude of the state’s projection onto this plane. A.4 Safety Function Computation The safety function U∞(q) represents the minimum control effort required to maintain trajectories start- ing from state q within a specified safe region Q 11 T able A...

  40. [51]

    PCA Coordinates z1, z2, z3

  41. [52]

    T ransformed Derivatives ˙x = σ(y − x) ˙y = x(ρ − z) − y ˙z = xy − βz

  42. [53]

    The goal is to keep the system’s trajectory within this region even when affected by bounded disturbances

    Engineered F eatures a) sin(atan2( z2, z1)), cos(atan2(z2, z1)) b) q z2 1 + z2 2 indefinitely [11]. The goal is to keep the system’s trajectory within this region even when affected by bounded disturbances. The ground-truth values for this function are com- puted using a recur...

    1919

  43. [1031]

    org / P18-1031/

    url: https : / / aclanthology . org / P18-1031/

  44. [2631]

    isbn: 9781450362016. doi: 10 . 1145 / 3292500.3330701

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.