Pith. sign in

REVIEW 5 major objections 4 minor 54 references

From Uncertain to Safe: Conformal Adaptation of Diffusion Models for Safe PDE Control

T0 review · 5 major / 4 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 conformal uncertainty quantile in the diffusion training loss yields PDE control sequences that satisfy safety constraints while keeping control error low.

desk verdict The empirical story is plausible, but the conformal safety guarantee is claimed for a pipeline that violates its own assumptions; that gap needs to be fixed or honestly reframed. read the letter →

arxiv 2502.02205 v4 pith:BCPHMSDM submitted 2025-02-04 cs.LG

Peiyan Hu , Xiaowei Qian , Wenhao Deng , Rui Wang , Haodong Feng , Ruiqi Feng , Tao Zhang , Long Wei
show 3 more authors
Yue Wang Zhi-Ming Ma Tailin Wu
This is my paper · ORCID
classification cs.LG
keywords safePDEcontroldiffusionmodelsconformalpredictionuncertaintyquantilesafetyconstraintsofflineBurgers'equationtokamakfusion
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

Diffusion-model controllers for PDEs can generate trajectories that look plausible but violate safety constraints because the model's predicted safety score carries unquantified error. This paper argues that an uncertainty quantile, estimated by conformal prediction on a held-out calibration set, can close that gap: when the quantile is added to the safety score before comparing with the bound, and the resulting margin is folded into a reweighted post-training loss and into inference-time guidance, the model is pushed toward distributions where its predictions are safe. On 1D Burgers' equation, 2D incompressible fluid, and tokamak fusion control, the authors report that SafeDiffCon is the only method in their comparison that satisfies all safety constraints while also achieving the lowest control objective among safe methods. The sympathetic reading is that uncertainty-aware reweighting, not better physics or more data, is what converts a diffusion model into a safety-respecting controller.

What carries the argument

The load-bearing object is the uncertainty quantile $Q(1-\alpha;\tilde{S})$, a single number extracted from the calibration set that represents how much the model's safety predictions can be trusted under distribution shift. It enters the function $W(u,w)=\max[s(u)+Q-s_0,0]+\gamma J(u,w)$, where $s_0$ is the safety bound and $J$ the control objective; $W$ is then used in three places: as the exponent in the reweighted post-training loss $e^{-W}\|\epsilon-\epsilon_\theta\|^2$, as the guidance $G=W$ during sampling, and as the fine-tuning loss summed over sampled trajectories. The mechanism that carries the argument is the reweighting: samples with large $W$---unsafe or high-cost ones---are down-weighted, and Theorem 4.2 shows that the model's stationary distribution becomes $p^*\propto p\,e^{-W}$. The validity of the safety margin rests on Theorem 4.1's coverage guarantee, which assumes weighted exchangeability between the calibration set and the model-generated test inputs.

What would settle it

Take one benchmark (e.g., the Burgers or tokamak task), fix $\alpha=0.1$, run SafeDiffCon with many random seeds, and compute the fraction of test trajectories whose true safety score falls outside the conformal interval---or, more directly, the fraction that violate $s_0$; if that fraction reliably exceeds $\alpha$, or if the coverage is below $1-\alpha$, the paper's central safety claim is refuted. A faster check is to recompute the uncertainty quantile on a fresh calibration set after post-training and fine-tuning; a large change indicates the quantile used during training was not calibrated for the deployed model.

Watch

Extended reading notes

Core claim

The central claim is that conformal prediction can turn a diffusion model's safety predictions into intervals that cover the true safety score, and that using the upper edge of those intervals as a penalty in training and sampling produces controls that respect hard safety bounds. The authors define the uncertainty quantile $Q(1-\alpha;\tilde{S})$ as the $(1-\alpha)(1+1/|D_{\mathrm{cal}}|)$-quantile of the shifted score set $\tilde{S}=\{\omega_{\mathrm{norm}}(u_i,w_i)|s(u_\theta(w_i))-s(u_i)|\}$, with weights $\omega_{\mathrm{norm}}\propto e^{-W(u_i,w_i)}$, and Theorem 4.1 states that the true safety score lies in the resulting conformal interval with probability at least $1-\alpha$. They then post-train the diffusion model with the reweighted loss $L_{\mathrm{post-train}}=\mathbb{E}_{k,(u,w),\epsilon}[e^{-W(u,w)}\|\epsilon-\epsilon_\theta(\cdot,k)\|^2]$, which by Theorem 4.2 yields $p^*(u,w)\propto p(u,w)e^{-W(u,w)}$, and they apply the same $W$ as guidance during inference, followed by fine-tuning on the sampled outputs. The reported upshot is that on all three PDE-constrained control tasks the method achieves a 0% unsafe-trajectory rate and the best control error among the methods that stay within the safety bound.

Load-bearing premise

The safety guarantee assumes the calibration set and test-time controls are exchangeable in the weighted sense, and that the model is fixed when the quantile is computed; the method updates the model after calibration, so this assumption is what must hold for Theorem 4.1 to apply to the deployed controller.

Editorial extensions

If this is right

  • If the coverage claim is correct, safe PDE control no longer requires an explicit constraint solver or online interaction: the safety margin is learned offline from a calibration set and carried through training and sampling.
  • The ablations show that post-training and inference-time fine-tuning are individually necessary; removing either raises the unsafe-trajectory rate on the benchmarks.
  • Setting the uncertainty quantile to zero (removing $Q$) produces violations, so the safety margin is doing real work beyond the point prediction of the safety score.

Reading between the lines

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

  • Because Theorem 4.1 is proved for a fixed $\theta$ and the algorithm changes $\theta$ after computing $Q$, the formal coverage guarantee does not apply to the final fine-tuned model; in practice the safety margin functions as a heuristic unless $Q$ is recomputed after the last parameter update.
  • The empirical 0% violation rates exceed the distributional guarantee, which suggests the margin is conservative; a useful stress test would be to lower $\alpha$ and see whether violations rise at the predicted rate.
  • Recomputing $Q$ on a fresh calibration set after fine-tuning, or updating it with an online conformal rule, would be the direct way to restore a formal guarantee and is a natural follow-up test.
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

5 major / 4 minor

Summary. The paper proposes SafeDiffCon, a method for safe PDE control that combines diffusion models with conformal-prediction-based uncertainty quantification. The method computes an uncertainty quantile Q from a calibration set, then uses Q in a reweighted post-training loss and in inference-time guidance and fine-tuning. The paper claims that the conformal interval defined in Eq. (9) covers the true safety score with probability at least 1−α (Theorem 4.1), and reports experiments on 1D Burgers' equation, 2D incompressible fluid, and a tokamak fusion control task in which SafeDiffCon is the only method with zero safety violations while achieving competitive control objectives.

Significance. If the conformal coverage guarantee were valid for the deployed pipeline, SafeDiffCon would be a significant contribution: it would give a principled way to turn a diffusion-based PDE controller into a safety-aware controller under distribution shift, with finite-sample coverage from a calibration set. The paper also ships code and evaluates on three nontrivial control tasks, reporting zero violation rates and an ablation study showing that post-training, fine-tuning, and the uncertainty quantile each contribute to safety. These empirical strengths are real. However, the central theoretical claim is not established for the actual algorithm: the calibration signal Q is computed from residuals and then used to modify the model and the weights, so the conformal assumptions are violated. As presented, the safety mechanism is better characterized as a heuristic penalty than as a calibrated 1−α guarantee. The paper merits a major revision rather than acceptance, because the load-bearing theory needs to be fixed or substantially reframed.

major comments (5)
  1. [§4.1 and §4.2, Eqs. (6), (11), (13)] The shifted score set S̃ in Eq. (6) is defined using weights ω_norm, but in Section 4.2 the weights are set to ω = C e^{−W} (Eq. 13), where W in Eq. (11) contains Q(1−α; S̃). Thus Q is the quantile of a weighted empirical distribution whose weights depend on Q itself. No fixed-point or contraction argument is supplied to show that this self-referential definition is well-posed. Theorem 4.1 cannot be invoked because it assumes a fixed, known likelihood ratio ω = p̃/p, not weights that depend on the calibration scores and on Q.
  2. [§4.2, §4.3, Algorithms 2 and 3] The coverage claim in Theorem 4.1 requires the model θ to be fixed when Q is computed and the calibration set to be independent of the test point. In SafeDiffCon, Dcal is reused every epoch of post-training (Algorithm 2, line 3) and every iteration of inference-time fine-tuning (Algorithm 3, line 3) to recompute Q, and the model is then updated with losses Eq. (12) and Eq. (15) that depend on Q. The final model is therefore not fixed and Dcal is not independent of the final predictions. Moreover, the test states are generated by u_θ(w) with θ being updated, so the distribution shift is not a pure covariate shift; the conditional law of the safety score changes during the procedure. Consequently, the zero-violation results are empirical outcomes of an adaptive penalty, not a calibrated 1−α guarantee.
  3. [Appendix A, proof of Theorem A.1 / Theorem 4.1] The proof derives two one-sided inequalities, P(s(u) ≤ s(u_θ(w)) + Q) ≥ 1−α and P(s(u_θ(w)) − Q ≤ s(u)) ≥ 1−α, and then concludes the two-sided interval has probability at least 1−α. This conclusion does not follow from the two displayed inequalities: a union bound gives only probability at least 1−2α for their intersection. If the intended argument is to apply Lemma 3 of Tibshirani et al. (2019) directly to the absolute residuals in S (Eq. 5), the proof should state that explicitly; as written, the proof is incomplete.
  4. [Theorem 4.3 and its proof, Appendix A] Theorem 4.3 does not establish the coverage of the deployed interval. The statement concerns equality of quantiles of weighted empirical distributions when DKL(p‖p_θ) ≤ ε, but the proof bounds |p(v)−p_θ(v)| by δ with 'probability' over a pair of fixed distributions, which is not a well-defined probabilistic statement. The constant C depends on p_min and δ, and δ is not a free parameter that can be made small independently of ε. Even if the quantile approximation were correct, it would only say that the weighted quantile is close to the unweighted quantile; it would not imply the coverage probability in Eq. (10), because coverage requires weighted exchangeability of the final scores.
  5. [Theorem 4.1 statement] The theorem assumes 'p is absolutely continuous with respect to p̃', but the weight in Eq. (7) is ω = p̃/p. Weighted conformal prediction under covariate shift requires p̃ to be absolutely continuous with respect to p (support(p̃) ⊆ support(p)) so that the likelihood ratio is well-defined. The stated assumption appears to be reversed; the same wording appears in Appendix A, Theorem A.1.
minor comments (4)
  1. [Eq. (8)] The denominator in the definition of ω_norm is written with the subscript '(u_i,w_i)∈Dcal' but should be a sum over (u_j,w_j)∈Dcal; as printed, the expression is mathematically ambiguous.
  2. [Appendix B.3, Tables 8 and 9] The caption of Table 8 says it studies the coverage probability α, but the table's first column is labeled 'Split Ratio', while Table 9 says it studies the split ratio and its first column is labeled 'α'. The captions and column labels appear to be swapped.
  3. [Appendix F.1] The text says 'Following Eq. (5.1)', but the safety score is defined in Eq. (17) of Section 5.1; the equation reference is incorrect.
  4. [Appendix H.1] The list of plasma parameters writes '1 − αp' where the paper's main text and objective (Eq. 18) use β_p; the notation should be made consistent.

Circularity Check

2 steps flagged · score 6.0 of 10

The paper's uncertainty quantile Q is self-referential: the shifted-score weights are e^{-W}, W contains Q, so Q is defined partly in terms of itself, and the conformal coverage guarantee is applied to a model that is updated using the calibration set after Q is computed.

  1. self definitional [Section 4.1-4.2, Eqs. (6), (11), (13)]
    "˜S := {ωnorm(ui, wi)∆si : ∆si ∈ S}. (6) ... W(u, w) = max[s(u) + Q(1 − α; ˜S) − s0, 0] + γJ (u, w), (11) ... In practical implementation, we let ω(ui, wi) = Ce −W(ui,wi), (13) and the constant C will be eliminated during normalization."

    Eq. 6 defines the shifted score set with weights ω_norm; Eq. 13 sets ω ∝ e^{-W}; Eq. 11 defines W using Q(1−α; ˜S). Hence Q is the quantile of scores weighted by e^{-W}, and W is a function of Q: Q = Quantile({e^{-max[s+Q-s0,0]-γJ} Δs}). The quantile therefore depends on the very quantity it is supposed to calibrate. Neither Theorem 4.1 (which assumes a fixed known likelihood ratio p_tilde/p) nor Theorem 4.3 (which compares two fixed weight vectors) establishes existence or uniqueness of this fixed point, so the 'uncertainty quantile' is not a well-defined conformal statistic and the coverage claim in Eq. 10 does not follow for this Q.

  2. fitted input called prediction [Section 4.3, Algorithm 3; Theorem 4.1, Eq. (10)]
    "Theorem 4.1. Assume that samples in the calibration set Dcal ∼ p are independent, and the test set (u, w) ∼ ˜p is also independent with the calibration set. ... P(s(u) ∈ CIθ(1 − α, Dcal)) ≥ 1 − α. (10) [Algorithm 3:] Compute the shifted score set ˜S with Dcal // Eq. 6; Get the uncertainty quantile Q(1 − α; ˜S); Sample the control sequence w with guidance G // Eq. 14; Take gradient descent step on ∇θLfine-tune // Eq. 15."

    Algorithm 3 recomputes Q from Dcal at every iteration and then updates θ by ∇θ L_fine-tune = Σ W(uθ,wθ), with W containing Q (Eqs. 11, 14, 15). The final generation w therefore depends on Dcal through both Q and θ, while Theorem 4.1 requires the test point (u,w) ~ p_tilde to be independent of Dcal. The claimed '≥ 1−α' coverage in Eq. 10 is thus asserted for a procedure whose own calibration set is used as a training signal; it is a fitted margin renamed as a probabilistic guarantee. Dcal is also reused across iterations, further breaking the weighted-exchangeability assumption.

full rationale

The core diffusion-control machinery (reweighted diffusion loss, guidance, fine-tuning) is not circular by itself: Theorem 4.2 states that the reweighted loss trains the model toward p e^{-W}, which is true by construction of the loss, and the empirical control results are real experimental outcomes. The circularity is concentrated in the conformal layer. The uncertainty quantile Q is computed from a shifted score set whose weights are e^{-W}, while W contains Q; this makes the quantile self-referential rather than an externally calibrated statistic. In addition, the 1−α coverage theorem is invoked for the final SafeDiffCon output, but the algorithm recomputes Q from Dcal and then fine-tunes the model using W(Q), so the calibration set is no longer independent of the test-time model. The ablation study shows that removing Q degrades safety, but that only confirms that the safety penalty is an active part of the objective; it does not validate the probabilistic coverage claim. The safety results remain interesting as empirical demonstrations of a heuristic safety margin, but the paper's stated guarantee that the conformal interval covers the true safety score with probability at least 1−α does not apply to the deployed pipeline. No load-bearing self-citation or imported uniqueness theorem was found; the external conformal prediction theory is standard and independent. The circularity is partial: it affects the derivation of the safety guarantee, not the empirical control comparison.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method inherits standard diffusion training assumptions and weighted-conformal assumptions. The most important additional ingredient is the circular dependence of Q on W and the reuse of the calibration set to update the model, which is not covered by the cited conformal theory. Free parameters are standard hyperparameters such as alpha, gamma, and guidance strengths that are chosen by hand and affect safety outcomes.

free parameters (4)
  • coverage probability alpha = 0.01 or 0.02 for main results; values 0.005-0.02 explored
    Sets the quantile level for the conformal interval in Eq. 9. The choice affects how conservative the safety penalty is. Tables 8 and 9 report sensitivity but their headers appear swapped.
  • objective weight gamma = 0.01 for the fluid ablation; value not clearly stated for all tasks
    Balances the control objective J and the safety term in W (Eq. 11). Table 10 shows that gamma = 0.1 and 0.3 produce 2% violation rates, so the reported zero violations depend on the chosen gamma.
  • guidance intensity and safety weight for inference = 100 and 10000 for the tokamak experiment (Table 13)
    Coefficients in the guided sampling step and in W. These hand-set values directly control how strongly the model is pushed toward safe regions.
  • post-training epochs, fine-tuning iterations, DDIM steps = DDIM steps = 100; N and M not specified for all experiments
    Algorithmic loop counts determine how far the model is pushed toward safe regions. Without exact values, reproduction is incomplete.
assumptions (5)
  • domain assumption Calibration and test samples are weighted exchangeable under a known likelihood ratio p_tilde/p
    Theorem 4.1 inherits weighted conformal prediction from Tibshirani et al. 2019. The implementation approximates p_tilde/p with e^{-W} (Eq. 13), so the assumption is not exactly satisfied.
  • ad hoc to paper The model theta is fixed when the calibration quantile Q is computed
    Split-conformal coverage requires calibration on a model that will not be modified afterward. Algorithms 2 and 3 alternate between updating theta and recomputing Q on Dcal, so this assumption is violated.
  • ad hoc to paper The pre-trained model p_theta is close enough to the data distribution p
    Justifies dropping the p_theta/p factor in Eq. 13 via Theorem 4.3. The theorem gives only a high-probability bound under a small KL divergence, and the paper does not measure this divergence.
  • standard math Reweighted diffusion training converges to p* proportional to p e^{-W}
    Theorem A.2 is a standard weighted denoising score-matching result. It assumes W is fixed, which is not strictly true when Q is updated during training.
  • domain assumption Surrogate simulators accurately represent the real environments
    Safety and objective are measured in a numerical Burgers solver, PhiFlow, and an LSTM-based KSTAR simulator (Appendix H). Errors in these simulators would change violation rates and objective values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Uncertain to Safe: Conformal Adaptation of Diffusion Models for Safe PDE Control." pith.science (2026). https://pith.science/paper/BCPHMSDM

@misc{pith2026250202205,
  author       = {Pith},
  title        = {Pith review of: From Uncertain to Safe: Conformal Adaptation of Diffusion Models for Safe PDE Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCPHMSDM}},
  note         = {Machine review of arXiv:2502.02205}
}
read the original abstract

The application of deep learning for partial differential equation (PDE)-constrained control is gaining increasing attention. However, existing methods rarely consider safety requirements crucial in real-world applications. To address this limitation, we propose Safe Diffusion Models for PDE Control (SafeDiffCon), which introduce the uncertainty quantile as model uncertainty quantification to achieve optimal control under safety constraints through both post-training and inference phases. Firstly, our approach post-trains a pre-trained diffusion model to generate control sequences that better satisfy safety constraints while achieving improved control objectives via a reweighted diffusion loss, which incorporates the uncertainty quantile estimated using conformal prediction. Secondly, during inference, the diffusion model dynamically adjusts both its generation process and parameters through iterative guidance and fine-tuning, conditioned on control targets while simultaneously integrating the estimated uncertainty quantile. We evaluate SafeDiffCon on three control tasks: 1D Burgers' equation, 2D incompressible fluid, and controlled nuclear fusion problem. Results demonstrate that SafeDiffCon is the only method that satisfies all safety constraints, whereas other classical and deep learning baselines fail. Furthermore, while adhering to safety constraints, SafeDiffCon achieves the best control performance. The code can be found at https://github.com/AI4Science-WestlakeU/safediffcon.

Figures

Figures reproduced from arXiv: 2502.02205 by the authors.

Figure 1
Figure 1. Overview of SafeDiffCon. First, we pre-train a diffusion model pθ on the training data. Then, combined with the uncertainty quantile, we post-train the model to steer its distribution to safer regions with better control objectives. Finally, to improve performance and safety for specific control tasks, we conduct inference-time fine-tuning, again incorporating the uncertainty quantile into the process. throughout th… view at source ↗
Figure 2
Figure 2. Visualizations of the 1D Burgers’ equation. The top row shows the original trajectory corresponding to the control target, and the bottom row is the trajectory controlled by SafeDiffCon. external force w(t, x). Given a target state ud(x), the control objective J is to minimize the control error between the final state uT and the target state ud. J := Z Ω |u(T, x) − ud(x)| 2dx. (16) The safety score is defined as: s(… view at source ↗
Figure 3
Figure 3. Visualization of the 2D incompressible fluid control by our SafeDiffCon. By controlling the fluid on the outside margin, the yellow smoke is successfully maneuvered to the center top exit while avoiding the red unsafe region. random targets within the reasonable ranges proposed in that study. The control sequences for training are collected using the reinforcement learning model trained in the previous work (Seo et … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of the 1D Burgers’ equation. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Visualizations of the 2D incompressible fluid control problem.. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 38 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    B., Witrant, E., Prieur, C., Bremond, S., Nouailletas, R., and Artaud, J.-F

    Argomedo, F. B., Witrant, E., Prieur, C., Bremond, S., Nouailletas, R., and Artaud, J.-F. Lyapunov-based distributed control of the safety-factor profile in a tokamak plasma. Nuclear Fusion, 53 0 (3): 0 033005, 2013

    work page 2013

  3. [3]

    A safe, accurate intravenous infusion control system

    Barros, E., des Santos, and M.V.D. A safe, accurate intravenous infusion control system. IEEE Micro, 18 0 (5): 0 12--21, 1998. doi:10.1109/40.735940

    work page doi:10.1109/40.735940 1998

  4. [4]

    Decision transformer: Reinforcement learning via sequence modeling

    Chen, L., Lu, K., Rajeswaran, A., Lee, K., Grover, A., Laskin, M., Abbeel, P., Srinivas, A., and Mordatch, I. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34: 0 15084--15097, 2021

    2021

  5. [5]

    M., and Burdick, J

    Cheng, R., Orosz, G., Murray, R. M., and Burdick, J. W. End-to-end safe reinforcement learning through barrier functions for safety-critical continuous control tasks. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pp.\ 3387--3395, 2019

    2019

  6. [6]

    Exact and robust conformal inference methods for predictive machine learning with dependent data

    Chernozhukov, V., W \"u thrich, K., and Yinchu, Z. Exact and robust conformal inference methods for predictive machine learning with dependent data. In Conference On learning theory, pp.\ 732--749. PMLR, 2018

    work page 2018

  7. [7]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., and Song, S. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, pp.\ 02783649241273668, 2023

    2023

  8. [8]

    Risk-constrained reinforcement learning with percentile risk criteria

    Chow, Y., Ghavamzadeh, M., Janson, L., and Pavone, M. Risk-constrained reinforcement learning with percentile risk criteria. Journal of Machine Learning Research, 18 0 (167): 0 1--51, 2018

    work page 2018

Show all 54 references
  1. [9]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=OnD9zGAGT0k

    2023

  2. [10]

    Safe nonlinear control using robust neural lyapunov-barrier functions

    Dawson, C., Qin, Z., Gao, S., and Fan, C. Safe nonlinear control using robust neural lyapunov-barrier functions. In Conference on Robot Learning, pp.\ 1724--1735. PMLR, 2022

    2022

  3. [11]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion Models Beat GANs on Image Synthesis , June 2021. URL http://arxiv.org/abs/2105.05233. arXiv:2105.05233 [cs, stat]

    2021 arXiv

  4. [12]

    Farahmand, A.-m., Nabi, S., and Nikovski, D. N. Deep reinforcement learning for partial differential equation control. In 2017 American Control Conference (ACC), pp.\ 3120--3127, 2017. doi:10.23919/ACC.2017.7963427

    2017

  5. [13]

    H., Brooks, J

    Federici, G., Skinner, C. H., Brooks, J. N., Coad, J. P., Grisolia, C., Haasz, A. A., Hassanein, A., Philipps, V., Pitcher, C. S., Roth, J., et al. Plasma-material interactions in current tokamaks and their implications for next step fusion reactors. Nuclear Fusion, 41 0 (12):...

    1967

  6. [14]

    and Allg \"o wer, F

    Findeisen, R. and Allg \"o wer, F. An introduction to nonlinear model predictive control. In 21st Benelux meeting on systems and control, volume 11, pp.\ 119--141. Veldhoven, 2002

    2002

  7. [15]

    E., Prett, D

    Garcia, C. E., Prett, D. M., and Morari, M. Model predictive control: Theory and practice—a survey. Automatica, 25 0 (3): 0 335--348, 1989

    1989

  8. [16]

    and Candes, E

    Gibbs, I. and Candes, E. Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems, 34: 0 1660--1672, 2021

    2021

  9. [17]

    A review of safe reinforcement learning: Methods, theory and applications

    Gu, S., Yang, L., Du, Y., Chen, G., Walter, F., Wang, J., and Knoll, A. A review of safe reinforcement learning: Methods, theory and applications. arXiv preprint arXiv:2205.10330, 2022

    2022 arXiv

  10. [18]

    Using trusted data to train deep networks on labels corrupted by severe noise

    Hendrycks, D., Mazeika, M., Wilson, D., and Gimpel, K. Using trusted data to train deep networks on labels corrupted by severe noise. Advances in neural information processing systems, 31, 2018

    2018

  11. [19]

    and Kunisch, K

    Hinze, M. and Kunisch, K. Second order methods for optimal control of time-dependent fluid flow. SIAM Journal on Control and Optimization, 40 0 (3): 0 925--946, 2001

    2001

  12. [20]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  13. [21]

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., and Fleet, D. J. Video diffusion models. arXiv:2204.03458, 2022

    2022 arXiv

  14. [22]

    Learning to control pdes with differentiable physics

    Holl, P., Thuerey, N., and Koltun, V. Learning to control pdes with differentiable physics. In International Conference on Learning Representations, 2020

    2020

  15. [23]

    Wavelet diffusion neural operator

    Hu, P., Wang, R., Zheng, X., Zhang, T., Feng, H., Feng, R., Wei, L., Wang, Y., Ma, Z.-M., and Wu, T. Wavelet diffusion neural operator. arXiv preprint arXiv:2412.04833, 2024

    2024 arXiv

  16. [24]

    Y., Shin, J

    Hwang, R., Lee, J. Y., Shin, J. Y., and Hwang, H. J. Solving pde-constrained control problems using operator learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pp.\ 4504--4512, 2022

    2022

  17. [25]

    Planning with diffusion for flexible behavior synthesis

    Janner, M., Du, Y., Tenenbaum, J., and Levine, S. Planning with diffusion for flexible behavior synthesis. Proceedings of Machine Learning Research, 162: 0 9902--9915, 17--23 Jul 2022

    2022

  18. [26]

    Johnson, M. A. and Moradi, M. H. PID control. Springer, 2005

    2005

  19. [27]

    Predicting disruptive instabilities in controlled fusion plasmas through deep learning

    Kates-Harbeck, J., Svyatkovskiy, A., and Tang, W. Predicting disruptive instabilities in controlled fusion plasmas through deep learning. Nature, 568 0 (7753): 0 526--531, 2019

    2019

  20. [28]

    Knight, J. C. Safety critical systems: challenges and directions. In Proceedings of the 24th international conference on software engineering, pp.\ 547--550, 2002

    2002

  21. [29]

    J., Heess, N., Precup, D., Kim, K.-E., and Guez, A

    Lee, J., Paduraru, C., Mankowitz, D. J., Heess, N., Precup, D., Kim, K.-E., and Guez, A. CO pti DICE : Offline constrained reinforcement learning via stationary distribution correction estimation. In International Conference on Learning Representations, 2022. URL https://openr...

    2022

  22. [30]

    G., Rinaldo, A., Tibshirani, R

    Lei, J., G'Sell, M. G., Rinaldo, A., Tibshirani, R. J., and Wasserman, L. A. Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113: 0 1094 -- 1111, 2016. URL https://api.semanticscholar.org/CorpusID:13741419

    2016

  23. [31]

    A., and Buzzicotti, M

    Li, T., Biferale, L., Bonaccorso, F., Scarpolini, M. A., and Buzzicotti, M. Synthetic lagrangian turbulence by generative diffusion models. Nature Machine Intelligence, pp.\ 1--11, 2024

    2024

  24. [32]

    H., and Chong, G

    Li, Y., Ang, K. H., and Chong, G. Pid control system analysis and design. IEEE Control Systems Magazine, 26 0 (1): 0 32--41, 2006. doi:10.1109/MCS.2006.1580152

    2006 arXiv

  25. [33]

    Safe offline reinforcement learning with real-time budget constraints

    Lin, Q., Tang, B., Wu, Z., Yu, C., Mao, S., Xie, Q., Wang, X., and Wang, D. Safe offline reinforcement learning with real-time budget constraints. In International Conference on Machine Learning, pp.\ 21127--21152. PMLR, 2023

    2023

  26. [34]

    Datasets and benchmarks for offline safe reinforcement learning

    Liu, Z., Guo, Z., Lin, H., Yao, Y., Zhu, J., Cen, Z., Hu, H., Yu, W., Zhang, T., Tan, J., et al. Datasets and benchmarks for offline safe reinforcement learning. arXiv preprint arXiv:2306.09303, 2023 a

    2023 arXiv

  27. [35]

    Constrained decision transformer for offline safe reinforcement learning

    Liu, Z., Guo, Z., Yao, Y., Cen, Z., Yu, W., Zhang, T., and Zhao, D. Constrained decision transformer for offline safe reinforcement learning. In International Conference on Machine Learning, pp.\ 21611--21630. PMLR, 2023 b

    2023

  28. [36]

    Maxime Cauchois, Suyash Gupta, A. A. and Duchi, J. C. Robust validation: Confident predictions even when distributions shift. Journal of the American Statistical Association, 0 0 (0): 0 1--66, 2024. doi:10.1080/01621459.2023.2298037. URL https://doi.org/10.1080/01621459.2023.2298037

    2024

  29. [37]

    and Nabi, S

    Mowlavi, S. and Nabi, S. Optimal control of pdes using physics-informed neural networks. Journal of Computational Physics, 473: 0 111731, 2023

    2023

  30. [38]

    Reinforcement learning with function-valued action spaces for partial differential equation control

    Pan, Y., Farahmand, A.-m., White, M., Nabi, S., Grover, P., and Nikovski, D. Reinforcement learning with function-valued action spaces for partial differential equation control. In Dy, J. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learni...

    2018

  31. [39]

    and Ramdas, A

    Podkopaev, A. and Ramdas, A. Distribution-free uncertainty quantification for classification under label shift. In de Campos, C. and Maathuis, M. H. (eds.), Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, volume 161 of Proceedings of Mac...

    2021

  32. [40]

    Pomerleau, D. A. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988

    1988

  33. [41]

    Adjoint-based optimization of pde systems with alternative gradients

    Protas, B. Adjoint-based optimization of pde systems with alternative gradients. Journal of Computational Physics, 227 0 (13): 0 6490--6510, 2008

    2008

  34. [42]

    Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control

    Rabault, J., Kuchta, M., Jensen, A., R \'e glade, U., and Cerardi, N. Artificial neural networks trained through deep reinforcement learning discover control strategies for active flow control. Journal of fluid mechanics, 865: 0 281--302, 2019

    2019

  35. [43]

    and Ariola, M

    Schuster, E. and Ariola, M. The role of controls in nuclear fusion. In Proceedings of the 45th IEEE Conference on Decision and Control, pp.\ 2263--2270. IEEE, 2006

    2006

  36. [44]

    Review on model predictive control: An engineering perspective

    Schwenzer, M., Ay, M., Bergs, T., and Abel, D. Review on model predictive control: An engineering perspective. The International Journal of Advanced Manufacturing Technology, 117 0 (5): 0 1327--1349, 2021

    2021

  37. [45]

    Feedforward beta control in the kstar tokamak by deep reinforcement learning

    Seo, J., Na, Y.-S., Kim, B., Lee, C., Park, M., Park, S., and Lee, Y. Feedforward beta control in the kstar tokamak by deep reinforcement learning. Nuclear Fusion, 61 0 (10): 0 106010, 2021

    2021

  38. [46]

    Development of an operation trajectory design algorithm for control of multiple 0d parameters using deep reinforcement learning in kstar

    Seo, J., Na, Y.-S., Kim, B., Lee, C., Park, M., Park, S., and Lee, Y. Development of an operation trajectory design algorithm for control of multiple 0d parameters using deep reinforcement learning in kstar. Nuclear Fusion, 62 0 (8): 0 086049, 2022

    2022

  39. [47]

    Soner, H. M. Stochastic optimal control in finance. Scuola normale superiore, 2004

    2004

  40. [48]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=St1giarCHLP

    2021

  41. [49]

    J., Barber, R

    Tibshirani, R. J., Barber, R. F., Cand \`e s, E. J., and Ramdas, A. Conformal prediction under covariate shift. In Neural Information Processing Systems, 2019. URL https://api.semanticscholar.org/CorpusID:115140768

    2019

  42. [50]

    Lion: Latent point diffusion models for 3d shape generation

    Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K., et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35: 0 10021--10039, 2022

    2022

  43. [51]

    Algorithmic learning in a random world, volume 29

    Vovk, V., Gammerman, A., and Shafer, G. Algorithmic learning in a random world, volume 29. Springer, 2005

    2005

  44. [52]

    A generative approach to control complex physical systems

    Wei, L., Hu, P., Feng, R., Feng, H., Du, Y., Zhang, T., Wang, R., Wang, Y., Ma, Z.-M., and Wu, T. A generative approach to control complex physical systems. Advances in Neural Information Processing Systems, 2024

    2024

  45. [53]

    Constraints penalized q-learning for safe offline reinforcement learning

    Xu, H., Zhan, X., and Zhu, X. Constraints penalized q-learning for safe offline reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 8753--8760, 2022

    2022

  46. [54]

    E., Zhan, X., and Liu, J

    Zheng, Y., Li, J., Yu, D., Yang, Y., Li, S. E., Zhan, X., and Liu, J. Safe offline reinforcement learning with feasibility-guided diffusion model. arXiv preprint arXiv:2401.10700, 2024

    2024 arXiv

Pith tools

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