Pith. sign in

REVIEW 3 major objections 6 minor 37 references

Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Two-timescale Bayesian optimization picks electron cyclotron heating profiles that avoid tearing instabilities, succeeding in four of eight live tokamak shots.

desk verdict Live result is not statistically supported, but the method and offline work are solid enough to deserve a careful referee. read the letter →

arxiv 2506.10287 v1 pith:VL3EP3BJ submitted 2025-06-12 cs.RO

Rohit Sonker , Alexandre Capone , Andrew Rothstein , Hiro Josep Farre Kaga , Egemen Kolemen , Jeff Schneider This is my paper
classification cs.RO
keywords BayesianoptimizationtokamakplasmacontroltearinginstabilityelectroncyclotronheatingGaussianprocessrecurrentneuralnetworkdynamicsmodelcontextualstabilization
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 claims that a two-timescale Bayesian optimization method, called DynaBO, can find electron cyclotron heating (ECH) profiles that stabilize the plasma in a tokamak using only a handful of live experiments. The method pairs a recurrent neural network trained on historical tokamak dynamics with a Gaussian process that is updated between shots; the neural model supplies the GP's prior mean, and Bayesian optimization chooses the next ECH profile. Offline tests on historical data show lower cumulative regret than several baselines, and live high-performance shots avoided tearing instabilities in four of eight runs, a 117% improvement over historical outcomes for the same configuration, with the average time to instability rising from 2424 ms to 3339 ms. If the report is right, it indicates that learning-based control can make meaningful progress on a volatile, high-confinement fusion scenario with very few expensive trials.

What carries the argument

The machinery is a two-timescale decomposition of the control problem. On the fast scale, an RPNN with a GRU cell (a gated recurrent memory unit) predicts the next plasma state as a multivariate normal distribution; a random-forest tearing-mode predictor then estimates the probability of a 2/1 tearing mode, a magnetic-island instability. Marginalizing those rollouts yields a prior estimate of the time to tearing, $\hat t_{TM}(\bar\beta_N, a_q)$, which serves as the GP's prior mean. On the slow scale, a Gaussian process with a squared-exponential kernel models the difference between the RPNN prior and observed outcomes, with inputs $(\bar\beta_N, a_q)$ and output $t_{TM}$; the posterior mean and variance feed a UCB acquisition function that picks the next ECH profile. The action space is deliberately low-dimensional: the ECH profile is approximated by a stationary Gaussian curve with three parameters (center, width, amplitude), which is projected onto whatever gyrotrons are available. This decomposition is what lets the method update its belief after each shot without retraining the neural network.

What would settle it

Run a sequence of tokamak shots in which only the ECH profile is varied while all other actuators are held fixed and measure the achieved $\bar\beta_N$: if $\bar\beta_N$ moves systematically with $a_q$, the independence assumption fails, whereas if it stays flat, the context/action split is sound.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the reliability of a data-driven plasma dynamics model and the data efficiency of a Gaussian process can be combined across two time scales to make live control decisions. A recurrent probabilistic neural network (RPNN) models high-frequency tokamak state transitions; a tearing-mode classifier converts rollouts into a predicted time to tearing, i.e., how long the plasma stays stable, and those predictions become the prior mean of a GP whose inputs are the target normalized plasma pressure $\bar\beta_N$ and a Gaussian-parameterized ECH profile $a_q$. The GP is updated after every shot with the measured ECH profile and measured $\bar\beta_N$, effectively treating the actuation as noisy, and the next action is selected by upper-confidence-bound acquisition. The paper reports that this pipeline avoided tearing instabilities in four of eight live high-performance shots, compared with a 77% historical tearing rate in similar shots, and increased the average stable time by 914 ms (from 2424 ms to 3339 ms).

Load-bearing premise

The load-bearing premise is that the achieved normalized plasma pressure $\bar\beta_N$ is independent of the ECH profile $a_q$; if ECH also changes $\beta_N$ (the appendix notes ECH can reduce density), the context is not exogenous and the GP's input separation breaks down.

Editorial extensions

If this is right

  • If the central claim is right, a small number of Bayesian optimization iterations can raise the chance of stabilizing a high-performance plasma from a historical success rate of about 23% to 50% in the same configuration.
  • Because the GP updates with measured (noisy) inputs, the method can keep working when hardware changes, gyrotrons fail, or actuator noise shifts the delivered ECH profile.
  • Because the search happens in Gaussian-parameterized ECH space rather than per-gyrotron angle space, the same optimization procedure stays valid as the number of available gyrotrons changes.
  • Offline, the RPNN-prior GP has lower cumulative regret than RPNN-only and vanilla GP baselines across RBF and Matérn kernels, except for a clearly misspecified linear kernel.

Reading between the lines

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

  • An unstated consequence is that any high-cost physical plant with abundant historical logs and a tiny live experiment budget could use the same prior-plus-GP structure; the paper's own suggestion of applying it to ramp-up design is one example.
  • The load-bearing independence assumption suggests a natural extension: if measured $\beta_N$ shifts systematically with the ECH profile in new shots, the GP should be re-conditioned on predicted rather than assumed $\beta_N$, or the context should be treated as an outcome of the action.
  • With eight live shots, the 4/8 success rate carries wide uncertainty; a future campaign that repeats recommended profiles and records failures would reveal whether the real improvement is close to the reported 117% figure.
  • Incorporating a physics-based island-evolution model into the neural prior could improve extrapolation outside the historical data; the paper lists this as future work, and the two-timescale structure would accommodate it without changing the GP update.
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

3 major / 6 minor

Summary. Sonker et al. present DynaBO, a Bayesian optimization framework for selecting feedforward electron cyclotron heating (ECH) profiles to avoid tearing instabilities in tokamaks. The method couples a recurrent probabilistic neural network (RPNN), trained on historical DIII-D data, with a Gaussian process that models the time-to-tearing as a function of the target normalized plasma pressure and a Gaussian-parameterized ECH profile. The RPNN provides the GP prior mean, and the GP is updated after each experimental shot; actions are chosen by contextual upper-confidence-bound optimization. The paper reports offline experiments on 281 historical high-qmin DIII-D shots, where DynaBO outperforms RPNN-only and vanilla GP baselines, and live DIII-D experiments with eight shots, in which tearing instabilities were avoided in four shots. The authors interpret these results as a 117% improvement over a historical 23% success rate, while also acknowledging in Section 4.2 that the number of live shots is too small to be statistically significant.

Significance. If the claims were fully supported, the work would be a valuable demonstration of data-efficient, model-based control in a real tokamak: it combines a learned high-frequency dynamics model with a low-frequency GP, and it reports genuine live experiments at DIII-D, which are rare and costly. The offline study includes multiple kernels, ablations, and acquisition functions, and the design choices around ECH profile parameterization and conversion to gyrotron angles are practical and clearly motivated. However, the central quantitative claim in the Abstract and Introduction is not supported by the presented evidence: the live comparison has very low power and is confounded by an unmatched historical baseline, and the offline evaluation has a temporal-overlap problem between RPNN training data and the evaluation set. These issues affect the main validation of the method and need to be addressed before the headline claims can be accepted.

major comments (3)
  1. [Abstract, Introduction, Section 4.2, Table 1] The headline claim of a 117% improvement is not statistically supported, and the paper itself concedes this. From Table 1, DynaBO avoided tearing in 4 of 8 live shots, versus 14 of 61 historical shots with similar settings (the 77% instability rate quoted in Section 4.2 corresponds to a 23% historical success rate). A Fisher exact test on these counts gives a two-sided p-value of approximately 0.19, and even the one-sided test is approximately 0.12; the observed 4/8 rate is therefore entirely consistent with the historical rate. The text in Section 4.2 states that 'this number of shots is too low to be statistically significant,' yet the Abstract and Introduction still present the 117% improvement as a headline result. The comparison is also not apples-to-apples: the historical baseline is described as 'very similar settings' rather than the same setting, Table 1 shows target beta_N varying from 3.14 to 3.45, and the Introduction acknowledges changes to other actuators. In addition, the mean stability-time comparison (3339 ms versus 2424 ms) is misleading if the historical 2424 ms is an average time-to-tearing for shots that developed tearing, while the DynaBO mean includes shots that never tore and lasted 4-5 s. The authors should remove or substantially soften the quantitative improvement claim, or provide a matched comparison with appropriate uncertainty quantification.
  2. [Section 4, Section 4.1, Appendix A.2.2] The offline evaluation in Section 4.1 likely leaks information from the RPNN training set. The RPNN is trained on DIII-D data collected between 2010 and 2019 (Section 4), while the 281-shot offline evaluation set is drawn from experiments between 2012 and 2023 (Appendix A.2.2). The paper does not report a temporal split or any exclusion of training shots from the evaluation set. Because the RPNN supplies the GP prior mean in Eq. (6), the offline results in Figure 2 can reward a prior that has already seen the test outcomes, inflating DynaBO's performance relative to the RPNN-only baseline and the vanilla GP baselines. The authors should quantify the overlap between the RPNN training data and the 281-shot evaluation set, and ideally rerun the offline comparison on a temporally disjoint test set (for example, shots executed after 2019) or otherwise demonstrate that the prior mean is not fitted to the evaluated outcomes.
  3. [Section 3.2, Appendix A.1] The assumption that the achieved normalized plasma pressure beta_N is independent of the ECH profile a_q is load-bearing but unexamined. The paper justifies this by saying that beta_N is largely determined by neutral beams through feedback control, but Appendix A.1 itself notes that ECH can reduce electron density via density pump-out, which can in turn affect beta_N. In the GP training data, beta_bar_N is the achieved average pressure and a_q is the achieved ECH profile; if ECH perturbs beta_bar_N, the context and action are correlated, the GP in Eqs. (6)-(7) is fitting a confounded mapping, and the contextual BO separation in Eq. (8) is no longer well-defined. The authors should test this assumption directly, for example by regressing achieved beta_bar_N on a_q within the 281-shot historical set and within the eight live shots, and should discuss the sensitivity of their conclusions to violations of this exogeneity assumption.
minor comments (6)
  1. [Appendix A.1] The citation '(Wang et al., 2017)' for density pump-out does not appear in the reference list; please add the reference or remove the citation.
  2. [Section 3.2, Eqs. (6)-(7)] The noise variance is introduced as sigma^2_no, but Eq. (7) writes sigma_no without the square, creating a notation inconsistency.
  3. [Section 4.2] The paper should define how time-to-tearing is computed for shots in which no tearing occurs; if stable shots are assigned the full shot length or capped at 4-5 s, this should be stated explicitly, because it affects the reported mean stability time and the offline regret in Eq. (9).
  4. [Figure 2] The main cumulative-regret comparison is shown without error bars or a statement of the number of random seeds; given the stochastic elements in the offline emulation, the variance across runs should be reported, as is done in Appendix A.4 for the acquisition-function comparison.
  5. [Section 4.2, Introduction] The Introduction says the method found stabilizing values 'despite changes to other actuators,' while Section 4.2 says the remaining settings were kept identical; this apparent contradiction should be clarified.
  6. [Section 3.2] The definition of the RPNN prior mean \(\hat{t}_{TM}\) has a formatting/notation issue: the expectation is not over an explicitly defined probability measure, and the condition \(T_t \ge 0.5\) is a deterministic threshold on a Bernoulli probability rather than an event; please clarify the stochastic rollout procedure.

Circularity Check

1 steps flagged · score 4.0 of 10

Offline validation is partially circular via overlapping RPNN/classifier training data; live DIII-D claim is not circular.

  1. fitted input called prediction [Section 4.1 (Offline Data Analysis); Appendix A.2.1-A.2.2]
    "We use a fixed RPNN in all experiments, trained using 15,000 one-step state transition observations collected between 2010 and 2019 at the DIII-D tokamak. ... we employ data from 281 past experiments carried out at the DIII-D tokamak between 2012 and 2023. ... After selecting an ECH profile, we treat the historical data point corresponding to that particular ECH profile as a new observation, which we use to update our GP model."

    The RPNN and the random-forest tearing classifier h are fit on 2010-2019 DIII-D observations (A.2.1: 'We add tearing mode labels to this dataset and train a random forest classifier'), while the offline emulator's ground-truth t_TM comes from 281 shots spanning 2012-2023, with no stated exclusion of overlapping shots. For overlapping shots, the prior mean t_hat is produced by models already trained on those trajectories' state transitions and tearing labels, and the same historical point's t_TM is then entered as the GP 'new observation'.

full rationale

The live DIII-D comparison is not circular: the GP is conditioned on 125 historical shots, then updated with eight genuinely new online measurements, and the 61-shot historical baseline is an external reference. The 117% figure is arithmetic (4/8 vs. 14/61 avoidance), and the paper itself concedes in Sec. 4.2 that 'this number of shots is too low to be statistically significant'—a statistical weakness, not a circular loop. The Sec. 3.2 assumption that achieved beta_N is independent of a_q is a model-misspecification risk (contradicted in part by the density pump-out noted in A.1), not a definitional circle. The one real circular element is the offline evaluation: the RPNN and the tearing classifier are fit on 2010-2019 DIII-D data, the 281-shot offline emulator spans 2012-2023 with no stated exclusion of the training shots, and the same historical point is then used as the 'new observation' to update the GP. For overlapping shots, DynaBO's prior mean is an in-sample fit and the reported offline advantage over baselines is inflated by construction. This makes the offline claim partially circular while leaving the live claim independently grounded.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The central claim rests on standard ML modeling assumptions and two domain assumptions that are load-bearing: the independence of βN from ECH and the unique specification of feedforward/feedback actions by β̄N. The GP hyperparameters and the tearing threshold are hand-chosen. No new physical entities are postulated. The most fragile elements are the independence assumption and the unhandled right-censoring of stable shots.

free parameters (6)
  • GP kernel length scale = Multiple values tested (0.01, 0.1, 1 for RBF; Matern nu 0.5, 2.5)
    The GP kernel length scale is a hyperparameter chosen by hand; no automatic selection via marginal likelihood is described.
  • GP noise variance σ_no = Not reported
    The noise variance in Eq (7) is a hyperparameter of the GP; its value is not given or estimated.
  • UCB exploration weight α = Not reported
    α in Eq (8) balances exploration and exploitation; the paper does not state its value or tuning procedure.
  • Tearing probability threshold 0.5 = 0.5
    The threshold for declaring a tearing mode in the RPNN rollouts (definition of t̂_TM) is set to 0.5 by hand.
  • ECH Gaussian parameterization (center, width, amplitude) = Optimized variables
    The ECH action space is restricted to stationary Gaussian profiles, parameterized by center, width, and amplitude; this is a modeling choice that limits the search space.
  • Bin margin ε for β̄N = 0.04
    The margin for binning RPNN predictions by β̄N in Appendix A.6 is chosen as 0.04.
assumptions (7)
  • domain assumption Tokamak dynamics follow the discrete-time stochastic system s_{t+1} ~ Π(s_t, a_t) (Eq 1)
    The problem is framed as a discrete-time controlled stochastic process; this is the starting model.
  • domain assumption Tearing mode occurrence follows a Bernoulli distribution T_t ~ Bernoulli(p(s_t, a_t)) (Eq 2)
    The probability of a tearing mode is modeled as a Bernoulli random variable depending on state and action.
  • domain assumption Achieved normalized plasma pressure βN is independent of the ECH profile a_q (Sec 3.2)
    This justifies treating β̄N as an exogenous context in the GP; it is questionable given ECH density pump-out effects noted in A.1.
  • domain assumption Feedforward and feedback actions are uniquely specified by the target β̄N (Sec 3.2)
    The paper approximates the full action vector by the target pressure to reduce the GP input space.
  • domain assumption ECH profiles can be approximated by a stationary Gaussian curve a_q ≈ a_ech_t (Sec 3.1)
    The continuous ECH profile is parameterized by three Gaussian parameters to enable optimization.
  • domain assumption Time-to-tearing t_TM is a valid proxy for the probability of a tearing mode (Sec 3.2)
    The GP predicts t_TM as a regression target; for stable shots this is right-censored, which the model does not handle.
  • standard math Squared-exponential kernel is appropriate for the GP (Sec 3.2)
    A standard kernel choice for continuous functions; the paper also tests other kernels for robustness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks." pith.science (2026). https://pith.science/paper/VL3EP3BJ

@misc{pith2026250610287,
  author       = {Pith},
  title        = {Pith review of: Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VL3EP3BJ}},
  note         = {Machine review of arXiv:2506.10287}
}
read the original abstract

Machine learning algorithms often struggle to control complex real-world systems. In the case of nuclear fusion, these challenges are exacerbated, as the dynamics are notoriously complex, data is poor, hardware is subject to failures, and experiments often affect dynamics beyond the experiment's duration. Existing tools like reinforcement learning, supervised learning, and Bayesian optimization address some of these challenges but fail to provide a comprehensive solution. To overcome these limitations, we present a multi-scale Bayesian optimization approach that integrates a high-frequency data-driven dynamics model with a low-frequency Gaussian process. By updating the Gaussian process between experiments, the method rapidly adapts to new data, refining the predictions of the less reliable dynamical model. We validate our approach by controlling tearing instabilities in the DIII-D nuclear fusion plant. Offline testing on historical data shows that our method significantly outperforms several baselines. Results on live experiments on the DIII-D tokamak, conducted under high-performance plasma scenarios prone to instabilities, shows a 50% success rate, marking a 117% improvement over historical outcomes.

Figures

Figures reproduced from arXiv: 2506.10287 by the authors.

Figure 1
Figure 1. DynaBO Pipeline to generate feedforward trajectory actions. Initial conditions and feedforward actuators are first used by the RPNN to generate rollouts through which we compute the prior mean of the objective function (time to tearing instability). Our Bayesian optimization algorithm uses this to optimize for actions (ECH). Noisy outputs from the Tokamak are then used to update the Gaussian process model used for B… view at source ↗
Figure 2
Figure 2. Cumulative Regret (log scale) achieved by DynaBO (green), DynaBO with a time-dependent kernel (gray), the RPNN only (blue), vanilla GP with a zero-mean prior (red) and vanilla GP with mean value as prior (orange) using six different kernels. ilar to our experiment), however, they use manual methods for selecting ECH profiles. The average time to tearing insta￾bility in these experiments is 2424 milliseconds, well be… view at source ↗
Figure 3
Figure 3. ECH Profiles queried by different methods during simulated offline runs using a Gaussian kernel. We see that DynaBO and DynaBO with a time-dependent GP explore the most, highlighting the importance of our dynamic model prior mean. Experiment ID Target β¯N Tearing Instability Stability Time (Shotnumber) Avoided (ms) 199599 3.37 Yes 4566 199601 3.27 Yes 4632 199602 3.27 No 2107 199603 3.27 No 2149 199604 3.27 Yes 4592… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Gyrotron action on the Plasma inside the Tokamak. The bottom 2 curves indicate the power absorbed (heating profile) and current driven in the plasma from the centre to outer region of the plasma. – FC layer: 512 × 128 • Output Heads: – Mean head: 128 × output dim – Log…
Figure 5
Figure 5. Figure 5: Identifying Tearing modes from Raw signal data. We use n1rms signal (denotes magnetic perturbations) and normalized plasma pressureβN to identify tearing modes. A sustained high n1rms signal denotes tearing modes. We label the drop in βN due to tearing mode formation. …
Figure 6
Figure 6. Figure 6: In this figure we also include the power injected (pinj) along with n1rms and βN . in 199606, we see a very late tearing mode which occurs just before power injected is dropped. Very low loss in βN is seen due to the tearing mode. Finally, in 199607 no tearing modes ar…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 28 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]

    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...

  3. [3]

    Data-driven profile prediction for diii-d

    Abbate, J., Conlin, R., and Kolemen, E. Data-driven profile prediction for diii-d. Nuclear Fusion, 61 0 (4): 0 046027, 2021

    work page 2021

  4. [4]

    Direct preemptive stabilization of neoclassical tearing modes by electron cyclotron current drive in the diii-d low-torque iter baseline scenario

    Bard \'o czi, L., La Haye, R., Strait, E., Logan, N., Smith, S., Richner, N., and Callen, J. Direct preemptive stabilization of neoclassical tearing modes by electron cyclotron current drive in the diii-d low-torque iter baseline scenario. Nuclear Fusion, 63 0 (9): 0 096021, 2023

    work page 2023

  5. [5]

    Machine learning for tokamak scenario optimization: combining accelerating physics models and empirical models

    Boyer, M., Wai, J., Clement, M., Kolemen, E., Char, I., Chung, Y., Neiswanger, W., and Schneider, J. Machine learning for tokamak scenario optimization: combining accelerating physics models and empirical models. In APS Division of Plasma Physics Meeting Abstracts, volume 2021, pp.\ PP11--164, 2021

    work page 2021

  6. [6]

    O., Boyer, M., Kolemen, E., and Schneider, J

    Char, I., Chung, Y., Neiswanger, W., Kandasamy, K., Nelson, A. O., Boyer, M., Kolemen, E., and Schneider, J. Offline contextual bayesian optimization. In Wallach, H., Larochelle, H., Beygelzimer, A., d Alch\' e -Buc, F., Fox, E., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019

    work page 2019

  7. [7]

    Offline model-based reinforcement learning for tokamak control

    Char, I., Abbate, J., Bard \'o czi, L., Boyer, M., Chung, Y., Conlin, R., Erickson, K., Mehta, V., Richner, N., Kolemen, E., et al. Offline model-based reinforcement learning for tokamak control. In Learning for Dynamics and Control Conference, pp.\ 1357--1372. PMLR, 2023 a

    work page 2023

  8. [8]

    Full shot predictions for the diii-d tokamak via deep recurrent networks

    Char, I., Chung, Y., Abbate, J., Kolemen, E., and Schneider, J. Full shot predictions for the diii-d tokamak via deep recurrent networks. In APS Division of Plasma Physics Meeting Abstracts, volume 2023, pp.\ UP11--096, 2023 b

    work page 2023

Show all 37 references
  1. [9]

    Deep reinforcement learning in a handful of trials using probabilistic dynamics models

    Chua, K., Calandra, R., McAllister, R., and Levine, S. Deep reinforcement learning in a handful of trials using probabilistic dynamics models. Advances in neural information processing systems, 31, 2018

    2018

  2. [10]

    Magnetic control of tokamak plasmas through deep reinforcement learning

    Degrave, J., Felici, F., Buchli, J., Neunert, M., Tracey, B., Carpanese, F., Ewalds, T., Hafner, R., Abdolmaleki, A., de Las Casas, D., et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602 0 (7897): 0 414--419, 2022

    2022

  3. [11]

    and Rasmussen, C

    Deisenroth, M. and Rasmussen, C. E. Pilco: A model-based and data-efficient approach to policy search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pp.\ 465--472, 2011

    2011

  4. [12]

    D., Eidietis, N., Barbour, N., Izacard, O., and Kolemen, E

    Fu, Y., Eldon, D., Erickson, K., Kleijwegt, K., Lupin-Jimenez, L., Boyer, M. D., Eidietis, N., Barbour, N., Izacard, O., and Kolemen, E. Machine learning control for disruption and tearing mode avoidance. Physics of Plasmas, 27 0 (2), 2020

    2020

  5. [13]

    Off-policy deep reinforcement learning without exploration

    Fujimoto, S., Meger, D., and Precup, D. Off-policy deep reinforcement learning without exploration. In International conference on machine learning, pp.\ 2052--2062. PMLR, 2019

    2019

  6. [14]

    U., et al

    Gantenbein, G., Zohm, H., Giruzzi, G., G \"u nter, S., Leuterer, F., Maraschek, M., Meskat, J., Yu, Q., Team, A. U., et al. Complete suppression of neoclassical tearing modes with current drive at the electron-cyclotron-resonance frequency in asdex upgrade tokamak. Physical Re...

    2000

  7. [15]

    Bayesian optimization

    Garnett, R. Bayesian optimization. Cambridge University Press, 2023

    2023

  8. [16]

    Agile but safe: Learning collision-free high-speed legged locomotion

    He, T., Zhang, C., Xiao, W., He, G., Liu, C., and Shi, G. Agile but safe: Learning collision-free high-speed legged locomotion. arXiv preprint arXiv:2401.17583, 2024

    2024 arXiv

  9. [17]

    V., Okabayashi, M., Lasnier, C., Hanson, J., Politzter, P., In, Y., Hyatt, A., Haye, R

    Holcomb, C., Ferron, J., Luce, T., Petrie, T., Park, J., Turco, F., Zeeland, M. V., Okabayashi, M., Lasnier, C., Hanson, J., Politzter, P., In, Y., Hyatt, A., Haye, R. L., and Lanctot, M. Steady state scenario development with elevated minimum safety factor on diii-d. Nuclear ...

    2014 doi

  10. [18]

    M., Logan, N

    Hu, Q. M., Logan, N. C., Yu, Q., and Bortolon, A. Effects of edge-localized electron cyclotron current drive on edge-localized mode suppression by resonant magnetic perturbations in diii-d. Nuclear Fusion, 64 0 (4): 0 046027, 2024

    2024

  11. [19]

    State-of-the-art neoclassical tearing mode control in diii-d using real-time steerable electron cyclotron current drive launchers

    Kolemen, E., Welander, A., La Haye, R., Eidietis, N., Humphreys, D., Lohr, J., Noraky, V., Penaflor, B., Prater, R., and Turco, F. State-of-the-art neoclassical tearing mode control in diii-d using real-time steerable electron cyclotron current drive launchers. Nuclear Fusion,...

    2014

  12. [20]

    Rma: Rapid motor adaptation for legged robots

    Kumar, A., Fu, Z., Pathak, D., and Malik, J. Rma: Rapid motor adaptation for legged robots. Robotics: Science and Systems XVII, 2021

    2021

  13. [21]

    Learning quadrupedal locomotion over challenging terrain

    Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., and Hutter, M. Learning quadrupedal locomotion over challenging terrain. Science robotics, 5 0 (47): 0 eabc5986, 2020

    2020

  14. [22]

    Disruptions in iter and strategies for their control and mitigation

    Lehnen, M., Aleynikova, K., Aleynikov, P., Campbell, D., Drewelow, P., Eidietis, N., Gasparyan, Y., Granetz, R., Gribov, Y., Hartmann, N., et al. Disruptions in iter and strategies for their control and mitigation. Journal of Nuclear materials, 463: 0 39--48, 2015

    2015

  15. [23]

    Offline reinforcement learning: Tutorial, review

    Levine, S., Kumar, A., Tucker, G., and Fu, J. Offline reinforcement learning: Tutorial, review. and Perspectives on Open Problems, 5, 2020

    2020

  16. [24]

    D., Char, I., Neiswanger, W., Kolemen, E., and Schneider, J

    Mehta, V., Barr, J., Abbate, J., Boyer, M. D., Char, I., Neiswanger, W., Kolemen, E., and Schneider, J. Automated experimental design of safe rampdowns via probabilistic machine learning. Nuclear Fusion, 64 0 (4): 0 046014, 2024

    2024

  17. [25]

    Integrated modeling applications for tokamak experiments with omfit

    Meneghini, O., Smith, S., Lao, L., Izacard, O., Ren, Q., Park, J., Candy, J., Wang, Z., Luna, C., Izzo, V., Grierson, B., Snyder, P., Holland, C., Penna, J., Lu, G., Raum, P., McCubbin, A., Orlov, D., Belli, E., Ferraro, N., Prater, R., Osborne, T., Turnbull, A., and Staebler,...

    2015

  18. [26]

    D., Tengku Hassan, T

    Mohamed, M., Zakuan, N. D., Tengku Hassan, T. N. A., Lock, S. S. M., and Mohd Shariff, A. Global development and readiness of nuclear fusion technology as the alternative source for clean energy supply. Sustainability, 16 0 (10): 0 4089, 2024

    2024

  19. [27]

    W., Bertelli, N., Shiraiwa, S., Delgado-Aparicio, L., Menard, J

    Ono, M., Berkery, J. W., Bertelli, N., Shiraiwa, S., Delgado-Aparicio, L., Menard, J. E., S \'a nchez-Villar, \'A ., Shah, K., Shevchenko, V. F., Idei, H., et al. Efficient eccd non-inductive plasma current start-up, ramp-up, and sustainment for an st fusion reactor. Nuclear F...

    2024

  20. [28]

    M., Babaeizadeh, M., Kozakowski, P., Milos, P., et al

    Osinski, B., Finn, C., Erhan, D., Tucker, G., Michalewski, H., Czechowski, K., Kaiser, L. M., Babaeizadeh, M., Kozakowski, P., Milos, P., et al. Model-based reinforcement learning for atari. ICLR, 1: 0 2, 2020

    2020

  21. [29]

    Disruptions in tokamaks

    Schuller, F. Disruptions in tokamaks. Plasma Physics and Controlled Fusion, 37 0 (11A): 0 A135, 1995

    1995

  22. [30]

    Avoiding fusion plasma tearing instability with deep reinforcement learning

    Seo, J., Kim, S., Jalalvand, A., Conlin, R., Rothstein, A., Abbate, J., Erickson, K., Wai, J., Shousha, R., and Kolemen, E. Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature, 626 0 (8000): 0 746--751, 2024

    2024

  23. [31]

    Automation and control of laser wakefield accelerators using bayesian optimization

    Shalloo, R., Dann, S., Gruse, J.-N., Underwood, C., Antoine, A., Arran, C., Backhouse, M., Baird, C., Balcazar, M., Bourgeois, N., et al. Automation and control of laser wakefield accelerators using bayesian optimization. Nature communications, 11 0 (1): 0 6355, 2020

    2020

  24. [32]

    J., Stevens, J., Li, J., Parasram, M., Damani, F., Alvarado, J

    Shields, B. J., Stevens, J., Li, J., Parasram, M., Damani, F., Alvarado, J. I. M., Janey, J. M., Adams, R. P., and Doyle, A. G. Bayesian reaction optimization as a tool for chemical synthesis. Nature, 590 0 (7844): 0 89--96, 2021

    2021

  25. [33]

    Overview and summary

    Shimada, M., Campbell, D., Mukhovatov, V., Fujiwara, M., Kirneva, N., Lackner, K., Nagami, M., Pustovitov, V., Uckan, N., Wesley, J., et al. Overview and summary. Nuclear Fusion, 47 0 (6): 0 S1, 2007

    2007

  26. [34]

    Simultaneous control of safety factor profile and normalized beta for jt-60sa using reinforcement learning

    Wakatsuki, T., Yoshida, M., Narita, E., Suzuki, T., and Hayashi, N. Simultaneous control of safety factor profile and normalized beta for jt-60sa using reinforcement learning. Nuclear Fusion, 63 0 (7): 0 076017, 2023

    2023

  27. [35]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  28. [36]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  29. [37]

    2| h_ § 1 ܺ6B gBoٖn F<h ?|v@u>=+|E@x& ݝbm) 㙸 B/[ in<-Jai=<T 9 6/ϱŹ UYvDlTx +]lK 9 MhsH= ̿ NюR+ g

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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