Pith. sign in

REVIEW 3 major objections 5 minor 29 references

How to craft a deep reinforcement learning policy for wind farm flow control

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A deep reinforcement learning policy that combines graph attention networks with multi-head self-attention increases wind farm energy production by up to 14% in high wake-loss conditions while training about 10 times faster than a fully…

desk verdict A carefully evaluated RL wake-steering architecture with a real but narrower contribution than the headline generalization claim. read the letter →

arxiv 2506.06204 v2 pith:GJDJ3NLN submitted 2025-06-06 cs.LG

Elie Kadoche , Pascal Bianchi , Florence Carton , Philippe Ciblat , Damien Ernst This is my paper
classification cs.LG
keywords deepreinforcementlearningwakesteeringwindfarmflowcontrolyawgraphattentionnetworkmulti-headself-attentionenergyproductionproximalpolicyoptimization
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 tries to establish that a single deep reinforcement learning policy, built by combining graph attention networks with multi-head self-attention blocks, can control yaw angles across an entire wind farm to increase energy production by up to 14% in high wake-loss conditions, while requiring roughly 10 times fewer training steps than a fully connected neural network. The authors argue that previous RL-based wake steering approaches failed to combine scalability, sample efficiency, and generalization across the full range of time-varying wind directions. If the claim holds, wake steering control becomes computationally cheap enough for real-time use and general enough to deploy without retuning per wind condition. The paper's evidence comes from a steady-state, low-fidelity simulator with 19 turbines and a wind process generated by an auto-regressive moving average model.

What carries the argument

The load-bearing mechanism is the V2 architecture: a graph attention network produces positional embeddings for the turbines, multi-head self-attention blocks let the policy attend to whichever turbines are most relevant for wake interactions, and a set of feed-forward layers maps the summed embeddings to a von Mises distribution over yaw actions. This is paired with a reward function that normalizes power gains against a wind-tracking baseline and exponentially discounts them when baseline wake losses are large, plus a training scheme that starts episodes from every one-degree wind direction to force coverage of the full circular range.

What would settle it

Run the trained V2 policy on a higher-fidelity unsteady wake simulator or on a different wind model, for example with abrupt direction shifts or diurnal speed variations, and measure the energy gain relative to standard wind tracking. If the gain does not remain positive across most wind directions, or if the reported up-to-14% improvement disappears, the claim of generalizing to any time-varying wind conditions is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that a model called V2, which splits the wind farm state into separate embeddings for wind, forecast, turbine positions, and turbine orientations, then reasons over the farm with a graph attention network followed by three multi-head self-attention blocks, learns a wake steering policy that generalizes across all 360 degrees of wind direction under unsteady and noisy conditions. The policy outputs a von Mises distribution over yaw offsets for each turbine, trained with PPO on a reward that compares power output to a perfect wind-tracking baseline and applies an exponential scaling by baseline wake losses. In testing on 360 wind directions, V2 consistently outperforms a fully connected network (V0) and a plain graph attention network (V1), matches or exceeds a strong heuristic optimizer in strong wake conditions with lower variance, and increases energy production by up to 14% relative to standard wind tracking. The paper claims this is the first deep RL-based wake steering controller to generalize effectively across any time-varying wind conditions in a steady-state, low-fidelity numerical simulation.

Load-bearing premise

The paper assumes that a simple one-step auto-regressive wind time series and the steady-state wake simulator faithfully represent the time-varying wind conditions in which the policy would be deployed, and it only tests generalization within that same wind process.

Editorial extensions

If this is right

  • A single trained V2 policy can provide yaw settings for all turbines in real time because evaluating it costs about 200 times less computation than the heuristic optimizer.
  • Training time drops by roughly a factor of 10 compared to a fully connected network, making RL-based wake steering practical for farms with dozens of turbines.
  • The policy's performance across the full 360-degree wind range means no separate controllers or lookup tables are needed for different wind conditions.
  • The reward design, which scales power gains by baseline wake losses, is what allows the same policy to perform well in both low-wake and high-wake regimes.

Reading between the lines

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

  • Inference: The same split-embedding plus self-attention recipe could transfer to other spatial control problems with a fixed sensor layout, such as coordinated robotics or airflow management, where learned attention weights would reveal which pairwise interactions dominate.
  • Inference: The baseline-relative, wake-scaled reward could be reused in any control task where the benefit of intervention depends on a measurable operating condition; the exponential term may act as an automatic curriculum.
  • Inference: Because the von Mises policy head respects the circular nature of angles, replacing a Gaussian policy head with it might improve other rotational control tasks before any architectural change is made.
  • Inference: The paper's generalization claim is untested beyond the single wind model used for training; an immediate next test is to train on that model and evaluate on a different one, such as wind with gusts or diurnal cycles, to see whether the attention mechanism generalizes or merely memorizes the training dynamics.
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 / 5 minor

Summary. The paper proposes a deep RL wake-steering controller (V2) for wind farms, combining a GAT-based positional encoding with multi-head self-attention, a von Mises policy head, and a reward that mixes an invalid-yaw penalty with a wake-loss-scaled power ratio. The controller is trained with PPO in a steady-state FLORIS simulation of a 19-turbine farm under an ARMA(1) wind process and is evaluated against standard wind tracking, Gauss-Seidel, and a heuristic forecast-based optimizer. The authors report roughly 10x sample-efficiency gains over an FNN policy, lower variance than the heuristic baseline, and up to 14% energy improvement in high-wake-loss conditions, and they claim complete generalization over time-varying wind conditions.

Significance. If the empirical claims hold, the work is a useful step toward practical RL-based wake steering: the 360-direction evaluation protocol and 10-seed training are careful, the reward ablation in Figure 9 gives useful design insight, and the proposed architecture is plausible and clearly motivated. However, the headline generalization claim is currently stronger than the evidence because training and testing use the same wind process, and the small discount factor makes the stated long-term-optimization mechanism questionable. The paper's value would increase substantially if the evaluation covered held-out wind dynamics and if the sample-efficiency claim were quantified.

major comments (3)
  1. [Section 4.1 and Section 5] The abstract and Section 5 claim that the policy "generalizes effectively across any time-varying wind conditions", but the experiments only train and test on a single ARMA(1) process with fixed parameters (K_t = (epsilon_t + K_{t-1} + 0.1 epsilon_{t-1}) mod 360 with epsilon_t ~ N(0,9), and the analogous speed process). The test protocol in Section 4.3 varies only the initial wind direction over a 1-degree grid and the random seeds; it does not vary the ARMA coefficients, noise levels, speed range, or the class of wind process (e.g., sinusoidal, measured data, or a different correlation structure). The policy is therefore evaluated in-distribution within one stochastic process, not "across any time-varying wind conditions". This is a load-bearing overstatement and should either be supported by held-out wind-dynamics experiments or replaced by a precise claim such as "across initial wind directions within the modeled ARMA process".
  2. [Section 4.2, Table 1] The discount factor is set to gamma = 0.1 and is not ablated. With this value, the total weight of all non-immediate rewards is only 0.1 + 0.01 + ... = 0.111, i.e., about 10% of the discounted return, making the policy effectively myopic. This is at odds with the paper's mechanism claims: Section 4.3 explains the GS baseline's poor performance by saying "wind direction can shift too rapidly relative to yaw constraints, making long-term optimization essential", and the V2 model is said to "leverage noisy wind forecasts to improve long-term performance", while the state contains a three-step forecast. A gamma = 0.1 policy cannot plausibly optimize over the three-hour horizon that the episode represents. Please either justify this choice with an ablation over gamma values or revise the claims about long-term optimization and forecast use.
  3. [Section 4.3, Figure 6] The claimed "approximately 10 times fewer training steps than a fully connected neural network" (Contribution 1, Section 1.3) is not supported by the reported measurements. All models are trained for the same 150 PPO steps, and no convergence criterion, time-to-threshold metric, or final-performance comparison that would quantify a 10x difference is reported. The training curves in Figure 6 are qualitative and, as presented, show that V2 converges faster and ends higher, but not that it needs 10x fewer steps to reach a given performance level. Please report a quantitative sample-efficiency measure (e.g., steps to reach V0's final mean performance) or temper the claim to "faster convergence".
minor comments (5)
  1. [Appendix A.5, Algorithm 1] The pseudocode has sign errors that would confuse an implementation: "Fix parameters theta_old <- -theta" should be "theta_old <- theta", and "Update parameters theta <- -theta - alpha * grad_theta L" should be "theta <- theta - alpha * grad_theta L".
  2. [Section 4.1] The mirroring function used to keep generated wind speeds in [V_min, V_max] is not defined; please describe it explicitly, since it affects the transition dynamics used in both training and testing.
  3. [Section 4.3] The statement that the V2 model is "roughly 200 times more computationally efficient" than the heuristic baseline needs a precise definition of the metric (e.g., wall-clock time per control step, number of FLORIS evaluations, or total optimization cost), otherwise it is not verifiable.
  4. [Section 2] The transition function is written as P : S x A -> S, which suggests a deterministic transition, but the environment is stochastic due to the ARMA wind process and observation noise; please write the transition as a probability kernel or otherwise clarify.
  5. [Section 4.2] The text says "we achieve a 70 speedup in training" in Section 4.2; this should read "70x speedup".

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain; the central V2 vs V0/V1 comparison and reward design are self-contained, with only minor non-load-bearing self-citations.

full rationale

The paper's derivation chain is self-contained. The reward in Eq. (3) maximizes a normalized power ratio relative to a perfect wind-tracking baseline computed inside FLORIS; no reward term, state variable, or architecture choice is defined in terms of the reported performance gains, so the 14% improvement is an empirical outcome rather than a construction. The V2 model is compared against a FNN (V0) and a GAT (V1) trained under the same PPO procedure and roughly matched parameter counts, and testing uses held-out random seeds on the same 1-degree wind-direction grid used in training; the central comparison is internal and not forced by any fitted parameter renamed as a prediction. The only self-citations are [9] (MARLYC) and [17] (the heuristic baseline). Reference [9] is cited only as prior work with a limited wind-direction range, and [17] is cited only to attribute the heuristic benchmark. That heuristic is a published, independently implementable algorithm and is not an input to the model, reward, or training loop, so the self-citation is not load-bearing. The broader claim of generalization 'across any time-varying wind conditions' is not established because both training and testing use the same single ARMA(1) wind process; however, that is an external-validity limitation, not a circular reduction. The paper itself acknowledges this in the conclusion, noting that the results are based on simplified, steady-state, low-fidelity models and that validation in higher-fidelity and unsteady environments remains future work. No equation reduces to itself, and no fitted input is renamed as a prediction.

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

The central claim rests on hand-chosen reward and training hyperparameters (notably p and gamma), a specific wind process, and a low-fidelity simulator. No new physical entities are introduced.

free parameters (4)
  • p = 3
    Exponential scaling parameter in reward (Eq. 3). Chosen by hand; ablation shows it changes performance, especially in low wake-loss conditions.
  • gamma = 0.1
    Discount factor; justified by steady-state assumption but not ablated. Affects how future rewards are weighted.
  • reward weights (w0, w1) = 1, 100
    Weights in reward (Eq. 1); set by hand without ablation.
  • ARMA noise scales = direction N(0,9), speed N(0,0.01), observation U(-3,3) and U(-0.1,0.1)
    Parameters of wind time series generator; chosen to mimic variability but not validated against real data.
assumptions (4)
  • domain assumption FLORIS Gaussian-curl hybrid model is a valid approximation of wake dynamics
    Used for all power computations; fidelity not validated against real farm data.
  • domain assumption ARMA(1) wind model with specified noise is representative of time-varying wind conditions
    The claimed generalization to any time-varying conditions rests on this model; no real wind data used.
  • domain assumption Graph edge construction (8-diameter threshold, upstream edges) captures relevant wake interactions
    Incorporated expert knowledge; not varied or validated.
  • standard math PPO with the listed hyperparameters converges to a near-optimal policy for this MDP
    Assumes RL training is effective; results are empirical.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How to craft a deep reinforcement learning policy for wind farm flow control." pith.science (2026). https://pith.science/paper/GJDJ3NLN

@misc{pith2026250606204,
  author       = {Pith},
  title        = {Pith review of: How to craft a deep reinforcement learning policy for wind farm flow control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJDJ3NLN}},
  note         = {Machine review of arXiv:2506.06204}
}
read the original abstract

Within wind farms, wake effects between turbines can significantly reduce overall energy production. Wind farm flow control encompasses methods designed to mitigate these effects through coordinated turbine control. Wake steering, for example, consists in intentionally misaligning certain turbines with the wind to optimize airflow and increase power output. However, designing a robust wake steering controller remains challenging, and existing machine learning approaches are limited to quasi-static wind conditions or small wind farms. This work presents a new deep reinforcement learning methodology to develop a wake steering policy that overcomes these limitations. Our approach introduces a novel architecture that combines graph attention networks and multi-head self-attention blocks, alongside a novel reward function and training strategy. The resulting model computes the yaw angles of each turbine, optimizing energy production in time-varying wind conditions. An empirical study conducted on steady-state, low-fidelity simulation, shows that our model requires approximately 10 times fewer training steps than a fully connected neural network and achieves more robust performance compared to a strong optimization baseline, increasing energy production by up to 14 %. To the best of our knowledge, this is the first deep reinforcement learning-based wake steering controller to generalize effectively across any time-varying wind conditions in a low-fidelity, steady-state numerical simulation setting.

Figures

Figures reproduced from arXiv: 2506.06204 by the authors.

Figure 1
Figure 1. Example of WFFC on a three turbines wind farm viewed from above. The first two turbines [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Graphical representation of a turbine i at time step t with wind direction Kt. After action a i t is being applied, the turbine has an updated yaw angle α˜ i t . At the end of the time step, wind direction evolves from Kt to Kt+1 and the next yaw angle α i t+1 is computed. 1. the policy computes the yaw settings such that πθ(st) = at; 2. absolute orientations are updated β i t+1 = (β i t + a i t ) mod 360; 3. yaw an… view at source ↗
Figure 3
Figure 3. V0 model, a FNN-based architecture. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: V1 model, based on GATs. Sub-Figure 4a illustrates an example input graph with a wind [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: V2 model architecture, incorporating GAT and MHSA blocks. FC layers refer to standard [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Training curves of each model, showing mean and variance over 10 different seeds. The V1 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Performance of each solution relative to the standard benchmark. The V2 model significantly [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Training time of the V2 model under different resource configurations. Parallelizing [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Ablation study for the V2 model. Without the exponential term, i.e., with [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 18 canonical work pages

  1. [1]

    Field test of wake steering at an offshore wind farm

    P. Fleming et al. “Field test of wake steering at an offshore wind farm”. In: Wind Energy Science 2.1 (2017), pp. 229–239. DOI: 10.5194/wes-2-229-2017 . URL: https://wes. copernicus.org/articles/2/229/2017/

    work page doi:10.5194/wes-2-229-2017 2017

  2. [2]

    Maximum power extraction for wind turbines through a novel yaw control solution using predicted wind directions

    Dongran Song et al. “Maximum power extraction for wind turbines through a novel yaw control solution using predicted wind directions”. In: Energy Conversion and Management 157 (2018), pp. 587–599. ISSN : 0196-8904. DOI: https://doi.org/10.1016/j.enconman. 2017 . 12 . 019. URL: https : / / www . sciencedirect . com / science / article / pii / S0196890417311676

    doi:10.1016/j.enconman 2018

  3. [3]

    Data-driven wind farm flow control and challenges towards field imple- mentation: A review

    Tuhfe Göçmen et al. “Data-driven wind farm flow control and challenges towards field imple- mentation: A review”. In: Renewable and Sustainable Energy Reviews 216 (2025), p. 115605. ISSN : 1364-0321. DOI: https : / / doi . org / 10 . 1016 / j . rser . 2025 . 115605. URL: https://www.sciencedirect.com/science/article/pii/S1364032125002783

    work page 2025

  4. [4]

    Model-free closed-loop wind farm control using reinforcement learning with recursive least squares

    Jaime Liew et al. “Model-free closed-loop wind farm control using reinforcement learning with recursive least squares”. In: Wind Energy (2023)

    work page 2023

  5. [5]

    Actor Critic Agents for Wind Farm Control

    Claire Bizon Monroc et al. “Actor Critic Agents for Wind Farm Control”. In: 2023 American Control Conference (ACC). 2023, pp. 177–183. DOI: 10.23919/ACC55779.2023.10156453

    arXiv 2023

  6. [6]

    Intelligent wind farm control via deep reinforcement learning and high-fidelity simulations

    Hongyang Dong, Jincheng Zhang, and Xiaowei Zhao. “Intelligent wind farm control via deep reinforcement learning and high-fidelity simulations”. In: Applied Energy 292 (2021), p. 116928. DOI: 10.1016/j.apenergy.2021.116928

    work page arXiv 2021

  7. [7]

    A Distributed Reinforcement Learning Yaw Control Approach for Wind Farm Energy Capture Maximization*

    P. Stanfel et al. “A Distributed Reinforcement Learning Yaw Control Approach for Wind Farm Energy Capture Maximization*”. In: 2020 American Control Conference (ACC). 2020, pp. 4065–4070. DOI: 10.23919/ACC45564.2020.9147946

    arXiv 2020

  8. [8]

    Deep Reinforcement Learning for Active Wake Control

    Grigory Neustroev et al. “Deep Reinforcement Learning for Active Wake Control”. In:Pro- ceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems . AAMAS ’22. Virtual Event, New Zealand: International Foundation for Autonomous Agents and Multiagent Systems, 2022, pp. 944–953. ISBN : 978-145-039-2-1-3-6

    work page 2022

Show all 29 references
  1. [9]

    MARLYC: Multi-Agent Reinforcement Learning Yaw Control

    Elie Kadoche et al. “MARLYC: Multi-Agent Reinforcement Learning Yaw Control”. In: Renewable Energy 217 (2023), p. 119129. ISSN : 0960-1481. DOI: https://doi.org/10. 1016/j.renene.2023.119129 . URL: https://www.sciencedirect.com/science/ article/pii/S0960148123010431

    2023

  2. [10]

    Deep reinforcement learning-based adaptive yaw control for wind farms in fluctuating winds

    Qiang Dong et al. “Deep reinforcement learning-based adaptive yaw control for wind farms in fluctuating winds”. In: Physics of Fluids 37.4 (Apr. 2025), p. 047157. ISSN : 1070-6631. DOI: 10 . 1063 / 5 . 0267200. eprint: https : / / pubs . aip . org / aip / pof / article - pdf /...

    2025 doi

  3. [11]

    FALCON- FArm Level CONtrol for wind turbines using multi-agent deep reinforcement learning

    Venkata Ramakrishna Padullaparthi et al. “FALCON- FArm Level CONtrol for wind turbines using multi-agent deep reinforcement learning”. In: Renewable Energy (2021). ISSN : 0960-

    2021

  4. [12]

    Learning to optimise wind farms with graph transformers

    Siyi Li et al. “Learning to optimise wind farms with graph transformers”. In: Applied Energy 359 (2024), p. 122758. ISSN : 0306-2619. DOI: https://doi.org/10.1016/j.apenergy. 2024 . 122758. URL: https : / / www . sciencedirect . com / science / article / pii / S0306261924001417

    2024 doi

  5. [13]

    Graph Attention Networks

    Petar Veliˇckovi´c et al. “Graph Attention Networks”. In: International Conference on Learning Representations. 2018. URL: https://openreview.net/forum?id=rJXMpikCZ

    2018

  6. [14]

    Attention is All you Need

    Ashish Vaswani et al. “Attention is All you Need”. In: Advances in Neural Informa- tion Processing Systems . Ed. by I. Guyon et al. V ol. 30. Curran Associates, Inc., 2017. URL: https : / / proceedings . neurips . cc / paper _ files / paper / 2017 / file / 3f5ee243547dee91fbd0...

    2017

  7. [15]

    NREL. FLORIS. V ersion 4.2.2. 2021. URL: https://github.com/NREL/floris

    2021

  8. [16]

    Serial-Refine Method for Fast Wake-Steering Yaw Optimization

    Paul A. Fleming et al. “Serial-Refine Method for Fast Wake-Steering Yaw Optimization”. In: Journal of Physics: Conference Series 2265.3 (May 2022), p. 032109. DOI: 10.1088/1742- 6596/2265/3/032109. URL: https://doi.org/10.1088/1742-6596/2265/3/032109

    2022 doi

  9. [17]

    On the importance of wind predictions in wake steering optimization

    Elie Kadoche et al. “On the importance of wind predictions in wake steering optimization”. In: Wind Energy Science 9.7 (2024), pp. 1577–1594. DOI: 10.5194/wes-9-1577-2024 . URL: https://wes.copernicus.org/articles/9/1577/2024/. 10

    2024 doi

  10. [18]

    Proximal Policy Optimization Algorithms

    John Schulman et al. Proximal Policy Optimization Algorithms. 2017. arXiv: 1707.06347 [cs.LG]

    2017 arXiv

  11. [19]

    High-Dimensional Continuous Control Using Generalized Advantage Estimation

    John Schulman et al. “High-Dimensional Continuous Control Using Generalized Advantage Estimation”. In: 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings . Ed. by Yoshua Bengio and Yann LeCun....

    2016 arXiv

  12. [20]

    Control-oriented model for secondary effects of wake steering

    J. King et al. “Control-oriented model for secondary effects of wake steering”. In: Wind Energy Science 6.3 (2021), pp. 701–714. DOI: 10.5194/wes-6-701-2021 . URL: https: //wes.copernicus.org/articles/6/701/2021/

    2021 doi

  13. [21]

    IEA Wind TCP Task 37: Definition of the IEA 15-Megawatt Offshore Reference Wind Turbine

    Evan Gaertner et al. “IEA Wind TCP Task 37: Definition of the IEA 15-Megawatt Offshore Reference Wind Turbine”. In: (Mar. 2020). DOI: 10.2172/1603478 . URL: https://www. osti.gov/biblio/1603478

    2020

  14. [22]

    Stable-Baselines3: Reliable Reinforcement Learning Implementations

    Antonin Raffin et al. “Stable-Baselines3: Reliable Reinforcement Learning Implementations”. In: Journal of Machine Learning Research 22.268 (2021), pp. 1–8. URL: http://jmlr.org/ papers/v22/20-1364.html

    2021

  15. [23]

    RLlib: Abstractions for Distributed Reinforcement Learning

    Eric Liang et al. “RLlib: Abstractions for Distributed Reinforcement Learning”. In: Interna- tional Conference on Machine Learning (ICML) . 2018. URL: https://arxiv.org/pdf/ 1712.09381

    2018 arXiv

  16. [24]

    RLlib Flow: Distributed Reinforcement Learning is a Dataflow Problem

    Zhanghao Wu et al. “RLlib Flow: Distributed Reinforcement Learning is a Dataflow Problem”. In: Conference on Neural Information Processing Systems (NeurIPS)

  17. [25]

    Array programming with NumPy

    Charles R. Harris et al. “Array programming with NumPy”. In: Nature 585.7825 (Sept. 2020), pp. 357–362. ISSN : 1476-4687. DOI: 10.1038/s41586-020-2649-2 . URL: https: //doi.org/10.1038/s41586-020-2649-2

    2020 doi

  18. [26]

    Ray: A Distributed Framework for Emerging AI Applications

    Philipp Moritz et al. “Ray: A Distributed Framework for Emerging AI Applications”. In:CoRR abs/1712.05889 (2017). arXiv: 1712.05889. URL: http://arxiv.org/abs/1712.05889. 11 A Appendix A.1 Detailed architectures V0 model The V0 model, a FNN-based architecture, is presented in ...

    2017 arXiv

  19. [29]

    During training, actions are sampled independently for each turbine as ai t ∼ V(µi, κi), while during evaluation, actions are set deterministically to the mode, ai t = µi

    and the concentration parameter as κi = softplus(pi 1). During training, actions are sampled independently for each turbine as ai t ∼ V(µi, κi), while during evaluation, actions are set deterministically to the mode, ai t = µi. The use of von Mises distributions ensures that t...

  20. [1481]

    DOI: 10.1016/j.renene.2021.09.023

    2021 doi

  21. [2021]

    URL: https : / / proceedings . neurips . cc / paper / 2021 / file / 2bce32ed409f5ebcee2a7b417ad9beed-Paper.pdf

    2021

Pith tools

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