pith. sign in

arxiv: 2606.24954 · v1 · pith:GOJXF37Bnew · submitted 2026-06-23 · 💻 cs.LG · cs.CL

Digital Twin-Driven Adaptive Sim-to-Real Alignment via Reinforcement Learning for Vibration-Based Bearing Health Monitoring Under Data Scarcity

Jinghan Wang , Yanjun Chen , Wei Zhang , Wentao Wu , Tianchen Liu , Gaoliang Peng This is my paper

Pith reviewed 2026-06-26 00:52 UTC · model grok-4.3

classification 💻 cs.LG cs.CL
keywords digital twinreinforcement learningbearing fault diagnosissim-to-real alignmentvibration monitoringproximal policy optimizationdata scarcitydomain adaptation
0
0 comments X

The pith

Feature alignment as a continuous-action MDP solved by PPO enables fault-specific sim-to-real corrections that preserve class separability in bearing monitoring.

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

The paper establishes that treating sim-to-real feature alignment as a sequential decision process allows reinforcement learning to apply adaptive, fault-type-specific transformations to simulated vibration signals. This approach addresses the limitation of global transformations that cannot handle heterogeneous gaps across fault classes without affecting separability. A reader would care because it provides a way to use abundant simulated data to supplement scarce real fault data in digital twin setups for machinery health monitoring. The method uses an asymmetry-aware strategy that keeps real normal data untouched while aligning faults. Experiments on multiple bearing datasets show the RL method outperforms static approaches and supports transfer across equipment.

Core claim

The paper claims that feature alignment for digital twin vibration signals can be solved by formulating it as a continuous-action Markov decision process and optimizing it with Proximal Policy Optimization. The learned policy generates fault-type-specific affine corrections based on the current state of the feature space. A dual-objective reward function balances minimizing the gap between simulated and real distributions while preserving the separability of different fault classes. This is combined with an asymmetry-aware data strategy that uses real data only for the normal class and aligned simulations for faults.

What carries the argument

A continuous-action Markov decision process solved via Proximal Policy Optimization (PPO) that issues fault-type-specific affine corrections to feature spaces.

If this is right

  • The RL policy resolves state-dependent alignment problems that one-shot optimizations cannot.
  • Class-specific corrections close heterogeneous gaps without distorting inter-class boundaries.
  • Asymmetry-aware augmentation improves diagnosis under data scarcity for fault events.
  • Cross-equipment monitoring achieves 92.8% accuracy via linear probing without encoder retraining.
  • Validation on XJTU-SY, CWRU, and slewing bearing testbed confirms the gains from RL-driven alignment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The MDP approach could be adapted for other domain adaptation tasks where gaps vary by class or state.
  • This might enable real-time adaptive alignment as operational conditions change.
  • Extending the dual reward to include more objectives like computational cost could be explored.
  • The transferable capability suggests potential for standardized monitoring systems across different machines.

Load-bearing premise

That solving the alignment via a continuous-action MDP with PPO and dual reward can handle state dependencies better than static methods without the reward distorting class separability.

What would settle it

A direct comparison where a standard domain adaptation method like adversarial training or MMD minimization achieves similar or higher accuracy on the same testbeds using the digital twin data.

Figures

Figures reproduced from arXiv: 2606.24954 by Gaoliang Peng, Jinghan Wang, Tianchen Liu, Wei Zhang, Wentao Wu, Yanjun Chen.

Figure 1
Figure 1. Figure 1: Overview of the proposed three-stage framework. The physical layer establishes a universal monitoring architecture for rotating machinery fault diagnosis. Various rotating equipment could be chosen as monitored objects, with vibration sensors capturing specific dynamic signatures [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simulated time-domain waveforms and frequency spectra for normal, inner race fault, outer race fault and ball fault states. The feature encoder employs a ResNet-10 architecture adapted for one-dimensional dual-channel vibration signals, with input shape (L=2048, C=2). The architecture comprises four residual blocks with progressively increasing channel dimensions (64→128→256→512), followed by global averag… view at source ↗
Figure 3
Figure 3. Figure 3: CA confidence ellipse plots of real and simulated feature distributions before and after RL alignment. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-class KDE distributions of real and simulated features along the first PCA projection axis before and after RL alignment. 4.3 Ablation Study Six model variants (M1–M6) are evaluated at  = 0.45 on the XJTU-SY dataset to quantify the contribution of each framework component. Results are summarized in Table II and illustrated in Figs. 5 and 6. Table II. Ablation Study Results on the XJTU-SY Dataset Model… view at source ↗
Figure 5
Figure 5. Figure 5: Test accuracy of ablation variants M1–M6, with incremental gains annotated between consecutive stages [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-class recall and F1 score of ablation variants M1–M6. Comparing M1 to M3 reveals the effect of pretraining and static alignment as baselines. Physics-simulated pretraining (M1→M2, +1.14%) and static MMD alignment (M2→M3, +1.69%) yield incremental gains, yet the cumulative improvement is only 2.83%, indicating a fundamental ceiling: a single global transform cannot simultaneously close heterogeneous, fa… view at source ↗
Figure 7
Figure 7. Figure 7: Overall test accuracy of the single model and [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-class test accuracy as a function of the fault-class real data ratio 𝝆. 4.5 Cross-Dataset Robustness Evaluation Two cross-dataset evaluations are conducted under a unified linear probing protocol: the XJTU-SY-trained encoder is frozen, and only a lightweight linear classifier is trained and evaluated on each target dataset. The first evaluation target is a self-built slewing bearing testbed, denoted as… view at source ↗
Figure 9
Figure 9. Figure 9: Self-built slewing bearing testbed for cross-dataset robustness validation. (a) Experimental platform comprising industrial site deployment (top left), disassembled bearing components (top right), detailed testbed layout (bottom left), and multi-channel data acquisition system with real-time waveform monitoring (bottom right). (b) Representative fault specimens across three fault types and three severity l… view at source ↗
Figure 10
Figure 10. Figure 10: Normalized confusion matrices of the full model under [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Per-class accuracy of four encoder variants under linear probing on CWRU and Handmade datasets. 5. Conclusion Reliable vibration-based health monitoring of rotating machinery requires diagnostic frameworks effective under real deployment conditions, where fault occurrences are structurally rare and sim-to-real discrepancies are mechanistically heterogeneous across fault classes. Existing global alignment … view at source ↗
read the original abstract

Vibration-based health monitoring of rotating machinery requires reliable fault diagnosis under operational data constraints, yet condition assessment remains challenged by structural scarcity of fault events and heterogeneous sim-to-real gaps in digital twin-generated signals. Each fault type generates impulses with distinct periodicity, amplitude modulation, and spectral character, making feature-space discrepancies fundamentally heterogeneous across fault classes. Existing domain adaptation methods apply a class-agnostic global transformation that cannot close all fault-specific gaps without distorting inter-class separability, while uniform source-target mixing introduces distributional noise into the data-abundant Normal class. These limitations stem from treating a sequential, state-dependent alignment problem as a one-shot optimization. Each corrective transformation simultaneously reshapes all class distributions, creating state dependencies that static gradient descent cannot resolve. We formulate feature alignment as a continuous-action Markov decision process solved via Proximal Policy Optimization, where the learned policy issues fault-type-specific affine corrections responsive to the current feature-space configuration, with a dual-objective reward balancing gap minimization against separability preservation. An asymmetry-aware strategy reserves real data for the Normal class while augmenting fault classes with policy-aligned simulated samples. Validation across XJTU-SY, CWRU, and a self-built slewing bearing testbed confirms the dominant gain from reinforcement learning-driven alignment, and cross-equipment linear probing achieves 92.8% without encoder retraining, demonstrating transferable monitoring capability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper claims that existing class-agnostic domain adaptation methods fail on heterogeneous sim-to-real gaps across fault classes in vibration-based bearing monitoring because each corrective transformation creates state-dependent effects that static gradient descent cannot resolve. It formulates alignment as a continuous-action MDP solved via PPO, with fault-type-specific affine corrections and a dual-objective reward that balances gap minimization against separability preservation, plus an asymmetry-aware strategy that reserves real Normal-class data. Validation on XJTU-SY, CWRU, and a self-built slewing bearing testbed is reported to show dominant gains from the RL approach, including 92.8% accuracy on cross-equipment linear probing without encoder retraining.

Significance. If the MDP formulation is shown to capture irreducible sequential dependencies that static methods cannot and the reward demonstrably preserves class boundaries, the work would provide a new tool for digital-twin-driven PHM under severe fault-data scarcity. The reported cross-equipment transfer result, if robustly controlled, would be of practical value for deployable monitoring systems.

major comments (1)
  1. [Abstract] Abstract: the central premise that 'state dependencies that static gradient descent cannot resolve' necessitate a continuous-action MDP via PPO is presented without any equations or analysis. No definition appears for the state representation, the action parameterization (fault-type-specific affine corrections), the transition dynamics, or the precise dual-objective reward (how gap minimization and separability are quantified and weighted). Without these, it cannot be verified whether the learned policy differs meaningfully from class-conditional static optimization or whether separability is preserved, undermining the claim that RL provides the dominant gain.
minor comments (1)
  1. [Abstract] The abstract introduces the 'asymmetry-aware strategy' and 'policy-aligned simulated samples' without indicating how the policy is applied at inference time or how the Normal class is exactly protected during training.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central premise that 'state dependencies that static gradient descent cannot resolve' necessitate a continuous-action MDP via PPO is presented without any equations or analysis. No definition appears for the state representation, the action parameterization (fault-type-specific affine corrections), the transition dynamics, or the precise dual-objective reward (how gap minimization and separability are quantified and weighted). Without these, it cannot be verified whether the learned policy differs meaningfully from class-conditional static optimization or whether separability is preserved, undermining the claim that RL provides the dominant gain.

    Authors: The abstract is written at a summary level per standard conventions for brevity and accessibility. The manuscript body supplies the requested definitions and analysis: state is the vector of per-class feature moments (Section 3.2); actions are class-specific affine parameters (a_c, b_c) applied to simulated features (Equation 4); transitions are the deterministic updates to the joint feature distribution after each action (Section 3.3); the reward is the weighted sum r = −MMD(s,t) + λ·FDR, where MMD quantifies gap and FDR is the Fisher discriminant ratio preserving separability (Equation 7). PPO is used precisely because each action alters the state for subsequent classes, creating dependencies absent from one-shot class-conditional optimization. We will revise the abstract to reference these components textually for improved clarity. revision: yes

Circularity Check

0 steps flagged

No circularity: formulation and validation remain independent of fitted inputs

full rationale

The paper introduces an MDP formulation solved via PPO for state-dependent feature alignment, with a dual-objective reward, but the abstract and description contain no equations or steps that reduce the claimed performance (e.g., 92.8% cross-equipment accuracy or dominant RL gain) to quantities fitted inside the same experiment or to self-citations. Validation draws on external public datasets (XJTU-SY, CWRU) plus a separate self-built testbed, providing independent benchmarks. No self-definitional, fitted-input-called-prediction, or load-bearing self-citation patterns appear; the central premise is presented as a modeling choice rather than a tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies insufficient detail to enumerate specific free parameters, axioms, or invented entities; the approach implicitly assumes heterogeneous per-class gaps and that a learned policy can issue responsive affine corrections without further specification.

pith-pipeline@v0.9.1-grok · 5798 in / 1126 out tokens · 18670 ms · 2026-06-26T00:52:31.085828+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 24 canonical work pages

  1. [1]

    A Survey on Fault Diagnosis of Rolling Bearings,

    B. Peng, Y. Bi, B. Xue, M. Zhang, and S. Wan, “A Survey on Fault Diagnosis of Rolling Bearings,” Algorithms, vol. 15, no. 10, p. 347, Oct. 2022, doi: 10.3390/a15100347

    work page doi:10.3390/a15100347 2022

  2. [2]

    Digital Twins-based prognostic and health management processes for rotating machinery: a review,

    J. Wang, G. Peng, W. Zhang, W. Wu, S. Li, and Z. Chen, “Digital Twins-based prognostic and health management processes for rotating machinery: a review,” Structural Health Monitoring, p. 14759217251368750, Sep. 2025, doi: 10.1177/14759217251368750

    work page doi:10.1177/14759217251368750 2025

  3. [3]

    Research on Digital Twin Modeling and Fault Diagnosis Methods for Rolling Bearings,

    J. Fan, L. Zhao, and M. Li, “Research on Digital Twin Modeling and Fault Diagnosis Methods for Rolling Bearings,” Sensors, vol. 25, no. 7, p. 2023, Jan. 2025, doi: 10.3390/s25072023

    work page doi:10.3390/s25072023 2023

  4. [4]

    Adaptive Intermediate Class -Wise Distribution Alignment: A Universal Domain Adaptation and Generalization Method for Machine Fault Diagnosis,

    Q. Qian, J. Luo, and Y. Qin, “Adaptive Intermediate Class -Wise Distribution Alignment: A Universal Domain Adaptation and Generalization Method for Machine Fault Diagnosis,” IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 3, pp. 4296–4310, Mar. 2025, doi: 10.1109/TNNLS.2024.3376449

    work page doi:10.1109/tnnls.2024.3376449 2025

  5. [5]

    Deep Transfer Network with Multi -Space Dynamic Distribution Adaptation for Bearing Fault Diagnosis,

    X. Zheng, Z. G u, C. Liu, J. Jiang, Z. He, and M. Gao, “Deep Transfer Network with Multi -Space Dynamic Distribution Adaptation for Bearing Fault Diagnosis,” Entropy, vol. 24, no. 8, p. 1122, Aug. 2022, doi: 10.3390/e24081122

    work page doi:10.3390/e24081122 2022

  6. [6]

    Simulation-to-real transfer learning for bearing fault diagnosis across working conditions: A hybrid approach combining physical modeling and data -driven techniques,

    Z. Han, W. Xia, W. Shen, Q. Zhu, H. Liu, and C. Zhang, “Simulation-to-real transfer learning for bearing fault diagnosis across working conditions: A hybrid approach combining physical modeling and data -driven techniques,” Advanced Engineering Informatics, vol. 69, p. 103998, Jan. 2026, doi: 10.1016/j.aei.2025.103998

    work page doi:10.1016/j.aei.2025.103998 2026

  7. [7]

    Joint distribution adaptation network with adversarial learning for rolling bearing fault diagnosis,

    K. Zhao, H. Jiang, K. Wang, and Z. Pei, “Joint distribution adaptation network with adversarial learning for rolling bearing fault diagnosis,” Knowledge-Based Systems , vol. 222, p. 106974, Jun. 2021, doi: 10.1016/j.knosys.2021.106974

    work page doi:10.1016/j.knosys.2021.106974 2021

  8. [8]

    A Hybrid Prognostics Approach for Estimating Remaining Useful Life of Rolling Element Bearings,

    B. Wang, Y. Lei, N. Li, and N. Li, “A Hybrid Prognostics Approach for Estimating Remaining Useful Life of Rolling Element Bearings,” IEEE Transactions on Reliability , vol. 69, no. 1, pp. 401 –412, Mar. 2020, doi: 10.1109/TR.2018.2882682

    work page doi:10.1109/tr.2018.2882682 2020

  9. [9]

    Bearing Data Center | Case School of Engineering

    “Bearing Data Center | Case School of Engineering.” Accessed: Apr. 04, 2026. [Online]. Available: https://engineering.case.edu/bearingdatacenter

    2026

  10. [10]

    Nonlinear dynamic modeling and vibration analysis for early fault evolution of rolling bearings,

    L. Zheng, Y. Xiang, and N. Luo, “Nonlinear dynamic modeling and vibration analysis for early fault evolution of rolling bearings,” Sci Rep, vol. 14, no. 1, p. 23687, Oct. 2024, doi: 10.1038/s41598-024-75126-5

    work page doi:10.1038/s41598-024-75126-5 2024

  11. [11]

    A multi -layer transcranial focused ultrasound model for neuromodulation procedure planning and insertion loss estimation

    Y. Zhang, X. Zhou, C. Gao, J. Lin, Z. Ren, and K. Feng, “Contrastive learning-enabled digital twin framework for fault diagnosis of rolling bearing,” Meas. Sci. Technol. , vol. 36, no. 1, p. 015026, Nov. 2024, doi: 10.1088/1361 - 6501/ad8f52

    work page doi:10.1088/1361 2024

  12. [12]

    Research on deep learning rolling bearing fault diagnosis driven by high-fidelity digital twins,

    J. Wu, Q. Shu, M. Li, G. Wang, and Y. Wei, “Research on deep learning rolling bearing fault diagnosis driven by high-fidelity digital twins,” Int J Interact Des Manuf, vol. 19, no. 2, pp. 1439–1450, Feb. 2025, doi: 10.1007/s12008- 024-01859-2

    work page doi:10.1007/s12008- 2025

  13. [13]

    A weighted DJP-MMD based deep transfer metric learning for the fault diagnosis of bearing under variable working conditions,

    Z. Xu, G. Ding, Y. Nie, X. Sun, and Z. Wang, “A weighted DJP-MMD based deep transfer metric learning for the fault diagnosis of bearing under variable working conditions,” Front. Mech. Eng., vol. 20, no. 2, p. 16, Apr. 2025, doi: 10.1007/s11465-025-0836-4

    work page doi:10.1007/s11465-025-0836-4 2025

  14. [14]

    A rolling bearing fault diagnosis framework based on multi - modal feature fusion and marginal-conditional alignment,

    Y. Cui, Z. Dong, W. Gao, C. Chang, and J. Wang, “A rolling bearing fault diagnosis framework based on multi - modal feature fusion and marginal-conditional alignment,” Meas. Sci. Technol., vol. 37, no. 3, p. 036112, Jan. 2026, doi: 10.1088/1361-6501/ae3198

    work page doi:10.1088/1361-6501/ae3198 2026

  15. [15]

    Data -Driven Incremental Model Predictive Control for Robot Manipulators,

    Y. Wang, Y. Zhou, F. Liu, M. Leibold, and M. Buss, “Data -Driven Incremental Model Predictive Control for Robot Manipulators,” IEEE/ASME Transactions on Mechatro nics, vol. 30, no. 6, pp. 4353 –4363, Dec. 2025, doi: 10.1109/TMECH.2024.3510729

    work page doi:10.1109/tmech.2024.3510729 2025

  16. [16]

    Domain-Adversarial Training of Neural Networks,

    Y. Ganin et al., “Domain-Adversarial Training of Neural Networks,” Journal of Machine Learning Research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  17. [17]

    Enhancing unsupervised bearing fault diagnosis through structured prediction in latent subspace,

    C. Liu, R. Hu, X. Fang, W. Luo, and C. Zhu, “Enhancing unsupervised bearing fault diagnosis through structured prediction in latent subspace,” Sci Rep, vol. 15, no. 1, p. 42146, Nov. 2025, doi: 10.1038/s41598-025-26013-0

    work page doi:10.1038/s41598-025-26013-0 2025

  18. [18]

    Speed -invariant prototypical network for rolling bearing fault diagnosis under variable speed conditions,

    J. Xing, X. Sun, Y. Song, Y. Li, and D. Wang, “Speed -invariant prototypical network for rolling bearing fault diagnosis under variable speed conditions,” Expert Systems with Applications , vol. 319, p. 132111, Jul. 2026, doi: 10.1016/j.eswa.2026.132111

    work page doi:10.1016/j.eswa.2026.132111 2026

  19. [19]

    Deep Transfer Learning for Bearing Fault Diagnosis: A Systematic Review Since 2016,

    X. Chen, R. Yang, Y. Xue, M. Huang, R. Ferrero, and Z. Wang, “Deep Transfer Learning for Bearing Fault Diagnosis: A Systematic Review Since 2016,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–21, 2023, doi: 10.1109/TIM.2023.3244237

    work page doi:10.1109/tim.2023.3244237 2016

  20. [20]

    A Novel Sensor Scheduling Algorithm Based on Deep Reinforcement Learning for Bearing-Only Target Tracking in UWSNs,

    L. Zheng, M. Liu, S. Zhang, and J. Lan, “A Novel Sensor Scheduling Algorithm Based on Deep Reinforcement Learning for Bearing-Only Target Tracking in UWSNs,” IEEE/CAA Journal of Automatica Sinica, vol. 10, no. 4, pp. 1077–1079, Apr. 2023, doi: 10.1109/JAS.2023.123159

    work page doi:10.1109/jas.2023.123159 2023

  21. [21]

    Structure-Enhanced DRL for Optimal Transmission Scheduling,

    J. Chen, W. Liu, D. E. Quevedo, S. R. Khosravirad, Y. Li, and B. Vucetic, “Structure-Enhanced DRL for Optimal Transmission Scheduling,” IEEE Transactions on Wireless Communications, vol. 23, no. 1, pp. 379 –393, Jan. 2024, doi: 10.1109/TWC.2023.3277861

    work page doi:10.1109/twc.2023.3277861 2024

  22. [22]

    Adaptive reinforcement learning for task scheduling in aircraft maintenance,

    C. Silva, P. Andrade, B. Ribeiro, and B. F. Santos, “Adaptive reinforcement learning for task scheduling in aircraft maintenance,” Sci Rep, vol. 13, no. 1, p. 16605, Oct. 2023, doi: 10.1038/s41598-023-41169-3

    work page doi:10.1038/s41598-023-41169-3 2023

  23. [23]

    Data -Informed Residual Reinforcement Learning for High -Dimensional Robotic Tracking Control,

    C. Li, F. Liu, Y. Wang, and M. Buss, “Data -Informed Residual Reinforcement Learning for High -Dimensional Robotic Tracking Control,” IEEE/ASME Transactions on Mechatronics, vol. 30, no. 3, pp. 1681–1691, Jun. 2025, doi: 10.1109/TMECH.2024.3412275

    work page doi:10.1109/tmech.2024.3412275 2025

  24. [24]

    Lightweight CNN architecture design for rolling bearing fault diagnosis,

    L. Jiang, C. Shi, H. Sheng, X. Li, and T. Yang, “Lightweight CNN architecture design for rolling bearing fault diagnosis,” Meas. Sci. Technol., vol. 35, no. 12, p. 126142, Sep. 2024, doi: 10.1088/1361-6501/ad7a1a

    work page doi:10.1088/1361-6501/ad7a1a 2024

  25. [25]

    A deep reinforcement learning-based intelligent fault diagnosis framework for rolling bearings under imbalanced datasets,

    Y. Li, Y. Wang, X. Zhao, and Z. Chen, “A deep reinforcement learning-based intelligent fault diagnosis framework for rolling bearings under imbalanced datasets,” Control Engineering Practice, vol. 145, p. 105845, Apr. 2024, doi: 10.1016/j.conengprac.2024.105845

    work page doi:10.1016/j.conengprac.2024.105845 2024

  26. [26]

    Multi-Agent Reinforcement Learning Control of a Hydrostatic Wind Turbine - Based Farm,

    Y. Huang, S. Lin, and X. Zhao , “Multi-Agent Reinforcement Learning Control of a Hydrostatic Wind Turbine - Based Farm,” IEEE Transactions on Sustainable Energy , vol. 14, no. 4, pp. 2406 –2416, Oct. 2023, doi: 10.1109/TSTE.2023.3270761

    work page doi:10.1109/tste.2023.3270761 2023

  27. [27]

    Proximal Policy Optimization Algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford , and O. Klimov, “Proximal Policy Optimization Algorithms,” arXiv.org. Accessed: Apr. 15, 2026. [Online]. Available: https://arxiv.org/abs/1707.06347v2

    Pith/arXiv arXiv 2026