Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Scenario-Based Hierarchical Reinforcement Learning for Automated Driving Decision Making

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A hierarchical reinforcement-learning agent trained on mixed synthetic and real-road scenarios reaches its goal in over 69 percent of episodes across every scenario type tested, in simulation.

desk verdict Solid system paper with a confounded ablation story; the main result is plausible, but the 'essential' claims don't follow from the experiments. read the letter →

arxiv 2506.23023 v1 pith:3ZMDKDWQ submitted 2025-06-28 cs.RO cs.AIcs.LG

M. Youssef Abdelhamid , Lennart Vater , Zlatan Ajanovic This is my paper · ORCID
classification cs.ROcs.AIcs.LG
keywords AutomatedDrivingDecision-MakingHierarchicalReinforcementLearningScenario-BasedTrainingSafetyGeneralizationAutonomousHighwaySimulation-BasedA2C
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 proposes SAD-RL, a framework for automated-driving decision making that trains a hierarchical reinforcement-learning policy inside a scenario-based simulator. The high-level policy chooses discrete driving maneuvers, the low-level controller executes them, and a shield blocks unsafe actions. The authors claim that an agent trained on a mixture of synthetic critical scenarios and real highway trajectories reaches its goal in over 69 percent of episodes on every test type, including scenarios it never saw during training. Ablation experiments are used to show that both hierarchical control and scenario diversity are needed, because removing either leaves the agent unable to reach the goal or unable to transfer. If correct, this would mean that carefully chosen training scenarios can replace massive random-driving datasets for learning a generalizable, safe highway policy.

What carries the argument

The load-bearing mechanism is the hierarchy itself: a high-level policy selects maneuver templates, a low-level SAD-pilot executes feasible trajectories, and a shield vetoes dangerous actions during training. The top-level action space is discrete, combining lateral intents with longitudinal intents, which makes credit assignment tractable and lets the lateral shield enforce safety. Around this, the scenario-based environment controls the training distribution: synthetic scenarios are verified for feasibility and filtered to allow at least 6.5 seconds for a lane change, while real-road scenarios are converted from a naturalistic highway trajectory dataset. This combination is what allows one agent to transfer across synthetic and real highway conditions.

What would settle it

Run the same SAD-RL training with reactive traffic models where other vehicles respond to the ego's actions, or with a scenario that forces a decision in less than 6.5 seconds; if goal-reaching rates collapse or collision and offroad rates rise sharply, the transfer claim is refuted. A weaker but concrete check is to measure safety-margin distributions on the existing test scenarios.

Watch

Extended reading notes

Core claim

The central claim is that decision-making for highway driving can be learned efficiently and safely by combining a two-level policy with a curated scenario curriculum. In the SAD-RL framework, a high-level A2C agent selects a discrete maneuver template, the SAD-pilot converts that choice into a feasible trajectory, and a shield vetoes high-risk actions. The training set mixes synthetically generated critical situations with naturalistic highway scenarios extracted from a drone-recorded trajectory dataset. Across the four test scenario types, the hybrid agent reaches goal-reaching rates of 82.0, 94.5, 69.5, and 75.0 percent for the real-road, Type A, Type B, and Cutout test sets respectively, so its goal rate exceeds 69 percent on every test type. The paper further claims that ablations confirm both components are necessary: without HRL the agent never reaches the goal and ends offroad, and without synthetic critical scenarios the real-road-only agent degrades sharply on unseen types.

Load-bearing premise

The framework's headline results hold only if the open-loop simulation, where other vehicles follow fixed trajectories, together with the 6.5-second scenario filter, faithfully represents the decision situations that matter for real automated driving.

Editorial extensions

If this is right

  • A single policy can transfer across synthetic hazard scenarios and naturalistic highway data without retraining.
  • Training on a few thousand curated synthetic scenarios plus a few hundred real-road scenarios yields cross-domain goal rates above 69 percent, in contrast to single-type agents that overfit their training domain.
  • Removing the hierarchical layer destroys learning, with the no-HRL agent reaching the goal zero times and ending offroad near 100 percent of the time.
  • The 6.5-second scenario filter defines the smallest reaction window the framework can handle, so scenarios requiring faster decisions are outside the method's current coverage.

Reading between the lines

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

  • I would expect this recipe to carry over to urban driving only after replacing the open-loop traffic model, since reactive driver responses are precisely what the current non-ego trajectory model omits.
  • A natural extension is to measure safety not only by termination reason but also by collision severity or time-to-collision margins, which would test whether the shield and policy leave a comfortable safety buffer.
  • The reported performance is bounded by the test suite and the 6.5-second filter; lowering the filter or adding perception noise would likely reveal the shortest scenario horizon and sensor error the policy can tolerate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SAD-RL, a hierarchical reinforcement learning framework for automated driving decision-making. A high-level A2C policy selects discrete maneuver options (lane changes, longitudinal actions), while a low-level SAD-Pilot controller executes them; a shield blocks unsafe actions. Training uses a CommonRoad-based simulator with synthetically generated critical scenarios (types A, B, Cutout) and real-road scenarios extracted from the highD dataset. The authors evaluate goal-reaching rates across four test scenario sets and report that the hybrid agent trained on both synthetic and real-road data, πwes+highD, achieves G(πwes+highD, t) > 69% for all test sets, outperforming single-source agents. Ablation studies are presented to support the conclusions that hierarchical RL and scenario diversity are both essential. The central empirical demonstration is a simulation study with a fixed, open-loop traffic model.

Significance. If the reported results hold, the paper offers a useful empirical demonstration that combining hierarchical RL with scenario-based training on mixed synthetic and naturalistic data can yield a single policy that performs reasonably across different scenario types in simulation. The evaluation is systematic in several respects: it uses distinct training/testing splits, multiple termination reasons (goal, collision, offroad, timeout, standstill), four RL algorithms, and cross-scenario testing. The work also has a practical angle in that it builds on the CommonRoad toolchain and the highD dataset, which supports reproducibility. However, the significance is limited by the absence of uncertainty quantification for the headline numbers and by the confounded ablations that underpin the two main design claims. The paper is a plausible engineering contribution rather than a decisive conceptual advance.

major comments (4)
  1. [V-B.1] The HRL ablation is confounded and does not establish that HRL is essential, as claimed in the abstract and in Sec. V-C. Removing the high-level policy also removes the SAD-Pilot low-level controller, disables the lateral shield, and changes the action space from discrete maneuver options to raw continuous steering and acceleration. The observed failure (offroad rate approaching 100%, zero goal-reaching) could be caused by A2C's difficulty with raw continuous control or by the absence of the shield, rather than by the absence of hierarchy. To support the claim, the ablation should keep the action interface and safety layer fixed and vary only the hierarchical structure, or add a flat continuous baseline with the shield enabled.
  2. [V-B.2] The scenario-diversity ablation confounds scenario source with training-data volume and distribution. The highD-only policy πhighD is trained on 300 real-road scenarios, whereas the synthetic-only policies are trained on 1000 to 3000 scenarios, and πwes+highD is trained on 3300 scenarios. The observed generalization differences could reflect dataset size or the absence of easy scenarios in the synthetic-only sets, rather than the intrinsic value of mixing real-road and synthetic data. A fair comparison would control for the number of training scenarios and the presence of easy versus challenging cases across conditions.
  3. [IV-B3 / Table II] The headline claim that πwes+highD is the most robust policy, with G(πwes+highD, t) > 69% for all t, rests on single point estimates with no variance, confidence intervals, or statistical tests, despite the earlier full-scale training using 10 seeds (Sec. V-A). Several differences in Table II are small (e.g., πB versus πfull on test set B: 69.5% versus 65.5%; πA versus πcut on test set B: 57.5% versus 56.0%), so without repeated-seed evaluation it is unclear whether these differences are meaningful. The authors should report mean and spread over seeds for the policies in Table II, or at minimum for the compared ablations.
  4. [III-A / III-E / Abstract] The scope of the safety and generalizability claims should be stated more carefully. The simulator is open-loop (non-ego vehicles follow predefined trajectories) and scenarios requiring lane changes with less than 6.5 seconds of decision time are filtered out. Consequently, the abstract's claim that the agent achieves 'safe behaviour' and Sec. V-C's claim of generalization are demonstrated only for this simulation setup, which does not include reactive traffic, perception uncertainty, or hazards requiring shorter response horizons. These are significant limitations for automated driving and should be explicitly acknowledged as such in the main text, not only implicitly in the methodology.
minor comments (5)
  1. [V-B.1] The ablation text is internally inconsistent: the first paragraph states 'Only A2C is evaluated', while the later paragraph says 'The algorithms tested in this case are PPO and A2C'. Please clarify which algorithms were actually run in the non-HRL ablation.
  2. [V-B.1 / Fig. 10] The caption of Fig. 10 says the agent was trained 'on one episode', but the text says training lasted 1 million timesteps. Please correct the caption to be consistent with the intended meaning.
  3. [Abstract] The abstract contains the typo 'challenging, but rate situations'; this should read 'rare situations' or similar.
  4. [IV-B2 / Table I] The overfitting test results in Table I are reported as single percentages with no indication of variance across seeds or evaluation episodes; adding this information would strengthen the comparison between the four algorithms.
  5. [III-E] The 6.5-second filtering threshold is central to the framework's scenario selection, but it is justified only qualitatively ('considers the time required by the SAD-pilot'). Please provide a concrete derivation or a sensitivity analysis around this value.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported goal-reaching results are independent simulation measurements, and the confounding in the ablations is an experimental-design concern rather than a definitional or fitted-input reduction.

full rationale

The paper's central quantitative claim is the empirical goal-reaching table (Table II), specifically G(piwes+highD, t) > 69% for all test scenario families. These numbers are measured on held-out scenario instances from CommonRoad/highD, and no reported quantity is defined in terms of the trained parameters in a way that makes the outcome true by construction. The hybrid policy is honestly described as 'A2C trained on the full WES set (3000 scenarios) + 300 highD scenarios', and the table reports its measured performance on separate test scenarios. The ablation studies in Sec. V-B are indeed not well-controlled: the 'non-HRL' ablation in Sec. V-B.1 simultaneously removes the SAD-Pilot, disables lateral shielding, and changes the action space from discrete maneuver options to raw continuous steering and acceleration, so the claim that HRL is critical is not uniquely established by that comparison. Similarly, Sec. V-B.2 compares pi_highD, trained on 300 real-road scenarios, against policies trained on 1000-3000 synthetic scenarios, confounding scenario source with dataset size and composition. However, these are threats to the validity of the paper's ablation-based conclusions, not circularity: the conclusions are not equivalent to the experimental inputs by definition, and no fitted parameter is renamed as a prediction. The self-citations present (e.g., [9] in related work) are contextual and do not carry the load of the central claim. Therefore no circular step can be exhibited with the required quotation-and-reduction standard, and the appropriate score is 0.

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

The central claims rest on modeling assumptions about simulation fidelity, scenario representativeness, and metric validity. The paper introduces no new physical or algorithmic entities beyond the SAD-RL combination itself, but the simulation proxy and the scenario filter are load-bearing assumptions the reader must accept.

free parameters (4)
  • Scenario filtering threshold = 6.5 seconds
    Section III-E: scenarios with lane-change decision time under 6.5 seconds are removed. This hand-chosen threshold shapes the training distribution and restricts the claim to situations with sufficient reaction time.
  • Reward values for sparse reward = not reported
    Section III-F1 states a sparse reward that rewards goal-reaching and penalizes other terminations, but the actual reward magnitudes are not given; these values influence all learning results.
  • A2C hyperparameters = not reported
    Section III-F2 says A2C was implemented with Stable-Baselines3, but learning rate, entropy coefficient, network sizes, and other hyperparameters are not listed; results could depend on them.
  • Shield risk thresholds = not reported
    Section III-D describes a shield that blocks high-risk actions, but the exact collision and offroad risk criteria are not specified; this safety layer is part of the central framework.
assumptions (4)
  • domain assumption Open-loop simulation with scripted non-ego trajectories is a sufficient proxy for real-world driving decision-making.
    Section III-A states non-ego vehicles follow predefined trajectories; the safety, generalizability, and sample-efficiency conclusions are drawn from this closed-loop but non-reactive simulation.
  • domain assumption The three synthetic scenario types (A, B, Cutout), inspired by UN R157, represent the critical situations needed for safety.
    Section III-B1 assumes that deceleration, cut-in, and cutout scenarios cover the safety-relevant behavior space, so training on them transfers to unseen driving.
  • ad hoc to paper Goal-reaching rate is the primary proxy for safe driving.
    Section IV-A uses the percentage of goal-reaching terminations as the main metric; episodes may also end by timeout or offroad, and the safe behavior claim relies on this metric.
  • domain assumption The SAD-Pilot low-level controller reliably executes any maneuver template the high-level policy selects within the filtered scenarios.
    Section III-C says the high-level agent selects maneuver templates executed by the SAD-pilot; the framework's correctness depends on the pilot's feasibility, which is not independently evaluated in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scenario-Based Hierarchical Reinforcement Learning for Automated Driving Decision Making." pith.science (2026). https://pith.science/paper/3ZMDKDWQ

@misc{pith2026250623023,
  author       = {Pith},
  title        = {Pith review of: Scenario-Based Hierarchical Reinforcement Learning for Automated Driving Decision Making},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZMDKDWQ}},
  note         = {Machine review of arXiv:2506.23023}
}
read the original abstract

Developing decision-making algorithms for highly automated driving systems remains challenging, since these systems have to operate safely in an open and complex environments. Reinforcement Learning (RL) approaches can learn comprehensive decision policies directly from experience and already show promising results in simple driving tasks. However, current approaches fail to achieve generalizability for more complex driving tasks and lack learning efficiency. Therefore, we present Scenario-based Automated Driving Reinforcement Learning (SAD-RL), the first framework that integrates Reinforcement Learning (RL) of hierarchical policy in a scenario-based environment. A high-level policy selects maneuver templates that are evaluated and executed by a low-level control logic. The scenario-based environment allows to control the training experience for the agent and to explicitly introduce challenging, but rate situations into the training process. Our experiments show that an agent trained using the SAD-RL framework can achieve safe behaviour in easy as well as challenging situations efficiently. Our ablation studies confirmed that both HRL and scenario diversity are essential for achieving these results.

Figures

Figures reproduced from arXiv: 2506.23023 by the authors.

Figure 1
Figure 1. SAD-RL system architecture showing the hierarchical [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Scenario Type A Concept—the ego vehicle is placed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Scenario Type B Concept—the challenger vehicle is [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Scenario Type Cutout Concept—two vehicles are [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Training performance of the A2C agent on a single [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 9
Figure 9. Figure 9: Moving average rate of training DQN agent for 1M [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 6
Figure 6. Figure 6: Moving average rate of training A2C agent for 1M [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Moving average rate of training PPO agent for 1M [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 10
Figure 10. Figure 10: Training performance of A2C on one episode without [PITH_FULL_IMAGE:figures/full_fig_p005_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

  1. [1]

    Improving passenger experience and trust in automated vehicles through user- adaptive hmis: “the more the better

    F. Hartwich, C. Hollander, D. Johannmeyer, and J. F. Krems, “Improving passenger experience and trust in automated vehicles through user- adaptive hmis: “the more the better” does not apply to everyone,” Frontiers in Human Dynamics , vol. 3, 2021. [Online]. Available: https://www.frontiersin.org/articles/10.3389/fhumd.2021.669030

    work page arXiv 2021

  2. [2]

    Scenario- and model-based systems engineering for highly automated driving,

    M. Kremer, S. Christiaens, C. Granrath et al. , “Scenario- and model-based systems engineering for highly automated driving,” ATZ Worldw , vol. 122, pp. 16–21, 2020. [Online]. Available: https://doi.org/10.1007/s38311-020-0330-x

    work page doi:10.1007/s38311-020-0330-x 2020

  3. [3]

    Mastering the game of go with deep neural networks and tree search,

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V . Panneershelvam, M. Lanctot et al., “Mastering the game of go with deep neural networks and tree search,” nature, vol. 529, no. 7587, pp. 484–489, 2016

    2016

  4. [4]

    Survey of deep reinforcement learning for motion planning of autonomous vehicles,

    S. Aradi, “Survey of deep reinforcement learning for motion planning of autonomous vehicles,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 2, pp. 740–759, 2022

    work page 2022

  5. [5]

    Combining deep reinforcement learning with rule-based constraints for safe highway driving,

    T. Liu, Q. Liu, H. Liu, and X. Ren, “Combining deep reinforcement learning with rule-based constraints for safe highway driving,” in 2022 China Automation Congress (CAC) . IEEE, 2022, pp. 2785–2790

    work page 2022

  6. [6]

    Efficient reinforcement learning for autonomous driving with parameterized skills and priors,

    L. Wang, J. Liu, H. Shao, W. Wang, R. Chen, Y . Liu, and S. L. Waslander, “Efficient reinforcement learning for autonomous driving with parameterized skills and priors,” arXiv preprint arXiv:2305.04412, 2023

    arXiv 2023

  7. [7]

    Imitation is not enough: Ro- bustifying imitation with reinforcement learning for challenging driving scenarios,

    Y . Lu, J. Fu, G. Tucker, X. Pan, E. Bronstein, R. Roelofs, B. Sapp, B. White, A. Faust, S. Whiteson et al. , “Imitation is not enough: Ro- bustifying imitation with reinforcement learning for challenging driving scenarios,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 7553–7560

    work page 2023

  8. [8]

    Dynamic trajectory planning with dynamic constraints: A’state-time space’approach,

    T. Fraichard, “Dynamic trajectory planning with dynamic constraints: A’state-time space’approach,” in Proceedings of 1993 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS’93), vol. 2. IEEE, 1993, pp. 1393–1400

    work page 1993

Show all 22 references
  1. [9]

    Search-based optimal motion planning for automated driving,

    Z. Ajanovic, B. Lacevic, B. Shyrokau, M. Stolz, and M. Horn, “Search-based optimal motion planning for automated driving,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 4523–4530

    2018

  2. [10]

    A minimax-based decision-making approach for safe maneuver planning in automated driving,

    M. Saraoglu, H. Jiang, M. Schirmer, ˙I. Mutlu, and K. Janschek, “A minimax-based decision-making approach for safe maneuver planning in automated driving,” in 2023 American Control Conference (ACC) . IEEE, 2023, pp. 4683–4690

    2023

  3. [11]

    Alvinn: An autonomous land vehicle in a neural network,

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

    1988

  4. [12]

    Symphony: Learning realistic and diverse agents for autonomous driving simulation,

    M. Igl, D. Kim, A. Kuefler, P. Mougin, P. Shah, K. Shiarlis, D. Anguelov, M. Palatucci, B. White, and S. Whiteson, “Symphony: Learning realistic and diverse agents for autonomous driving simulation,” in 2022 Inter- national Conference on Robotics and Automation (ICRA) . IEEE, ...

    2022

  5. [13]

    Safe reinforcement learning via shielding,

    M. Alshiekh, R. Bloem, R. Ehlers, B. K ¨onighofer, S. Niekum, and U. Topcu, “Safe reinforcement learning via shielding,” in Proceedings of the AAAI conference on artificial intelligence , vol. 32, no. 1, 2018

    2018

  6. [14]

    Safetynet: Safe planning for real-world self-driving vehicles using machine-learned policies,

    M. Vitelli, Y . Chang, Y . Ye, A. Ferreira, M. Wołczyk, B. Osi ´nski, M. Niendorf, H. Grimmett, Q. Huang, A. Jain et al. , “Safetynet: Safe planning for real-world self-driving vehicles using machine-learned policies,” in 2022 International Conference on Robotics and Automatio...

    2022

  7. [15]

    Commonroad: Composable benchmarks for motion planning on roads,

    M. Althoff, M. Koschi, and S. Manzinger, “Commonroad: Composable benchmarks for motion planning on roads,” in Proc. of the IEEE Intelligent Vehicles Symposium, 2017, pp. 719–726

    2017

  8. [16]

    Un regulation no. 157 - automated lane keeping systems (alks),

    United Nations Economic Commission for Europe, “Un regulation no. 157 - automated lane keeping systems (alks),” https://unece.org/sites/default/files/2023-12/R157e.pdf, 2023, accessed: 2024-07-09

    2023

  9. [17]

    A framework for definition of logical scenarios for safety assurance of automated driving,

    H. Weber, J. Bock, J. Klimke, C. Roesener, J. Hiller, R. Krajewski, A. Zlocki, and L. Eckstein, “A framework for definition of logical scenarios for safety assurance of automated driving,” Traffic injury prevention, vol. 20, no. sup1, pp. S65–S70, 2019

    2019

  10. [18]

    Commonroad drivability checker: Simplifying the development and validation of motion planning algorithms,

    C. Pek, V . Rusinov, S. Manzinger, M. C. ¨Uste, and M. Althoff, “Commonroad drivability checker: Simplifying the development and validation of motion planning algorithms,” in 2020 IEEE intelligent vehicles symposium (IV) . IEEE, 2020, pp. 1013–1020

    2020

  11. [19]

    The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,

    R. Krajewski, J. Bock, L. Kloeker, and L. Eckstein, “The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems,” in 2018 21st International Conference on Intelligent Transportation Systems (ITS...

    2018

  12. [20]

    Automatic traffic scenario conversion from openscenario to commonroad,

    Y . Lin, M. Ratzel, and M. Althoff, “Automatic traffic scenario conversion from openscenario to commonroad,” arXiv preprint arXiv:2305.10080 , 2023

    2023 arXiv

  13. [21]

    Gymnasium: A standard interface for reinforcement learning environments,

    M. Towers, A. Kwiatkowski, J. Terry, J. U. Balis, G. De Cola, T. Deleu, M. Goul ˜ao, A. Kallinteris, M. Krimmel, A. KG et al. , “Gymnasium: A standard interface for reinforcement learning environments,” arXiv preprint arXiv:2407.17032, 2024

    2024 arXiv

  14. [22]

    Stable-baselines3: Reliable reinforcement learning implementa- tions,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dor- mann, “Stable-baselines3: Reliable reinforcement learning implementa- tions,” Journal of machine learning research , vol. 22, no. 268, pp. 1–8, 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.