Pith. sign in

REVIEW 2 major objections 2 minor 43 references

Mesh-RL: Coupled subgrid reinforcement learning

T0 review · 2 major / 2 minor · reviewed 2026-06-26 · grok-4.3

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

Pith's one-line read Mesh-RL partitions environments into overlapping subgrids and enforces boundary-consistent updates to accelerate long-range credit assignment in reinforcement learning.

desk verdict Mesh-RL applies overlapping subgrids and boundary-consistent TD updates to speed value propagation in grid RL, but the abstract leaves the enforcement mechanism too vague to confirm it avoids altering the Bellman operator. read the letter →

arxiv 2606.26333 v1 pith:BERNRBH2 submitted 2026-06-24 cs.LG

Behnam Gheshlaghi , Bahador Rashidi , Shahin Atakishiyev This is my paper · ORCID
classification cs.LG
keywords reinforcementlearningdomaindecompositiontemporaldifferencecreditassignmentgridworldvaluepropagationsubgridstability
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

Reinforcement learning in large or sparse-reward settings suffers from slow spread of value information because temporal-difference updates occur only locally. Mesh-RL divides the state space into overlapping subgrids and requires that updates remain consistent at their shared boundaries. This structure supports localized learning while keeping value estimates coherent across the whole space. A sympathetic reader would care because faster credit assignment could shorten training time and raise final performance without changes to rewards or core update rules. Tests on hazard-dense grid worlds show gains in speed, reward, and stability for several standard algorithms.

What carries the argument

Spatial domain-decomposition into overlapping subgrids that enforces boundary-consistent temporal-difference updates.

What would settle it

An experiment on a large hazard-dense grid where value fails to reach distant states faster than baseline or where boundary mismatches appear in the learned values would falsify the central claim.

Watch

Extended reading notes

Core claim

Mesh-RL accelerates long-range credit assignment by partitioning the environment into overlapping subgrids and enforcing boundary-consistent temporal-difference updates, improving convergence speed, cumulative reward, and learning stability across Q-learning, SARSA, and Dyna-Q on hazard-dense grid-world environments with varying geometries and mesh resolutions, without modifying the reward function, Bellman operator, or adding explicit planning.

Load-bearing premise

Boundary-consistent temporal-difference updates across overlapping subgrids will produce globally coherent value propagation without introducing new inconsistencies or requiring extra tuning that cancels the gains.

Editorial extensions

If this is right

  • Mesh-RL improves convergence speed, cumulative reward, and learning stability across Q-learning, SARSA, and Dyna-Q.
  • Higher mesh resolutions sustain exploration, prevent premature convergence, and accelerate value propagation to distant states.
  • Dyna-Q receives additional gains from the decomposition even though it already uses internal planning.
  • The method works across different environment geometries and mesh resolutions in hazard-dense grid worlds.

Reading between the lines

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

  • The decomposition could scale to larger state spaces by increasing the number of subgrids while keeping overlap fixed.
  • It may reduce reliance on hierarchical architectures for tasks with long credit chains by handling propagation through spatial structure alone.
  • Varying overlap size independently of resolution in follow-up tests could isolate which factor drives the observed speedups.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Mesh-RL, a spatial domain-decomposition framework for reinforcement learning that partitions environments into overlapping subgrids and enforces boundary-consistent temporal-difference updates. This is claimed to accelerate long-range credit assignment in sparse-reward settings without modifying the reward function, Bellman operator, or introducing explicit planning. Evaluations on hazard-dense grid-worlds show consistent gains in convergence speed, cumulative reward, and stability across Q-learning, SARSA, and Dyna-Q, with higher mesh resolutions further improving value propagation.

Significance. If the central claim is substantiated, the work would offer a principled mechanism for improving sample efficiency in RL by adapting finite-element and domain-decomposition ideas from scientific computing. The explicit commitment to release source code is a strength for reproducibility. The approach could be relevant for large or sparse-reward domains where standard TD propagation is slow, provided the boundary-consistency mechanism does not implicitly alter the update rule.

major comments (2)
  1. [Abstract and Method] Abstract and Method section: The claim that overlapping subgrids enforce boundary-consistent TD updates 'without modifying the Bellman operator' is load-bearing for the central contribution. Standard local TD backups (Q-learning, SARSA) do not automatically guarantee global coherence across overlaps; any enforcement step (averaging, projection, or message passing at boundaries) must be shown mathematically to be equivalent to unmodified local backups. The manuscript provides no explicit update equation or pseudocode demonstrating this equivalence, leaving the key assumption unverified.
  2. [Experiments] Experiments section: The abstract reports consistent improvements in convergence speed, cumulative reward, and learning stability but supplies no error bars, number of independent runs, or statistical tests. Without these, it is impossible to determine whether the reported gains are robust or whether they could be offset by any additional tuning required for boundary consistency.
minor comments (2)
  1. [Abstract] The abstract mentions 'varying geometries and mesh resolutions' but does not specify the concrete grid sizes or resolution values used in the reported experiments; adding these details would improve clarity.
  2. The manuscript states it will release source code; confirming the repository link or DOI in the camera-ready version would strengthen reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment below and will revise the paper to improve clarity and rigor where indicated.

read point-by-point responses

  1. Referee: [Abstract and Method] Abstract and Method section: The claim that overlapping subgrids enforce boundary-consistent TD updates 'without modifying the Bellman operator' is load-bearing for the central contribution. Standard local TD backups (Q-learning, SARSA) do not automatically guarantee global coherence across overlaps; any enforcement step (averaging, projection, or message passing at boundaries) must be shown mathematically to be equivalent to unmodified local backups. The manuscript provides no explicit update equation or pseudocode demonstrating this equivalence, leaving the key assumption unverified.

    Authors: We agree that an explicit mathematical formulation is required to substantiate the claim of equivalence. In the revised Method section we will add the precise boundary-consistent update equations together with pseudocode, demonstrating that local TD backups remain unmodified while global coherence is achieved solely through the overlapping subgrid structure. revision: yes

  2. Referee: [Experiments] Experiments section: The abstract reports consistent improvements in convergence speed, cumulative reward, and learning stability but supplies no error bars, number of independent runs, or statistical tests. Without these, it is impossible to determine whether the reported gains are robust or whether they could be offset by any additional tuning required for boundary consistency.

    Authors: We acknowledge the absence of these statistical details in the current version. The revised manuscript will report results from multiple independent runs (with the exact number stated), include error bars, and provide appropriate statistical tests to confirm the robustness of the observed improvements. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected; derivation chain is self-contained

full rationale

The abstract and description present Mesh-RL as a novel spatial domain-decomposition approach inspired by finite element methods, with boundary-consistent TD updates claimed to accelerate credit assignment without altering the reward function or Bellman operator. No equations, fitted parameters, or self-citations appear in the provided text that would reduce any claimed prediction or result to a redefinition of inputs by construction. The central premise relies on the proposed mechanism producing global coherence from local updates, which is presented as an independent contribution rather than a tautology or renamed known result. This qualifies as a normal non-finding of circularity.

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

Abstract provides no explicit free parameters, axioms, or invented entities; mesh resolution is mentioned as varying but its status as a tuned hyperparameter cannot be confirmed without methods section.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mesh-RL: Coupled subgrid reinforcement learning." pith.science (2026). https://pith.science/paper/BERNRBH2

@misc{pith2026260626333,
  author       = {Pith},
  title        = {Pith review of: Mesh-RL: Coupled subgrid reinforcement learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BERNRBH2}},
  note         = {Machine review of arXiv:2606.26333}
}
read the original abstract

Reinforcement learning in large or sparse-reward environments suffers from slow temporal-difference reward propagation, as value information spreads only locally across the state space. We propose Mesh-RL, a spatial domain-decomposition framework inspired by the finite element method and domain decomposition theory, which partitions the environment into overlapping subgrids and enforces boundary-consistent temporal-difference updates. Such an approach enables localized learning while ensuring globally coherent value propagation. Unlike hierarchical or model-based approaches, Mesh-RL accelerates long-range credit assignment without modifying the reward function, Bellman operator, or introducing explicit planning mechanisms. We evaluate Mesh-RL on hazard-dense grid-world environments with varying geometries and mesh resolutions. Across Q-learning, SARSA, and Dyna-Q, Mesh-RL consistently improves convergence speed, cumulative reward, and learning stability. Higher mesh resolutions sustain exploration, prevent premature convergence, and substantially accelerate value propagation to distant states. While Dyna-Q already benefits from internal planning, it still achieves additional gains under structured decomposition. Overall, Mesh-RL introduces a principled spatial domain-decomposition mechanism for accelerating temporal-difference learning. Our framework bridges finite element method-inspired boundary-consistency techniques from scientific computing with reinforcement learning to improve sample efficiency in sparse-reward environments. We will release source code of the study.

Figures

Figures reproduced from arXiv: 2606.26333 by the authors.

Figure 1
Figure 1. A graphical illustration of Mesh-RL. improving TD learning in long-horizon tasks. Successor feature representations [15] factorize value functions into predictive state occupancies, enabling faster generalization of reward signals across related tasks. Option-critic architectures [16] learn temporal abstractions end-to-end, reducing effective planning horizons and improving stability in sparse-reward domains. Explor… view at source ↗
Figure 2
Figure 2. Performance on a 10 × 30 grid with 50 random holes (M = 2 and M = 6) and baseline methods. Since Q∗ satisfies the Bellman optimality equation Q∗ = T Q∗ , it follows that QG = T QG. □ Proposition 2 (Contraction Property). The Mesh-RL operator TM, which consists of local Bellman updates, boundary substitution, and stitching on overlapping states, is a γ-contraction under the max norm: ∥TMQ − TMQ ′ ∥∞ ≤ γ∥Q − Q ′ ∥∞. S… view at source ↗
Figure 3
Figure 3. Value heatmaps on a 10 × 30 grid with 50 holes, for baselines and Mesh-RL (M = 2 and M = 6). The mesh decomposition strategy follows a systematic spatial partitioning approach: for M meshes, the environment is divided into M contiguous subregions, each containing approximately N/M states, where N is the total number of states. This partitioning aligns with principles from parallel computing and domain decomposition,… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance on a 20 × 20 grid with 50 random holes (M = 2 and M = 6) and baseline methods. synchronization ensures coherent propagation of reward information across the grid, effectively reducing the learning horizon to distant states. Cumulative reward plots reinforce…
Figure 5
Figure 5. Figure 5: Value heatmaps on a 20 × 20 grid with 50 holes, for baselines and Mesh-RL (M = 2 and M = 6). 6.3 Summary of Mesh Resolution Effects Across all experimental setups, increasing the mesh resolution from M = 2 to M = 6 consistently improves performance for TD-based methods…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 7 canonical work pages

  1. [1]

    Analysis of Temporal-Difference Learning with Function Approximation.Advances in Neural Information Processing Systems, 9, 1996

    John Tsitsiklis and Benjamin Van Roy. Analysis of Temporal-Difference Learning with Function Approximation.Advances in Neural Information Processing Systems, 9, 1996

    1996

  2. [2]

    MIT Press Cambridge, 1998

    Richard S Sutton and Andrew G Barto.Reinforcement learning: An Introduction, volume 1. MIT Press Cambridge, 1998

    1998

  3. [3]

    Temporal Difference Learning: Why It Can Be Fast and How It Will Be Faster

    Patrick Schnell, Luca Guastoni, and Nils Thuerey. Temporal Difference Learning: Why It Can Be Fast and How It Will Be Faster. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  4. [4]

    Prioritized sweeping: Reinforcement learning with less data and less time.Machine Learning, 13(1):103–130, 1993

    Andrew W Moore and Christopher G Atkeson. Prioritized sweeping: Reinforcement learning with less data and less time.Machine Learning, 13(1):103–130, 1993

    1993

  5. [5]

    Reinforcement Learning with Hierarchies of Machines.Ad- vances in Neural Information Processing Systems, 10, 1997

    Ronald Parr and Stuart Russell. Reinforcement Learning with Hierarchies of Machines.Ad- vances in Neural Information Processing Systems, 10, 1997

    1997

  6. [6]

    Recent Advances in Hierarchical Reinforcement Learning.Discrete Event Dynamic Systems, 13(4):341–379, 2003

    Andrew G Barto and Sridhar Mahadevan. Recent Advances in Hierarchical Reinforcement Learning.Discrete Event Dynamic Systems, 13(4):341–379, 2003

    2003

  7. [7]

    Hierarchical Rein- forcement Learning: A Comprehensive Survey.ACM Computing Surveys (CSUR), 54(5):1–35, 2021

    Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchical Rein- forcement Learning: A Comprehensive Survey.ACM Computing Surveys (CSUR), 54(5):1–35, 2021

    2021

  8. [8]

    Discovering Temporal Structure: An Overview of Hierarchical Reinforcement Learning.arXiv preprint arXiv:2506.14045, 2025

    Martin Klissarov, Akhil Bagaria, Ziyan Luo, George Konidaris, Doina Precup, and Marlos C Machado. Discovering Temporal Structure: An Overview of Hierarchical Reinforcement Learning.arXiv preprint arXiv:2506.14045, 2025. 12

    work page arXiv 2025

Show all 43 references
  1. [9]

    Hierarchical Reinforcement Learning with the MAXQ Value Function Decomposition.Journal of Artificial Intelligence Research, 13:227–303, 2000

    Thomas G Dietterich. Hierarchical Reinforcement Learning with the MAXQ Value Function Decomposition.Journal of Artificial Intelligence Research, 13:227–303, 2000

    2000

  2. [10]

    Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning.Artificial Intelligence, 112(1- 2):181–211, 1999

    Richard S Sutton, Doina Precup, and Satinder Singh. Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning.Artificial Intelligence, 112(1- 2):181–211, 1999

    1999

  3. [11]

    Discovery of options via meta-learned subgoals

    Vivek Veeriah, Tom Zahavy, Matteo Hessel, Zhongwen Xu, Junhyuk Oh, Iurii Kemaev, Hado P van Hasselt, David Silver, and Satinder Singh. Discovery of options via meta-learned subgoals. Advances in Neural Information Processing Systems, 34:29861–29873, 2021

    2021

  4. [12]

    Multi-layer abstraction for nested generation of options (mango) in hierarchical reinforcement learning.IFAC-PapersOnLine, 59(26):25–30, 2025

    Alessio Arcudi, Davide Sartor, Alberto Sinigaglia, Vincent François-Lavet, and Gian Anto- nio Susto. Multi-layer abstraction for nested generation of options (mango) in hierarchical reinforcement learning.IFAC-PapersOnLine, 59(26):25–30, 2025

    2025

  5. [13]

    Value Iteration Networks.Advances in Neural Information Processing Systems, 29, 2016

    Aviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine, and Pieter Abbeel. Value Iteration Networks.Advances in Neural Information Processing Systems, 29, 2016

    2016

  6. [14]

    RUDDER: Return Decomposition for Delayed Rewards

    Jose A Arjona-Medina, Michael Gillhofer, Michael Widrich, Thomas Unterthiner, Johannes Brandstetter, and Sepp Hochreiter. RUDDER: Return Decomposition for Delayed Rewards . Advances in Neural Information Processing Systems, 32, 2019

    2019

  7. [15]

    Successor Features for Transfer in Reinforcement Learning.Advances in Neural Information Processing Systems, 30, 2017

    André Barreto, Will Dabney, Rémi Munos, Jonathan J Hunt, Tom Schaul, Hado P van Hasselt, and David Silver. Successor Features for Transfer in Reinforcement Learning.Advances in Neural Information Processing Systems, 30, 2017

    2017

  8. [16]

    The Option-Critic Architecture

    Pierre-Luc Bacon, Jean Harb, and Doina Precup. The Option-Critic Architecture. InProceedings of the AAAI Conference on Artificial Intelligence, volume 31, 2017

    2017

  9. [17]

    Go-explore: a new approach for hard-exploration problems.arXiv preprint arXiv:1901.10995, 2019

    Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O Stanley, and Jeff Clune. Go-explore: a new approach for hard-exploration problems.arXiv preprint arXiv:1901.10995, 2019

    1901

  10. [18]

    Springer Science & Business Media, 2004

    Andrea Toselli and Olof Widlund.Domain decomposition methods-algorithms and theory, volume 34. Springer Science & Business Media, 2004

    2004

  11. [19]

    Butterworth- Heinemann Oxford, UK:, 2013

    Robert Leroy Taylor and Olgierd Cecil Zienkiewicz.The finite element method. Butterworth- Heinemann Oxford, UK:, 2013

    2013

  12. [20]

    Divide-and- conquer reinforcement learning.arXiv preprint arXiv:1711.09874, 2017

    Dibya Ghosh, Avi Singh, Aravind Rajeswaran, Vikash Kumar, and Sergey Levine. Divide-and- conquer reinforcement learning.arXiv preprint arXiv:1711.09874, 2017

    2017 arXiv

  13. [21]

    State space decomposition and subgoal creation for transfer in deep reinforcement learning.arXiv preprint arXiv:1705.08997, 2017

    Himanshu Sahni, Saurabh Kumar, Farhan Tejani, Yannick Schroecker, and Charles Isbell. State space decomposition and subgoal creation for transfer in deep reinforcement learning.arXiv preprint arXiv:1705.08997, 2017

    2017 arXiv

  14. [22]

    State-space decomposition for reinforcement learning

    Esther Wong, Kin Leung, and Tony Field. State-space decomposition for reinforcement learning. Dept. Comput., Imperial College London, London, UK, Rep, 2021

    2021

  15. [23]

    Q-Cut—Dynamic Discovery of Sub-goals in Reinforcement Learning

    Ishai Menache, Shie Mannor, and Nahum Shimkin. Q-Cut—Dynamic Discovery of Sub-goals in Reinforcement Learning. InEuropean Conference on Machine Learning, pages 295–306. Springer, 2002

    2002

  16. [24]

    On the bottleneck concept for options discovery: Theoretical underpinnings and extension in continuous state spaces.Masters thesis, McGill University, 2014

    Pierre-Luc Bacon. On the bottleneck concept for options discovery: Theoretical underpinnings and extension in continuous state spaces.Masters thesis, McGill University, 2014

    2014

  17. [25]

    First results with Dyna, an integrated architecture for learning, planning and reacting.Neural Networks for Control, 179, 1990

    Richard S Sutton. First results with Dyna, an integrated architecture for learning, planning and reacting.Neural Networks for Control, 179, 1990

    1990

  18. [26]

    Proto-Value Functions: Developmental Reinforcement Learning

    Sridhar Mahadevan. Proto-Value Functions: Developmental Reinforcement Learning. In Proceedings of the 22nd International Conference on Machine Learning, pages 553–560, 2005

    2005

  19. [27]

    Policy invariance under reward transforma- tions: Theory and application to reward shaping

    Andrew Y Ng, Daishi Harada, and Stuart Russell. Policy invariance under reward transforma- tions: Theory and application to reward shaping. InICML, volume 99, pages 278–287. Citeseer, 1999. 13

    1999

  20. [28]

    Dealing with sparse rewards in reinforcement learning.arXiv preprint arXiv:1910.09281, 2019

    Joshua Hare. Dealing with sparse rewards in reinforcement learning.arXiv preprint arXiv:1910.09281, 2019

    1910

  21. [29]

    Reinforcement Learning for Adaptive Mesh Refinement

    Jiachen Yang, Tarik Dzanic, Brenden Petersen, Jun Kudo, Ketan Mittal, Vladimir Tomov, Jean-Sylvain Camier, Tuo Zhao, Hongyuan Zha, Tzanio Kolev, et al. Reinforcement Learning for Adaptive Mesh Refinement. InInternational Conference on Artificial Intelligence and Statistics, pa...

    2023

  22. [30]

    Swarm Reinforcement Learning For Adaptive Mesh Refinement.Advances in Neural Information Processing Systems, 36:73312–73347, 2023

    Niklas Freymuth, Philipp Dahlinger, Tobias Würth, Simon Reisch, Luise Kärger, and Gerhard Neumann. Swarm Reinforcement Learning For Adaptive Mesh Refinement.Advances in Neural Information Processing Systems, 36:73312–73347, 2023

    2023

  23. [31]

    Multi-agent reinforcement learning for subgrid-scale modeling of environmental turbulence

    Rambod Mojgani, Daniel Wälchli, Yifei Guan, Petros Koumoutsakos, and Pedram Hassanzadeh. Multi-agent reinforcement learning for subgrid-scale modeling of environmental turbulence. In 76th Annual Meeting of the Division of Fluid Dynamics. APS, 2023

    2023

  24. [32]

    Enhancing data efficiency in reinforcement learning: a novel imagination mechanism based on mesh information propagation.arXiv preprint arXiv:2309.14243, 2023

    Zihang Wang and Maowei Jiang. Enhancing data efficiency in reinforcement learning: a novel imagination mechanism based on mesh information propagation.arXiv preprint arXiv:2309.14243, 2023

    2023

  25. [33]

    Mesh-rft: Enhancing mesh generation via fine-grained reinforcement fine-tuning.arXiv preprint arXiv:2505.16761, 2025

    Jian Liu, Jing Xu, Song Guo, Jing Li, Jingfeng Guo, Jiaao Yu, Haohan Weng, Biwen Lei, Xianghui Yang, Zhuo Chen, et al. Mesh-rft: Enhancing mesh generation via fine-grained reinforcement fine-tuning.arXiv preprint arXiv:2505.16761, 2025

    2025

  26. [34]

    Courier Corporation, 2003

    Thomas JR Hughes.The finite element method: linear static and dynamic finite element analysis. Courier Corporation, 2003

    2003

  27. [35]

    Klaus-Jurgen Bathe, 2006

    Klaus-Jürgen Bathe.Finite element procedures. Klaus-Jurgen Bathe, 2006

    2006

  28. [36]

    McGraw- Hill New York, 2005

    Junuthula Narasimha Reddy.An introduction to the finite element method, volume 3. McGraw- Hill New York, 2005

    2005

  29. [37]

    Parallel domain decomposition software

    William Gropp and Barry Smith. Parallel domain decomposition software. InDomain-Based Parallelism and Problem Decomposition Methods in Computational Science and Engineering, pages 97–106. SIAM, 1995

    1995

  30. [38]

    Learning from delayed rewards.Ph

    Christopher JCH Watkins. Learning from delayed rewards.Ph. D. thesis, King’s College, University of Cambridge, 1989

    1989

  31. [39]

    Q-learning.Machine Learning, 8(3):279–292, 1992

    Christopher JCH Watkins and Peter Dayan. Q-learning.Machine Learning, 8(3):279–292, 1992

    1992

  32. [40]

    University of Cambridge, Department of Engineering, Cambridge, UK, 1994

    Gavin A Rummery and Mahesan Niranjan.On-line Q-learning using connectionist systems, volume 37. University of Cambridge, Department of Engineering, Cambridge, UK, 1994

    1994

  33. [41]

    Integrated Modeling and Control Based on Reinforcement Learning and Dynamic Programming.Advances in Neural Information Processing Systems, 3, 1990

    Richard S Sutton. Integrated Modeling and Control Based on Reinforcement Learning and Dynamic Programming.Advances in Neural Information Processing Systems, 3, 1990

    1990

  34. [42]

    Schwarz methods over the course of time.Electron

    Martin J Gander et al. Schwarz methods over the course of time.Electron. Trans. Numer. Anal, 31(5):228–255, 2008

    2008

  35. [43]

    Generalization in Reinforcement Learning: Safely Approxi- mating the Value Function.Advances in Neural Information Processing Systems, 7, 1994

    Justin Boyan and Andrew Moore. Generalization in Reinforcement Learning: Safely Approxi- mating the Value Function.Advances in Neural Information Processing Systems, 7, 1994. 14

    1994

Pith tools

Reviewed June 26, 2026 · model on record in the stance chip above.