Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Strategically Linked Decisions in Long-Term Planning and Reinforcement Learning

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 ↗ Review JSON

Pith's one-line read A strategic link score quantifies when decisions exist only to enable later decisions.

desk verdict A clean, simple measure of strategic dependence between decisions; the GridWorld and Shortcuts results are convincing, but the traffic 'road closure' experiment does not actually instantiate the score as defined. read the letter →

arxiv 2505.16833 v1 pith:7JD7JGVN submitted 2025-05-22 cs.LG

Alihan Hüyük , Finale Doshi-Velez This is my paper · ORCID
classification cs.LG
keywords strategiclinkscorelong-termplanningreinforcementlearningplanning-levelexplanationssafepolicyimprovementinversehorizoninterventionalmeasurement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes that long-horizon strategy in decision-making can be measured decision by decision. Its strategic link score defines the strength of the dependency between a set-up action and a later pay-off action as the drop in the probability of the set-up when the pay-off is forbidden. The paper argues this score is not tied to a particular planner or to rewards: it can be computed for any agent whose planning can be intervened on, and it supports three uses—explaining black-box policies, making action recommendations safer by grouping linked changes, and estimating how far ahead an emergent system plans, as demonstrated in a traffic simulator. A reader should care because, if the score works as claimed, it gives a general, intervention-based language for the structure of strategies, not just for the policies they produce.

What carries the argument

The central object is the strategic link score of Equation (2): for a planner $P^{\dagger}$, the score between a set-up decision $(s,a)$ and a pay-off decision $(\tilde s,\tilde a)$ is $\pi^{\dagger}(a|s) - \pi^{\dagger:\{\pi(\tilde a|\tilde s)=0\}}(a|s)$, the drop in the set-up action's probability when the pay-off action is constrained to have probability zero. This object carries the argument by turning the counterfactual question "what would the planner do if the follow-up were unavailable" into a single number. Higher-order applications build on it: scoring all pairs along a high-probability trajectory yields planning-level explanations; grouping recommendations whose pairwise scores are large yields strategy-aware recommendation sets; and measuring pre-intervention versus post-intervention flow rates in a continuous-action traffic setting yields the same score, with continuous decisions handled by constraining a region of pay-off decisions instead of a single point.

What would settle it

Set the simulated drivers in UXsim to follow a known optimal routing policy and close the arterial road past J10; the optimal policy yields a peak strategic link at J1. If the measured pre/post intervention flow differences do not match the link scores computed by directly constraining the optimal policy, then the road closure is not implementing the abstract pay-off constraint and the traffic-based horizon estimate does not measure Equation (2).

Watch

Extended reading notes

Core claim

The central claim is that the strategic link score $S^{\dagger}_{(s,a)\to(\tilde s,\tilde a)} = \pi^{\dagger}(a|s) - \pi^{\dagger:\{\pi(\tilde a|\tilde s)=0\}}(a|s)$ measures how much a decision is made as a setup for a later decision. The paper calls $(s,a)$ the set-up decision and $(\tilde s,\tilde a)$ the pay-off decision, and interprets the difference in the planner's action probabilities as the drop in the set-up's likelihood once the pay-off is unavailable. A key demonstration is that two planners can produce exactly the same optimal policy while carrying completely different strategic links, so link scores expose the planning process rather than just the policy. The paper then shows the score works across three applications: explaining agents by scoring pairs along a likely trajectory, improving worst-case recommendation safety by grouping strategically linked changes, and measuring the planning horizon of emergent driver behavior through road-closure interventions in a traffic simulator.

Load-bearing premise

The traffic experiment treats closing the arterial road past J10 as equivalent to imposing the policy constraint that the pay-off decision at J10 is unavailable to a single emergent planner, while a real road closure also changes the road network's transition dynamics and each driver's alternatives, so the measured flow-rate difference is not automatically the strategic link score defined by Equation (2).

Editorial extensions

If this is right

  • Explanations become planning-level: a user can see not only what a policy does, but which actions exist purely to unlock later actions, including dependencies that are invisible from the policy alone.
  • Recommendation systems become safer under partial adoption: strategically linked changes are presented as a single all-or-nothing group, avoiding the worst case where one change is adopted alone and performance drops, while independent recommendations can still be chosen individually.
  • The planning horizon of an unobservable agent can be estimated from interventions alone: the time distance between strongly linked decisions, measured through road closures or analogous prohibitions, characterizes how far ahead the agent plans.
  • Demonstrations are sufficient to recover strategic structure: inverse reinforcement learning infers a reward function that supports counterfactual policies, and the resulting link scores remain accurate even when the reward function itself becomes unidentifiable, because uniformly random policies imply no strategic links.
  • The score applies beyond reward-based RL, so emergent, multi-agent, or otherwise opaque planning systems can be analyzed as long as one can intervene to remove a follow-up decision.

Reading between the lines

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

  • Beyond the paper, pairwise link scores from demonstrations will inherit the assumptions of the planner used during inverse RL; the same demonstrations could support different strategic-link conclusions under a different planner model, so explanations should be reported together with the assumed counterfactual planner.
  • The paper's own limitation suggests a natural extension: strategies that are redundant across multiple pay-offs, such as 'A is taken to enable either B or C', are missed by pairwise constraints, and defining joint constraints over sets of pay-off decisions would capture such higher-order structure.
  • Negative link scores, such as the one measured at J1 in the traffic experiment, point to feedback-mediated strategies—actions taken to avoid congestion caused by the very route the pay-off decision depends on, suggesting the score could detect indirect environment-mediated dependencies in emergent multi-agent behavior.
  • Because the score is an interventional quantity, applying it to real systems requires interventions that block only the pay-off decision rather than changing the environment's dynamics; when that separation fails, the measured quantity is not the strategic link score as defined.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 introduces a formal quantity called the strategic link score, defined in Eq. (2) as the drop in the probability of a set-up decision (s,a) when a pay-off decision (s̃,ã) is made unavailable through the policy constraint π(ã|s̃)=0. The score is meant to capture whether an early action is taken mainly to enable a later action. The authors illustrate the definition with a toy MDP, then present three applications: planning-level explanations of soft-optimal RL agents in GridWorld, strategy-aware safe recommendations in a procedurally generated Shortcuts environment, and measurement of the planning horizon of drivers in a UXsim traffic simulation through road closures. The paper also studies how well the scores can be recovered from demonstrations via maximum entropy IRL.

Significance. If the central definition holds, the strategic link score is a simple and broadly applicable tool for planning-level analysis, and the first two applications are convincing demonstrations: the toy example cleanly separates two planners with identical policies but different strategic dependencies, the GridWorld experiments show that Eq. (2) produces interpretable explanations, and the Shortcuts experiments show that strategy-aware grouping can improve both average and worst-case recommendation outcomes. The IRL experiments are also a useful practical check. However, the third application does not implement Eq. (2): the traffic intervention changes the environment dynamics rather than constraining a fixed planner's policy, and the continuous-action treatment is internally inconsistent with Section 4. The planning-horizon conclusion is therefore not established by the formal definition.

major comments (3)
  1. [Section 5.3] The traffic experiment does not implement Eq. (2). The strategic link score in Eq. (2) compares two policies of the same planner in the same environment E: the unconstrained policy π† and the policy planned under the constraint π(ã|s̃)=0. In Figures 9 and 10, the post-intervention routing policy is instead measured after physically closing the arterial road past J10, which changes the transition dynamics, travel times, congestion, and information available to drivers. Even though the closure removes the J10 arterial action, it does so by altering the environment, so the pre/post flow differences in Figure 10c conflate the counterfactual policy drop with direct dynamics effects such as spillback and forced diversions. Consequently, the claim that the strongest link at J9 reveals the drivers' effective planning horizon is not a consequence of the proposed score, and the negative link at J1 is not shown to be a strategic avoidance rather than a congestion response. This identification is load-bearing for the third application; the authors should either implement a genuine policy-constraint intervention or substantially reframe the traffic analysis as an environment-intervention study that is outside Eq. (2).
  2. [Section 4 and Section 5.3] The continuous-action treatment is internally inconsistent. Section 4 states that for continuous states and actions a point constraint such as π(ã|s̃)=0 'would practically have no impact' and therefore requires constraining a region S̃×Ã, yet Section 5.3 defines S†_{JX→J10}=π†(JX)-π†:{π(J10)=0}(JX) with a point constraint and reports that the J10 score is one 'by definition.' A measure-zero constraint on a continuous action distribution is not equivalent to the road closure used in the simulation. The paper should either specify the region constraint used in the traffic example (for example, constraining π(J10) to a low-flow interval, as suggested in Section 4) or justify why the point constraint is well-defined for this emergent policy; as written, the traffic scores are not an instance of the formal definition.
  3. [Section 5.2 and Appendix A.2] The strategy-aware recommendation algorithm depends on a significance threshold, set in Appendix A.2 to 'halfway between 0 and 1/5,' and the safety and effectiveness comparison in Figure 7 depends on the groups produced by that threshold. The paper gives a rationale for the 1/5 bound but no sensitivity analysis, so it is unclear how robust the claimed advantage of Strategy-Aware over Pick-and-Choose and All-or-Nothing is to the threshold choice. Please add a sensitivity study or replace the heuristic threshold with a more principled criterion.
minor comments (5)
  1. [Section 5.1] There is a typo in 'maximum entorpy IRL (Ziebart et al., 2008)' in Section 5.1, Approach 2; it should be 'maximum entropy.'
  2. [Appendix A.1] The appendix says that when constraining a decision the reward is set to r(i,j,{f'k},a)←∞, but the main text and the intended constraint require −∞; this sign error should be corrected.
  3. [Figure 4 caption] The caption refers to 'the second layout (Figure 4a)' when describing the correlated-keys layout; it should refer to Figure 4c.
  4. [Section 2] The sentence 'explainability in RL, which can be address at state, policy, or planning levels' contains a grammatical error; 'address' should be 'addressed.'
  5. [General] The paper does not include a reproducibility statement or a link to the code and data used for the simulations; adding these would make the experimental results easier to verify, especially for the UXsim traffic experiment.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the strategic link score is a definition, and each application evaluates it against independent environment structure or ground-truth rewards; self-citations are contextual only.

full rationale

The paper's central quantity, S in Equation (2), is introduced as a definition rather than as a fitted or predicted value: it is the difference between the unconstrained policy probability and the policy probability under an explicit constraint. The GridWorld and Shortcuts experiments compute this quantity from a specified soft-optimal planner or from a reward recovered by maximum-entropy IRL, then compare the resulting links against known environment structure such as key-door dependencies, so the conclusions are not baked into the inputs. The IRL validation compares link scores derived from inferred rewards with link scores computed from the true reward, reporting error rather than assuming agreement, which is an external check. The traffic application measures the score as the difference between pre- and post-intervention routing policies, which operationalizes the definition; the physical road closure also changes environment dynamics, a possible validity limitation that the paper partially acknowledges in its Limitations paragraph, but this does not make the derivation circular because no target conclusion is imposed by construction. Related-work citations to prior papers by the authors are contextual and are not used to justify the definition or the main results. No step in the claimed derivation reduces to its own input.

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

No new physical or formal entities are introduced beyond the strategic link score definition itself, which is a quantity computed from existing policy probabilities and constrained counterfactual policies. The score is well-defined in discrete settings, but the traffic application rests on the ad hoc assumption that a road closure is a policy constraint, and the IRL application rests on standard MaxEnt IRL assumptions.

free parameters (4)
  • significance threshold for strategic link grouping = 0.1 (half of 1/5)
    In Algorithm 2, a link score is treated as significant if it is above this threshold; chosen by hand as the midpoint between 0 and the maximum possible score for a single preparation action, not grounded in statistical theory.
  • arterial free flow speed = 20 * sqrt(J/(J+1))
    Chosen as the geometric mean of the speeds that would make arterial and highway routes equal or arterial always preferable; this creates the strategic tension needed for the traffic experiment and is a modeling choice, not measured from data.
  • preparation action cost C = 0.1
    Set so that shortcuts after preparation are strictly better than moving; a hyperparameter of the Shortcuts environment that controls the trade-off the recommendations exploit.
  • inverse temperature beta = 100
    Near-deterministic soft-optimal policies in GridWorld and Shortcuts; the magnitude affects the score values but not the definition of the score.
assumptions (4)
  • standard math Soft value iteration converges to the unique soft-optimal policy for the given reward and temperature.
    Used throughout to compute π† and constrained policies; referenced from Haarnoja et al. (2017).
  • domain assumption The emergent routing behavior of the traffic simulator can be represented as a stationary policy π†(JX) equal to the frequency of drivers taking the arterial road, with these frequencies behaving like action probabilities under a single planner.
    Section 5.3 introduces π† and derives strategic link scores from pre/post intervention flow rates; this requires treating many independent drivers as one planning entity.
  • ad hoc to paper Closing the arterial road at J10 is equivalent to imposing the constraint π(J10)=0 on that policy while leaving all other environment dynamics and rewards fixed.
    Section 5.3, 'Intervention', uses the road closure to compute S†_JX→J10; but the closure physically changes the road network and driver options, not just the probability of one action at one state.
  • domain assumption Demonstrations are generated by a soft-optimal planner with known transition dynamics, so MaxEnt IRL recovers a reward whose implied counterfactual policies reproduce the true strategic link scores.
    Section 5.1 Approach 2 relies on this for the IRL-based link computation; IRL is otherwise unidentifiable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Strategically Linked Decisions in Long-Term Planning and Reinforcement Learning." pith.science (2026). https://pith.science/paper/7JD7JGVN

@misc{pith2026250516833,
  author       = {Pith},
  title        = {Pith review of: Strategically Linked Decisions in Long-Term Planning and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7JD7JGVN}},
  note         = {Machine review of arXiv:2505.16833}
}
read the original abstract

Long-term planning, as in reinforcement learning (RL), involves finding strategies: actions that collectively work toward a goal rather than individually optimizing their immediate outcomes. As part of a strategy, some actions are taken at the expense of short-term benefit to enable future actions with even greater returns. These actions are only advantageous if followed up by the actions they facilitate, consequently, they would not have been taken if those follow-ups were not available. In this paper, we quantify such dependencies between planned actions with strategic link scores: the drop in the likelihood of one decision under the constraint that a follow-up decision is no longer available. We demonstrate the utility of strategic link scores through three practical applications: (i) explaining black-box RL agents by identifying strategically linked pairs among decisions they make, (ii) improving the worst-case performance of decision support systems by distinguishing whether recommended actions can be adopted as standalone improvements or whether they are strategically linked hence requiring a commitment to a broader strategy to be effective, and (iii) characterizing the planning processes of non-RL agents purely through interventions aimed at measuring strategic link scores - as an example, we consider a realistic traffic simulator and analyze through road closures the effective planning horizon of the emergent routing behavior of many drivers.

Figures

Figures reproduced from arXiv: 2505.16833 by the authors.

Figure 1
Figure 1. An example strategy in a navigation task. Picking up the key takes extra time early on but unlocks a major shortcut later. If the shortcut were to be blocked, going for the key would no longer be optimal, showing that picking up the key and taking the shortcut are strategically linked—the key is picked up specifically to be able to take the shortcut. A fundamental goal in reinforcement learning (RL) is to find long-… view at source ↗
Figure 2
Figure 2. Dynamics of the toy example. For initial state S1, the action sequence (A2, A2) is optimal under both reward functions rα and rβ. For rα, the two ac￾tions are strategically linked as A2|S1 is not optimal unless A2|S2. For rβ, there is no strategic link as A2|S1 is optimal regardless of the action taken at state S2. A Toy Example. Consider a simple environment with two states S = {S1, S2} and two actions A = {A1, A2}… view at source ↗
Figure 3
Figure 3. Strategic link scores for a simple maze layout. Creating a shortcut by picking up the key to unlocking the door is strategic (a), since blocking the door results in the key not being picked up (b), while constraining an unrelated action does not lead to the same outcome (c). By looking at the strategic link scores between all the decisions along the optimal trajectory (d), the link between the key (“K”) and the door… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Strategic link scores for layout with independent vs. correlated keys. When the keys unlock separate shortcuts (a), they is no strategically link between them (b). If one key is skipped, collecting the other still remains optimal. When the keys jointly unlock a single …
Figure 5
Figure 5. Figure 5: Strategic link scores inferred from demonstrations become more accurate with increasing variation, following a similar trend to reward inference (until policies are almost uniformly random and rewards become unidentifiable, strategic link scores remain accurate as reco…
Figure 6
Figure 6. Figure 6: An example environment that has 5 nodes, 3 shortcuts, and 4 preparation actions. After taking the required preparation actions, the agent can jump via shortcuts to move more efficiently towards the target. 0 1 2 3 4 5 0.5 1.0 1.5 Number of Recommendations Implemented F…
Figure 7
Figure 7. Figure 7: Performance following various recommenda￾tion methods. Pick-and-Choose is not safe, potentially leading to worse performance than to begin with. All￾or-Nothing is safe but not effective unless a large num￾ber of recommendations are implemented. Strategic￾Aware is both …
Figure 8
Figure 8. Figure 8: The traffic scenario. ‘Entry’ to ‘exit’, drivers need to decide whether to stay on an arterial road or divert to the highway at junctions ‘J1’ to ‘J10’. Despite having a lower speed limit, the arterial route is shorter hence overall quicker. But, if one of its segments…
Figure 9
Figure 9. Figure 9: Vehicle counts. The rate at which vehicles pass through each junction and still stay on the arterial road changes significantly after the intervention at J10. Simulation. We run a simulation for 50k time steps, where the intervention to close off the ar￾terial road at …
Figure 10
Figure 10. Figure 10: Strategic link scores for the simulated drivers. Using the count data in [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Strategic link scores for the optimal routing behavior. Unlike the behavior of the simulated drivers, the optimal response to the road clo￾sure past J10 would have been at J1. Results. First, we see that the link score for J10 is one. This is by definition of link sco…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages

  1. [1]

    Abbeel, P. and A. Y. Ng, ``Apprenticeship learning via inverse reinforcement learning,'' in International Conference on Machine Learning, 2004

    work page 2004

  2. [2]

    Amir, D. and O. Amir, `` HIGHLIGHTS : Summarizing agent behavior to people,'' in International Conference on Autonomous Agents and Multiagent Systems, 2018

    work page 2018

  3. [3]

    Silvestri, G

    Chen, Z., F. Silvestri, G. Tolomei, J. Wang, H. Zhu, and H. Ahn, ``Explain the explainer: Interpreting model-agnostic counterfactual explanations of a deep reinforcement learning agent,'' IEEE Transactions on Artificial Intelligence, vol. 5, no. 4, pp. 1443--1457, 2022

    work page 2022

  4. [4]

    Dazeley, P

    Cruz, F., R. Dazeley, P. Vamplew, and I. Moreira, ``Explainable robotic systems: Understanding goal-driven actions in a reinforcement learning scenario,'' Neural Computing and Applications, vol. 35, no. 25, pp. 18\,113--18\,130, 2023

    work page 2023

  5. [5]

    Harrison, L

    Ehsan, U., B. Harrison, L. Chan, and M. O. Riedl, ``Rationalization: A neural machine translation approach to generating natural language explanations,'' in AAAI/ACM Conference on AI, Ethics, and Society, 2018

    work page 2018

  6. [6]

    Erwig, M., A. Fern, M. Murali, and A. Koul, ``Explaining deep adaptive programs via reward decomposition,'' in IJCAI/ECAI Workshop on Explainable Artificial Intelligence, 2018

    work page 2018

  7. [7]

    Dennis, S

    Gleave, A., M. Dennis, S. Legg, S. Russell, and J. Leike, ``Quantifying differences in reward functions,'' in International Conference on Learning Representations, 2021

    work page 2021

  8. [8]

    Greydanus, S., A. Koul, J. Dodge, and A. Fern, ``Visualizing and understanding Atari agents,'' in International Conference on Machine Learning, 2018

    work page 2018

Show all 31 references
  1. [9]

    Haarnoja, T., H. Tang, P. Abbeel, and S. Levine, ``Reinforcement learning with deep energy-based policies,'' in International Conference on Machine Learning, 2017

    2017

  2. [10]

    Udluft, and T

    Hein, D., S. Udluft, and T. A. Runkler, ``Interpretable policies for reinforcement learning by genetic programming,'' Engineering Applications of Artificial Intelligence, vol. 76, pp. 158--169, 2018

    2018

  3. [11]

    Demmler, S

    Huber, T., M. Demmler, S. Mertes, M. L. Olson, and E. Andr \'e , ``Ganterfactual-rl: Understanding reinforcement learning agents' strategies through visual counterfactual explanations,'' arXiv preprint arXiv:2302.12689, 2023

    2023 arXiv

  4. [12]

    Iyer, R., Y. Li, H. Li, M. Lewis, R. Sundar, and K. Sycara, ``Transparency and explanation in deep reinforcement learning neural networks,'' in AAAI/ACM Conference on AI, Ethics, and Society, 2018

    2018

  5. [13]

    Juozapaitis, Z., A. Koul, A. Fern, M. Erwig, and F. Doshi-Velez, ``Explainable reinforcement learning via reward decomposition,'' in IJCAI/ECAI Workshop on Explainable Artificial Intelligence, 2019

    2019

  6. [14]

    Poupart, and J

    Khan, O., P. Poupart, and J. Black, ``Minimal sufficient explanations for factored markov decision processes,'' in International Conference on Automated Planning and Scheduling, 2009

    2009

  7. [15]

    Trichelair, and R

    Laroche, R., P. Trichelair, and R. T. Des Combes, ``Safe policy improvement with baseline bootstrapping,'' in International Conference on Machine Learning, 2019

    2019

  8. [16]

    Schulte, W

    Liu, G., O. Schulte, W. Zhu, and Q. Li, ``Toward interpretable deep reinforcement learning with linear model U -trees,'' in Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2018

    2018

  9. [17]

    Miller, L

    Madumal, P., T. Miller, L. Sonenberg, and F. Vetere, ``Explainable reinforcement learning through a causal lens,'' in Proceedings of the AAAI conference on artificial intelligence, vol. 34, no. 03, 2020, pp. 2493--2500

    2020

  10. [18]

    Olson, M. L., R. Khanna, L. Neal, F. Li, and W.-K. Wong, ``Counterfactual state explanations for reinforcement learning agents via generative deep learning,'' Artificial Intelligence, vol. 295, p. 103455, 2021

    2021

  11. [19]

    Seo, T., `` UXsim : lightweight mesoscopic traffic flow simulator in pure Python ,'' Journal of Open Source Software, 2025

    2025

  12. [20]

    Benac, S

    Sharma, A., L. Benac, S. Parbhoo, and F. Doshi-Velez, ``Decision-point guided safe policy improvement,'' arXiv preprint arXiv:2410.09361, 2024

    2024 arXiv

  13. [21]

    Xiong, and R

    Shu, T., C. Xiong, and R. Socher, ``Hierarchical and interpretable skill acquisition in multi-task reinforcement learning,'' in International Conference on Learning Representations, 2018

    2018

  14. [22]

    Gombolay, T

    Silva, A., M. Gombolay, T. Killian, I. Jimenez, and S.-H. Son, ``Optimization methods for interpretable differentiable decision trees applied to reinforcement learning,'' in International Conference on Artificial Intelligence and Statistics, 2020

    2020

  15. [23]

    H \"u y \"u k, D

    Sun, H., A. H \"u y \"u k, D. Jarrett, and M. van der Schaar, ``Accountability in offline reinforcement learning: Explaining decisions with a corpus of examples,'' Conference on Neural Information Processing Systems, 2023

    2023

  16. [24]

    Topin, N. and M. Veloso, ``Generation of policy-level explanations for reinforcement learning,'' in AAAI Conference on Artificial Intelligence, 2019

    2019

  17. [25]

    van Diggelen, K

    van der Waa, J., J. van Diggelen, K. van den Bosch, and M. Neerincx, ``Contrastive explanations for reinforcement learning in terms of expected consequences,'' arXiv preprint arXiv:1807.08706, 2018

    2018 arXiv

  18. [26]

    Murali, R

    Verma, A., V. Murali, R. Singh, P. Kohli, and S. Chaudhuri, ``Programmatically interpretable reinforcement learning,'' in International Conference on Machine Learning, 2018

    2018

  19. [27]

    Wu, J., H. Wu, Z. Qiu, J. Wang, and M. Long, ``Supported policy optimization for offline reinforcement learning,'' Conference on Neural Information Processing Systems, 2022

    2022

  20. [28]

    Parbhoo, W

    Yao, J., S. Parbhoo, W. Pan, and F. Doshi-Velez, ``Policy optimization with sparse global contrastive explanations,'' arXiv preprint arXiv:2207.06269, 2022

    2022 arXiv

  21. [29]

    Russell, and S

    Yau, H., C. Russell, and S. Hadfield, ``What did you think would happen? explaining agent behaviour through intended outcomes,'' in Conference on Neural Information Processing Systems, 2020

    2020

  22. [30]

    Ziebart, B. D., A. L. Maas, J. A. Bagnell, A. K. Dey et al., ``Maximum entropy inverse reinforcement learning.'' in AAAI Conference on Artificial Intelligence, 2008

    2008

  23. [31]

    write newline

    " write newline "" initialize.prev.this.status FUNCTION begin.bib " write newline preamble empty 'skip preamble write newline if " thebibliography " number.label int.to.str * " " * write newline " [1] #1 " write newline " [1] #1 " write newline " url@samestyle " write newline ...

Pith tools

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