Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Causal-Inspired Multi-Agent Decision-Making via Graph Reinforcement Learning

T0 review · 4 major / 6 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 causal filter learned inside a variational graph auto-encoder lowers collision rates and raises rewards for an autonomous vehicle at an unsignalized intersection.

desk verdict The causal filter's target A*_av is undefined and the causal graph contradicts itself, so the paper's central claim is unsupported despite a new pipeline comparison. read the letter →

arxiv 2507.23080 v1 pith:2R6QAXXW submitted 2025-07-30 cs.MA

Jing Wang , Yan Jin , Fei Ding , Chongfeng Wei This is my paper
classification cs.MA
keywords CausalDisentanglementRepresentationLearningGraphReinforcementVariationalAuto-EncoderD3QNAutonomousDrivingUnsignalizedIntersectionMulti-AgentDecision-MakingCollisionAvoidance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that an autonomous vehicle at an unsignalized intersection makes safer and higher-reward decisions when its reinforcement-learning policy receives causal features extracted from the surrounding traffic graph instead of raw correlations. The proposed CGRL framework uses causal disentanglement representation learning inside a variational graph auto-encoder to separate features that genuinely influence the ego vehicle's optimal decision from spurious, confounded features, then feeds the causal features into a graph neural network-based dueling double deep Q-network policy. This would give graph reinforcement learning a concrete remedy for its known weaknesses: dependence on observed correlations and poor sample efficiency in multi-agent traffic. The paper supports the claim with simulator experiments across left-turn, straight, and right-turn tasks, where CGRL reports the lowest collision rate and the highest average reward among seven compared methods.

What carries the argument

The load-bearing mechanism is the causal filter inside a variational graph auto-encoder. A two-layer graph convolutional encoder produces latent representations $Z$; the filter splits them into causal features $Z_c$ and spurious features $Z_s$, then scores dependencies with matrix-based Rényi-$\alpha$ entropy, computed from the eigenspectrum of Gram matrices. The training objective minimizes $-I_\alpha(Z_c;A^*_{av}|Z_s) + I_\alpha(Z_c;Z_s) + \lambda_1 L_{VGAE} + \lambda_2 \|A_c\|_1/\|A\|_1$, which strengthens the direct causal path from $Z_c$ to the optimal decision $A^*_{av}$, removes the confounding influence of $Z_s$, and keeps the representation close to the VGAE reconstruction. The filtered causal features are decoded into a causal adjacency matrix $A_c = \sigma(Z_c Z_c^T)$ that serves as the policy state input.

What would settle it

Train the full CGRL pipeline twice in the same simulator with identical seeds, once using the causal objective of Eq. (32) with the correct $A^*_{av}$ and once with $A^*_{av}$ replaced by an action sampled uniformly from the three-action space. If collision rate and average reward stay essentially unchanged, the causal term is not what produces the reported gains; if they degrade sharply, the filter depends on a meaningful reference decision.

Watch

Extended reading notes

Core claim

The paper's central claim is that a graph neural network-based reinforcement learning policy for a single autonomous ego vehicle at an unsignalized intersection performs better when its state representation is restricted to causally disentangled features learned by a variational graph auto-encoder. The authors model the traffic scene as a graph, encode it into latent features, and use a causal filter that minimizes $I_\alpha(Z_c;Z_s)$ while maximizing $I_\alpha(Z_c;A^*_{av}|Z_s)$, where $Z_c$ and $Z_s$ are the causal and spurious latent features and $A^*_{av}$ is the ego vehicle's optimal decision. Intervening on the spurious features is meant to block the backdoor path $Z_c \leftarrow G \rightarrow Z_s \rightarrow A^*_{av}$. The resulting causal features are decoded into a causal adjacency matrix that becomes the state input to a GCN-GAT-dueling D3QN policy. In 2000 testing episodes, CGRL reports collision rates of 13.50% for left turns, 11.35% for straight crossings, and 7.00% for right turns, against best baseline rates of 19.40%, 14.20%, and 9.20%, and the highest average reward in all three tasks.

Load-bearing premise

The whole method leans on a quantity the paper calls the optimal decision-making of the ego vehicle, but it never says how that quantity is obtained; if it is just the agent's own sampled action, the causal filter is fit to the policy and then fed back into that same policy, creating a circular loop.

Editorial extensions

If this is right

  • CGRL's reported collision rates are 13.50% for left turns, 11.35% for straight crossings, and 7.00% for right turns, lower than all six baseline policies in the same scenario.
  • CGRL attains the highest average reward on all three maneuvers, which the paper reads as evidence that the causal filter improves both safety and efficiency rather than trading one for the other.
  • The same CGRL architecture is applied to left-turn, straight, and right-turn tasks without task-specific changes, indicating the causal-feature extraction is not maneuver-specific.
  • GCN-GAT-D3QN is the baseline that combines the same GCN, GAT, and D3QN components; the reported gap between CGRL and that baseline is the contribution of the causal module.

Reading between the lines

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

  • The paper does not test this, but the causal graph it assumes, with $Z_c$ and $Z_s$ split around the decision target $A^*_{av}$, is not tied to intersection geometry, so the same filter could be carried to roundabouts, lane merges, or highway ramps.
  • Because $A^*_{av}$ is never sourced, a natural next experiment is to supply it from a rule-based expert or from a separate trained policy and compare; the result would show whether the method is learning causality or just recycling its own policy's choices.
  • The reported numbers come from one simulator with one ego vehicle and 15 human-driven vehicles; a reader should not treat them as evidence about real intersections until the framework is tested with sensor noise, perception errors, and heterogeneous traffic.
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 / 6 minor

Summary. The paper proposes CGRL, a framework that combines causal disentanglement representation learning (CDRL) with graph reinforcement learning for an autonomous ego vehicle (AEV) navigating an unsignalized intersection. The method uses a variational graph auto-encoder (VGAE) to learn latent representations, then applies an information-theoretic causal filter that separates causal features Zc from spurious features Zs by maximizing conditional mutual information I(Zc; A*_av | Zs), where A*_av is called the optimal decision-making of the AEV. The extracted causal features are fed into a GCN-GAT-D3QN policy. Experiments in the highway-env simulator compare CGRL against six graph-based DQN variants across left-turn, straight, and right-turn tasks, reporting collision rate, average reward, and average velocity.

Significance. If the causal filter truly extracts invariant causal features that improve safety and efficiency, the framework would be a useful step toward causality-aware decision-making in multi-agent autonomous driving. The paper provides a complete algorithmic description and a comparative evaluation in a standard simulator, which is a constructive basis for further work. However, the central causal claim rests on the undefined variable A*_av and on a causal graph that is internally inconsistent, so the reported improvements cannot currently be attributed to causal learning. The empirical evaluation also lacks error bars and ablations, making the headline performance claims unverifiable from the presented data.

major comments (4)
  1. [Section IV-D1 and IV-D3, Eq. (32)] The variable A*_av, described as the 'optimal decision-making' of the AEV, is never defined or sourced. In a standard MDP without demonstrations or an external expert, the only available candidate for A*_av is the agent's own action sampled from the online Q-network; if that is the case, the causal filter in Eq. (32) fits Zc to the very policy it then trains, creating a self-referential loop with no independent causal target. The authors must specify how A*_av is obtained (e.g., an oracle policy, human demonstrations, or a separate expert) and justify that this quantity is genuinely optimal; without this, the entire causal contribution in Eq. (32) is unsupported.
  2. [Section IV-D1, causal graphical model] The causal graph is stated as 'Zc ← G → Zs → A*_av', which contains no directed edge from Zc to A*_av, yet the text claims that Zc 'exerts a direct causal influence' on A*_av. Under the stated graph, Zc and A*_av are d-separated given Zs, so the conditional mutual information I(Zc; A*_av | Zs) in Eq. (28) would be zero and the maximization objective in Eq. (32) would be vacuous. The graph must include the edge Zc → A*_av, and the text describing Zc as a direct cause of A*_av must be made consistent with the graph, or the claims about causal influence must be revised.
  3. [Section IV-D3, Eq. (28)] The equality I(Zc → A*_av | do(Zs)) = I(Zc; A*_av | Zs) is not a valid do-calculus identity; intervening on Zs is generally not equivalent to conditioning on Zs. If the authors intend to identify the causal effect of Zc on A*_av by adjusting for the confounder Zs, they should invoke the backdoor criterion and verify that the graph satisfies the required conditions, rather than writing an equality that conflates do(Zs) with conditioning. As written, the intervention claim is unjustified.
  4. [Section VI-B, Table III] The claim that CGRL 'significantly outperforms' the baselines is not supported by error bars, multiple seeds, or statistical tests. The table reports single-point estimates (apparently over 2000 testing episodes), so the reader cannot determine whether the differences, such as 13.50% versus 20.15% collision rate for left turns, are meaningful or within noise. In addition, there is no ablation that removes the causal filter while keeping the same GCN-GAT-D3QN network and training procedure, so the observed improvements cannot be attributed to the causal features rather than to the added VGAE module and the extra loss terms in Eq. (32).
minor comments (6)
  1. [Abstract] The sentence 'make autonomous vehicles' navigation safer' should be 'makes autonomous vehicles' navigation safer and more efficient' to agree with the singular subject 'this study'.
  2. [Section I] The phrase 'This limitation arises these methods heavily rely' is missing a preposition or conjunction; it should be 'This limitation arises because these methods heavily rely' or similar.
  3. [Section IV-A4] The sentence 'where ωc_t, ωhs_t, ωor_t and ωtc_t are 1 predefined as respectively' is garbled; it should read 'where the weights ωc_t, ωhs_t, ωor_t, and ωtc_t are each predefined as 1, respectively' or similar.
  4. [Section IV-D2, Eq. (20)] The same weight matrix W1 is used for both μ and log σ² in the second GCN layer; in a standard VGAE these are separate weight matrices, so the authors should clarify whether the shared weight is intentional or a typographical error.
  5. [Section IV-D3, Eqs. (25)-(27)] The matrix-based Rényi entropy notation is imprecise: Sα should be defined on the eigenvalues of a normalized Gram matrix rather than on raw Zc, and the use of the Kronecker product D = Zc ⊗ Zs for the joint entropy requires an explicit explanation of how D is related to the joint Gram matrix.
  6. [Section VI-A] The statement that CGRL shows 'superior learning efficiency' is based only on a visual inspection of Figure 3; the authors should provide quantitative learning curve statistics or error bars to support this claim.

Circularity Check

2 steps flagged · score 7.0 of 10

CGRL's causal filter is fitted to an undefined target A*_av; in the implemented D3QN the only computable target is the agent's own argmax action, so the extracted causal features are a self-referential re-encoding of the policy and Table III cannot be credited to independent causal information.

  1. fitted input called prediction [Section IV-D (Eq. 32), with A*_av introduced in Section IV-D1 and actions defined in Eqs. (15)-(16), Section IV-C]
    "This includes graph data G, two isolated features derived from the hidden space in VGAE: causal features Zc and spurious features Zs, as well as the optimal decision-making of the autonomous vehicle A∗ av ... To obtain causal features Zc from the latent feature matrix Z, the overall loss function can be defined based on [35]: min −Iα(Zc; A∗ av|Zs) + Iα(Zc; Zs) + λ1LVGAE + λ2 ∥Ac∥1 ∥A∥1 ... The term −Iα(Zc; A∗ av|Zs) is minimized to maximize the causal influence between Zc and A∗ av."

    A∗_av, the target of the causal objective, is never defined or sourced. No expert policy, oracle, or external dataset is provided. In the D3QN implementation, the only decision variable computed is the online network's argmax (Eqs. 15-16). If A∗_av is that argmax, Eq. (32) maximizes mutual information between Zc and the current policy's own actions; Zc is then used to build Ac, which 'serves as the state input for training the CGRL decision-making algorithm' (Eq. 31). The causal features are therefore fitted to the policy and fed back into the same policy, so the improvement over baselines in Table III is not evidence for an independent causal mechanism.

  2. self definitional [Section IV-D3, Eq. (28)]
    "To identify direct causal connections between Zc and A∗ av. we can apply the do-calculus on the variable Zs to block the backdoor path. Accordingly, we propose an intervention mechanism that leverages a conditional MI method to quantify causal influence and identify the causal relationship by intervening on Zs. I (Zc → A∗ av| do (Zs)) = I (Zc; A∗ av|Zs)"

    The paper defines the causal intervention do(Zs) as equivalent to ordinary conditioning on Zs, which is not a consequence of do-calculus. Combined with the graph Zc ← G → Zs → A∗_av, which has no directed edge Zc → A∗_av, this turns the asserted 'direct causal influence' into a definitional relabeling of conditional dependence: the causal quantity is defined to be the conditional MI, so the extraction objective does not verify a causal effect but assumes it.

full rationale

The central causal contribution of CGRL is the extraction of causal features Zc that influence the AEV's optimal decision-making. That extraction is driven entirely by Eq. (32), whose only supervision signal is A∗_av. The paper never specifies where A∗_av comes from. The implemented agent has no external teacher: it is a D3QN trained by TD updates, and the only actions it ever computes are Qonline's argmax and the target-network evaluation. Therefore, in any concrete implementation consistent with the paper, A∗_av must be the agent's own decision variable. Maximizing I(Zc;A∗_av|Zs) then fits Zc to the current policy, and the same Zc is converted into a causal adjacency matrix and fed back as the policy's state input. This is a closed loop: the 'causal' representation is a re-encoding of the policy's existing choices, not an independent causal ground truth. Consequently, the collision-rate and reward improvements reported in Table III cannot be attributed to a genuinely causal mechanism; they could arise from the altered architecture, the extra VGAE loss, or the self-referential regularization. This is a constructive reduction of the paper's key claim to its own training signal, not a self-citation issue or a mere disagreement with the community. The paper's other components (GCN-GAT encoder, dueling D3QN, benchmark comparison) are standard and non-circular, but the signature causal contribution is unfalsifiable as stated because A∗_av is undefined. The score reflects that the central claim partially reduces by construction (7/10).

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The method introduces several free parameters and ad hoc modeling choices: two unspecified loss weights, an undocumented speed-reward map, arbitrary adjacency thresholds, an unspecified Rényi order, and an unspecified latent dimension. The causal graph and the intervention-by-conditioning step are assumptions with no validation. The two latent splits Zc and Zs are postulated entities with no independent falsifiable evidence, and the training objective appears to fit them to the agent's own evolving policy, making the causal claim circular.

free parameters (5)
  • λ1, λ2 loss weights = not reported
    Weights on VGAE loss and sparsity regularizer in Eq (32); no values or tuning procedure are given.
  • x0, x1, y0, y1 speed-reward map = not reported
    Defines the piecewise-linear high-speed reward in Eq (6); values are never given.
  • graph edge thresholds = Δx < 10 m, Δy < 30 m
    Determines the adjacency matrix in Section IV-A; the choice is arbitrary and not justified.
  • Rényi entropy order α = not reported
    Order of the matrix-based Rényi entropy in Eqs (25)-(30); unspecified.
  • VGAE latent dimension L = not reported
    Dimension of the latent representation Z in the VGAE; unspecified.
assumptions (4)
  • ad hoc to paper Causal graph Zc ← G → Zs → A*_av correctly represents the data-generating process.
    Assumed in Section IV-D1 without justification or validation, and the described graph is internally inconsistent regarding the role of Zs as a confounder.
  • ad hoc to paper Conditioning on Zs via conditional MI I(Zc; A*|Zs) implements the do-intervention do(Zs).
    Eq (28) equates intervention with conditioning; this is only valid under strong assumptions (e.g., no unmeasured confounders) that are not stated or checked.
  • domain assumption Invariant representations correspond to causal factors.
    Stated in the introduction as a premise for the whole CDRL approach; no evidence is provided that invariance to spurious features identifies true causes in this traffic setting.
  • standard math Matrix-based Rényi entropy gives accurate MI estimates for the high-dimensional features.
    Imported from [38]; no verification is provided that the estimator is accurate or stable for the specific feature distributions in this problem.
invented entities (2)
  • Causal features Zc
    purpose: Latent representation claimed to capture direct causes of optimal decisions.
    No external validation; only evaluated indirectly via downstream RL performance, with no ground-truth causal graph or intervention test.
  • Spurious features Zs
    purpose: Latent representation claimed to capture non-causal confounders or noise.
    No independent test distinguishes spurious from causal; the split is driven by the objective, which may encode policy bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal-Inspired Multi-Agent Decision-Making via Graph Reinforcement Learning." pith.science (2026). https://pith.science/paper/2R6QAXXW

@misc{pith2026250723080,
  author       = {Pith},
  title        = {Pith review of: Causal-Inspired Multi-Agent Decision-Making via Graph Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2R6QAXXW}},
  note         = {Machine review of arXiv:2507.23080}
}
read the original abstract

Since the advent of autonomous driving technology, it has experienced remarkable progress over the last decade. However, most existing research still struggles to address the challenges posed by environments where multiple vehicles have to interact seamlessly. This study aims to integrate causal learning with reinforcement learning-based methods by leveraging causal disentanglement representation learning (CDRL) to identify and extract causal features that influence optimal decision-making in autonomous vehicles. These features are then incorporated into graph neural network-based reinforcement learning algorithms to enhance decision-making in complex traffic scenarios. By using causal features as inputs, the proposed approach enables the optimization of vehicle behavior at an unsignalized intersection. Experimental results demonstrate that our proposed method achieves the highest average reward during training and our approach significantly outperforms other learning-based methods in several key metrics such as collision rate and average cumulative reward during testing. This study provides a promising direction for advancing multi-agent autonomous driving systems and make autonomous vehicles' navigation safer and more efficient in complex traffic environments.

Figures

Figures reproduced from arXiv: 2507.23080 by the authors.

Figure 1
Figure 1. an unsignalized intersection scenario. This study addresses the decision-making issue in au￾tonomous driving at an unsignalized intersection, as depicted in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The framework of CGRL algorithm. 2) Adjacency Matrix: The adjacency matrix is a square matrix that represents vehicle interactions. Each element eij in the matrix indicates whether a pair of vehicles i and j is adjacent. If vehicles i and j are within the specified range and interacting, eij = 1; otherwise, if there is no interaction or proximity, eij = 0. A =           e11 e12 · · · · · · e1n e21 e22 · · … view at source ↗
Figure 3
Figure 3. Training Rewards and Loss Curves Across Three Driving Tasks: Left Turn, Straight, and Right Turn. The reward and [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 33 canonical work pages

  1. [1]

    Graph neural networks and reinforcement learning: A survey,

    P. Adibi, B. Shoushtarian, and J. Chanussot, “Graph neural networks and reinforcement learning: A survey,” 2023

    work page 2023

  2. [3]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio’, and Y . Bengio, “Graph attention networks,” ArXiv, vol. abs/1710.10903, 2017

    arXiv 2017

  3. [4]

    Semi-supervised classification with graph convolutional networks,

    T. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” ArXiv, vol. abs/1609.02907, 2016

    arXiv 2016

  4. [5]

    Exploring Causal Learning through Graph Neural Networks: An In-depth Review

    S. Job, X. Tao, T. Cai, H. Xie, L. Li, J. Yong, and Q. Li, “Exploring causal learning through graph neural networks: An in-depth review,” ArXiv, vol. abs/2311.14994, 2023

    work page Pith review arXiv 2023

  5. [6]

    When graph neural network meets causality: Opportunities, methodologies and an outlook,

    W. Jiang, H. Liu, and H. Xiong, “When graph neural network meets causality: Opportunities, methodologies and an outlook,” 2023

    work page 2023

  6. [7]

    A survey on causal reinforcement learning,

    Y . Zeng, R. Cai, F. Sun, L. Huang, and Z. Hao, “A survey on causal reinforcement learning,” ArXiv, vol. abs/2302.05209, 2023

    arXiv 2023

  7. [8]

    Causal reinforcement learning: A survey,

    Z.-H. Deng, J. Jiang, G. Long, and C. Zhang, “Causal reinforcement learning: A survey,” Trans. Mach. Learn. Res. , vol. 2023, 2023

    work page 2023

  8. [9]

    Causal Multi-Agent Reinforcement Learning: Review and Open Problems

    S. J. Grimbly, J. Shock, and A. Pretorius, “Causal multi-agent reinforce- ment learning: Review and open problems,” ArXiv, vol. abs/2111.06721, 2021

    work page Pith review arXiv 2021

Show all 41 references
  1. [10]

    Variational graph auto-encoders,

    T. Kipf and M. Welling, “Variational graph auto-encoders,” ArXiv, vol. abs/1611.07308, 2016

    2016 arXiv

  2. [11]

    Challenges and opportunities in deep reinforcement learning with graph neural networks: A comprehensive review of algorithms and applications,

    S. Munikoti, D. Agarwal, L. Das, M. Halappanavar, and B. Natarajan, “Challenges and opportunities in deep reinforcement learning with graph neural networks: A comprehensive review of algorithms and applications,” IEEE Transactions on Neural Networks and Learning Systems, vol. ...

    2024

  3. [12]

    Graph convolution-based deep reinforcement learning for multi-agent decision-making in mixed traffic environments,

    Q. Liu, Z. Li, X. Li, J. Wu, and S. Yuan, “Graph convolution-based deep reinforcement learning for multi-agent decision-making in mixed traffic environments,” ArXiv, vol. abs/2201.12776, 2022

    2022 arXiv

  4. [13]

    Generalized single-vehicle- based graph reinforcement learning for decision-making in autonomous driving,

    F. Yang, X. Li, Q. Liu, Z. Li, and X. Gao, “Generalized single-vehicle- based graph reinforcement learning for decision-making in autonomous driving,” Sensors (Basel, Switzerland) , vol. 22, 2022

    2022

  5. [14]

    Multi-agent decision-making modes in uncertain interactive traffic scenarios via graph convolution-based deep reinforcement learning,

    X. Gao, X. Li, Q. Liu, Z.-H. Li, F. Yang, and T. Luan, “Multi-agent decision-making modes in uncertain interactive traffic scenarios via graph convolution-based deep reinforcement learning,” Sensors (Basel, Switzerland), vol. 22, 2022

    2022

  6. [15]

    Graph neural network and reinforcement learning for multi-agent cooperative control of con- nected autonomous vehicles,

    S. Chen, J. Dong, P. Y . J. Ha, Y . Li, and S. Labi, “Graph neural network and reinforcement learning for multi-agent cooperative control of con- nected autonomous vehicles,” Computer-Aided Civil and Infrastructure Engineering, vol. 36, pp. 838 – 857, 2021

    2021

  7. [16]

    Cooperative behavioral plan- ning for automated driving using graph neural networks,

    M. Klimke, B. V ¨olz, and M. Buchholz, “Cooperative behavioral plan- ning for automated driving using graph neural networks,” ArXiv, vol. abs/2202.11376, 2022

    2022 arXiv

  8. [17]

    Efficient connected and automated driving system with multi-agent graph rein- forcement learning,

    T. Shi, J. Wang, Y . Wu, L. Miranda-Moreno, and L. Sun, “Efficient connected and automated driving system with multi-agent graph rein- forcement learning,” 2020

    2020

  9. [18]

    Dq-gat: Towards safe and efficient autonomous driving with deep q-learning and graph attention networks,

    P. Cai, H. Wang, Y . Sun, and M. Liu, “Dq-gat: Towards safe and efficient autonomous driving with deep q-learning and graph attention networks,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, pp. 21 102–21 112, 2021

    2021

  10. [19]

    Drl-gat-sa: Deep reinforcement learning for autonomous driving planning based on graph attention networks and simplex architecture,

    Y . Peng, G. Tan, H. Si, and J. Li, “Drl-gat-sa: Deep reinforcement learning for autonomous driving planning based on graph attention networks and simplex architecture,” J. Syst. Archit., vol. 126, no. C, May

  11. [20]

    Causal based q-learning,

    A. M ´endez-Molina, I. Feliciano-Avelino, E. F. Morales, and L. E. Sucar, “Causal based q-learning,” Res. Comput. Sci. , vol. 149, pp. 95–104, 2020

    2020

  12. [21]

    Efficient reinforcement learning with prior causal knowledge,

    Y . Lu and A. Tewari, “Efficient reinforcement learning with prior causal knowledge,” in CLEaR, 2022

    2022

  13. [22]

    Counterfactual policy evaluation for decision- making in autonomous driving,

    P. Hart and A. Knoll, “Counterfactual policy evaluation for decision- making in autonomous driving,” arXiv: Learning, 2020

    2020

  14. [23]

    Causality-driven hierarchical struc- ture discovery for reinforcement learning,

    S. Peng, X. Hu, R. Zhang, K. Tang, J. Guo, Q. Yi, R. Chen, X. Zhang, Z. Du, L. Li, Q. Guo, and Y . Chen, “Causality-driven hierarchical struc- ture discovery for reinforcement learning,” ArXiv, vol. abs/2210.06964, 2022

    2022 arXiv

  15. [24]

    Constructing bayesian network models of gene expression networks from microarray data,

    P. Spirtes, C. Glymour, R. Scheines, S. A. Kauffman, V . Aimale, and F. C. Wimberly, “Constructing bayesian network models of gene expression networks from microarray data,” 2000

    2000

  16. [25]

    Multi-Channel Causal Variational Autoencoder,

    S. Al-Ali and I. Balelli, “Multi-Channel Causal Variational Autoencoder,” Aug. 2024, working paper or preprint. [Online]. Available: https://hal.science/hal-04666466

    2024

  17. [26]

    On causally disen- tangled representations,

    A. G. Reddy, B. G. L, and V . N. Balasubramanian, “On causally disen- tangled representations,” in AAAI Conference on Artificial Intelligence , 2021

    2021

  18. [27]

    Weakly supervised disentangled generative causal representation learning,

    X. Shen, F. Liu, H. Dong, Q. Lian, Z. Chen, and T. Zhang, “Weakly supervised disentangled generative causal representation learning,” J. Mach. Learn. Res. , vol. 23, pp. 241:1–241:55, 2020

    2020

  19. [28]

    Causal- vae: Disentangled representation learning via neural structural causal models,

    M. Yang, F. Liu, Z. Chen, X. Shen, J. Hao, and J. Wang, “Causal- vae: Disentangled representation learning via neural structural causal models,” 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9588–9597, 2020

    2021

  20. [29]

    Concept-free causal disentanglement with variational graph auto-encoder,

    J. Feng, L. Zhang, and L. Yang, “Concept-free causal disentanglement with variational graph auto-encoder,” ArXiv, vol. abs/2311.10638, 2023

    2023 arXiv

  21. [30]

    Cadet: A causal dis- entanglement approach for robust trajectory prediction in autonomous driving,

    M. Pourkeshavarz, J. Zhang, and A. Rasouli, “Cadet: A causal dis- entanglement approach for robust trajectory prediction in autonomous driving,” 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14 874–14 884, 2024

    2024

  22. [31]

    How attentive are graph attention networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” ArXiv, vol. abs/2105.14491, 2021

    2021 arXiv

  23. [32]

    Dueling network architectures for deep reinforcement learning,

    Z. Wang, T. Schaul, M. Hessel, H. Hasselt, M. Lanctot, and N. Freitas, “Dueling network architectures for deep reinforcement learning,” in Proceedings of The 33rd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. F. Balcan and K. Q...

    2016

  24. [33]

    Information flows in causal networks,

    N. Ay and D. Polani, “Information flows in causal networks,” Adv. Complex Syst., vol. 11, pp. 17–41, 2008

    2008

  25. [34]

    Ci-gnn: A granger causality-inspired graph neural network for interpretable brain network-based psychiatric diagnosis,

    K. Zheng, S. Yu, and B. Chen, “Ci-gnn: A granger causality-inspired graph neural network for interpretable brain network-based psychiatric diagnosis,” Neural networks : the official journal of the International Neural Network Society , vol. 172, p. 106147, 2023

    2023

  26. [35]

    Orphicx: A causality-inspired latent variable model for interpreting graph neural networks,

    W. Lin, H. Lan, H. Wang, and B. Li, “Orphicx: A causality-inspired latent variable model for interpreting graph neural networks,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13 719–13 728, 2022

    2022

  27. [36]

    Estimation of renyi entropy and mutual information based on generalized nearest-neighbor graphs,

    D. P ´al, B. P ´oczos, and C. Szepesvari, “Estimation of renyi entropy and mutual information based on generalized nearest-neighbor graphs,” in Neural Information Processing Systems , 2010

    2010

  28. [37]

    Mine: Mutual information neural estimation,

    I. Belghazi, S. Rajeswar, A. Baratin, R. D. Hjelm, and A. C. Courville, “Mine: Mutual information neural estimation,” ArXiv, vol. abs/1801.04062, 2018

    2018 arXiv

  29. [38]

    Measures of entropy from data using infinitely divisible kernels,

    L. G. S. Giraldo, M. Rao, and J. C. Pr ´ıncipe, “Measures of entropy from data using infinitely divisible kernels,” IEEE Transactions on Information Theory, vol. 61, pp. 535–548, 2012

    2012

  30. [39]

    An environment for autonomous driving decision-making,

    E. Leurent, “An environment for autonomous driving decision-making,” https://github.com/eleurent/highway-env, 2018

    2018

  31. [40]

    Congested traffic states in empirical observations and microscopic simulations,

    M. Treiber, A. Hennecke, and D. Helbing, “Congested traffic states in empirical observations and microscopic simulations,” Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics , vol. 62 2 Pt A, pp. 1805–24, 2000

    2000

  32. [41]

    Reasoning graph-based reinforce- ment learning to cooperate mixed connected and autonomous traffic at unsignalized intersections,

    D. Zhou, P. Hang, and J. Sun, “Reasoning graph-based reinforce- ment learning to cooperate mixed connected and autonomous traffic at unsignalized intersections,” Transportation Research Part C: Emerging Technologies, 2024

    2024

  33. [2022]

    Available: https://doi.org/10.1016/j.sysarc.2022.102505

    [Online]. Available: https://doi.org/10.1016/j.sysarc.2022.102505

    2022

Pith tools

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