Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Strategic Cyber Defense via Reinforcement Learning-Guided Combinatorial Auctions

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Neural auctions can plan cyber defenses before attacks arrive

desk verdict A plausible neural-auction application that overclaims cyber-defense viability because the promised downstream test is missing. read the letter →

arxiv 2509.10983 v1 pith:4UVMQWVS submitted 2025-09-13 cs.GT

Mai Pham , Vikrant Vaze , Peter Chin This is my paper
classification cs.GT
keywords cyberdefenseplanningcombinatorialauctionsreinforcementlearningQ-valuesmechanismdesignincentivecompatibilitydifferentiableoptimizationresourceallocation
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 using combinatorial auctions to allocate limited defensive actions, such as Analyze, Remove, or Restore, across network hosts before a cyber attack episode begins. Host valuations are derived from the Q-values of a reinforcement-learned defensive agent, letting the auction inherit long-term strategic foresight. The authors train CAFormer, a Transformer-based differentiable auction, to maximize revenue while approximately discouraging misreporting, and they compare it against oracle and greedy baselines in the DARPA CAGE Challenge 2 environment. The method achieves revenue competitive with an exact solver, near-zero regret, and allocations that preserve host-priority order under strategic lies. The authors argue this is the first neural-auction application to upstream cyber defense planning, offering a modular, interpretable, and decentralized alternative to centralized planning.

What carries the argument

The load-bearing mechanism is CAFormer, a permutation-equivariant Transformer-based neural auction trained with a combined revenue and regret loss to output probabilistic bundle allocations and normalized payments. Its inputs are private valuations v_i(S) built from RL Q-values: each action's Q-value is decomposed as Q_i(s,a) = V_i(s) + A_i(s,a), normalized to [0,1], and then a quadratic curvature term ±θ(|S|-1)^2·ε_i(S) is added to model bundle-level interactions. This valuation model converts a single centralized policy's critic values into host-specific combinatorial preferences, enabling the auction to allocate scarce defensive actions under feasibility and incentive constraints.

What would settle it

Run the same auction pipeline with Q-values replaced by (a) a second, independently trained defensive agent and (b) random valuations while keeping the curvature noise term unchanged. If the near-zero regret and the strong Blue-activity correlation (r=0.964) persist in case (b), the results are an artifact of the auction's generality rather than the RL signal; if case (a) fails, the Q-value source is the essential ingredient.

Watch

Extended reading notes

Core claim

The paper's central claim is that an auction mechanism can serve as an upstream planning layer for cyber defense: each host acts as a bidder with private valuations over bundles of defensive actions, and those valuations are extracted from the Q-values of a trained PPO agent operating in the CAGE 2 environment. These Q-values, normalized via min-max scaling, encode long-term expected utility and are augmented with structured quadratic curvature terms to model submodular (diminishing returns) or supermodular (synergistic) interactions among actions. CAFormer, trained end-to-end with revenue maximization and adversarial regret minimization, produces probabilistic allocations that are approxima

Load-bearing premise

The central premise is that Q-values from one centralized defensive agent faithfully represent each host's private, host-specific value for action bundles; if that proxy fails, the auction mechanism still trains correctly but the cyber-defense conclusions do not follow.

Editorial extensions

If this is right

  • If the Q-value proxy holds, the auction framework can pre-allocate defensive actions across heterogeneous hosts with revenue close to an exact oracle, making it a practical upstream planner for budget-constrained cyber defense.
  • The near-zero regret achieved under adversarial misreporting implies that self-interested hosts (or their learned proxies) have little to gain from lying, supporting the use of such mechanisms in decentralized or multi-stakeholder settings.
  • The preserved host-priority ordering under misreporting means the allocation is robust even when valuations are noisy or adaptive, a property not guaranteed by standard greedy heuristics.
  • The strong correlation between allocation scores and defender activity suggests the mechanism can act as an interpretable explanation tool, highlighting which hosts are operationally important before an episode unfolds.
  • The architecture is modular: swapping in alternative valuation sources (e.g., resilience scores) or changing the objective from revenue to social welfare or fairness would not require re-designing the core mechanism.

Reading between the lines

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

  • If the Q-value proxy is valid, the same pipeline could transfer to other simulated or real cyber environments that use an RL-trained defender, making the auction a drop-in strategic-planning layer; a testable prediction is that allocations would track whichever agent's critic is used, which a comparison of two independently trained defenders would reveal.
  • The curvature noise term ε_i(S) is a hand-chosen stochastic perturbation; if its magnitude (controlled by θ) dominates the Q-value signal, allocations would be driven by random noise, not strategic value—one can check this by computing the ratio of the curvature term to the Q-value spread and correlating allocations with Q-values alone.
  • The revenue objective, rather than social welfare, implicitly embeds an operational cost assumption; in practice this means the planner may favor aggressive, high-payment bundles even when a cheaper bundle would suffice, a behavior worth probing in budget-constrained settings where payments are virtual.
  • The reported Blue-activity correlation of r=0.964 is based on only a handful of hosts (six or so), so the strength of the alignment claim may be overstated; a larger network or repeated episodes would provide a more stable estimate.
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 an upstream cyber-defense planning layer for the DARPA CAGE Challenge 2 (CC2) environment in which host-specific valuations for bundles of defensive actions are constructed from Q-values of a trained PPO Blue agent and fed into CAFormer, a Transformer-based combinatorial auction trained with adversarial regret minimization. The experimental section reports Q-value distributions, training convergence of revenue and regret against an oracle-VCG baseline, allocation heatmaps under truthful and perturbed reports, and correlations between allocation scores and per-host Red/Blue action counts. The paper claims that this demonstrates the viability of auction-based planning in cyber defense and highlights interpretability benefits of RL-derived value structures.

Significance. The interdisciplinary combination is original: connecting neural combinatorial auctions to an upstream planning layer for a realistic cyber-defense simulator is a reasonable idea, and the auction-training component follows the well-established RegretNet/CAFormer recipe. The paper is honest about some limitations, especially in Section VII. However, the central viability claim is currently unsubstantiated: no experiment executes the auction allocations in CC2 or measures any downstream defense outcome. The statistical evidence for operational alignment is also problematic and partly circular. If the paper is reframed as a mechanism-level feasibility study and the missing downstream evaluation is added, it could make a useful contribution; as written, the main claim overreaches.

major comments (4)
  1. [§I, §VI] The introduction promises evaluation of 'downstream impact on cyber defense effectiveness,' but Section VI reports only auction revenue, regret, allocation heatmaps, and correlations. There is no experiment in which CAFormer allocations are executed in CC2, and no comparison of episode reward, compromised hosts, or operational resilience against a no-auction or heuristic baseline. Consequently, the abstract's claim that the results 'demonstrate the viability of auction-based planning in cyber defense' is unsupported even if all reported auction metrics are correct. Please either add a downstream CC2 evaluation or substantially narrow the claims to mechanism-level feasibility.
  2. [§VI-D, Table I] The correlation evidence is not reliable as reported. With 13 hosts, Pearson r=0.964 would have a two-tailed p-value far below 0.0361, and Spearman rho=0.800 with p=0.20 is similarly implausible; the table appears to mix one-tailed/two-tailed tests or incorrect sample sizes. More importantly, the 'alignment' finding is substantially by construction: the valuations are Q-values from the same CardiffUni PPO whose per-host Blue action counts are the correlation target. The allocation scores therefore inherit the policy's action preferences; this does not independently demonstrate alignment with operational priorities. Please report N, exact test variant, confidence intervals, and address the circularity explicitly.
  3. [§IV-B, §VII] The load-bearing modeling assumption is that centralized Q-values from a single PPO policy can serve as private, host-specific utilities over defensive action bundles. Section VII concedes that these are 'global return under a shared policy rather than private agent utilities,' which undercuts the private-values auction interpretation. In addition, the curvature term in Eq. (4) is not specified numerically: lambda, mu, gamma, and the epsilon_i(S) perturbation are never reported. Since epsilon_i(S) ~ U(0,1) and theta multiplies (|S|-1)^2, a large theta can dominate the Q-value signal and make the allocation nearly random. Please provide the actual hyperparameter values, a sensitivity analysis, or a validation of the Q-value proxy (e.g., decentralized Q-learning or cross-check against host-level outcomes).
  4. [§VI-C, Abstract] The abstract claims 'robustness to misreporting,' but the evidence in Section VI-C is a visual comparison of heatmaps. No quantitative stability metric is reported for the preservation of host ordering under truthful versus misreported valuations. This is a testable claim and is central to the incentive-compatibility narrative; please report, for example, rank correlation or allocation-distance statistics across multiple misreporting instances, with standard errors.
minor comments (6)
  1. [§IV-B, Figure 4] Spelling is inconsistent: 'Analyze' in the text but 'Analyse' in Figure 4 and in bundle names. Pick one convention.
  2. [§I, References] The introduction says auctions have been applied to 'mitigation of denial-of-service (DoS) attacks' and cites [7], but [7] is the cloud-resource paper; the DoS auction is [21]. Please correct the citation.
  3. [§VI-B, Figure 3] The caption states standard deviation <= 0.001, but no error bars or shaded regions are shown. Specify what is averaged and where the variation occurs.
  4. [§VI-A] The KS test result is reported without the sample size underlying it. Add the number of Q-value samples per host type, since with large samples even tiny KS statistics become significant.
  5. [§VI-D, Table I] The p-values are formatted inconsistently (0.0361 in the text vs 0.04 in the table) and the notation for Pearson r is missing in the table header. Please unify and include exact p-values.
  6. [§VI-C, Figure 4] The greedy baseline rows are said not to sum to one; this makes direct visual comparison with the learned allocations difficult. Consider presenting the greedy baseline as a normalized frequency matrix as well.

Circularity Check

1 steps flagged · score 6.0 of 10

The Blue-activity 'alignment' finding is largely inherited from the Q-value source by construction; the auction mechanism itself is not circular.

  1. self definitional [Section VI-D, Table I; Section IV-A; Section VII]
    "we analyze the correlation between the aggregated allocation scores and the number of Red and Blue actions each host receives during the simulation. ... we observe a strong positive correlation between allocation scores and the number of defensive actions by Blue in all hosts (Pearson r=0.964, p=0.0361), suggesting that the mechanism tends to focus attention on hosts where defenders are more active."

    The allocation scores are computed from valuations derived from Q_i(a) of the same CardiffUni PPO agent that generates the 'number of Blue actions' used as the correlation target. The agent's policy selects actions based on these Q-values/advantages, so hosts with more Blue activity are, by construction, hosts the agent values highly. The auction then allocates bundles using those same Q-values. Thus the strong Blue correlation is not an emergent property of the auction 'implicitly aligning with operational priorities'; it is a mechanical consequence of using one policy's Q-values as both the auction input and the benchmark for defensive activity. The paper concedes in Section VII that 'centralized Q-values reflect the global return under a shared policy rather than private agent utilities

full rationale

The core auction mechanism is self-contained: CAFormer is trained here on Q-derived valuations with its own revenue and regret objectives, and the comparisons to oracle-VCG and greedy heuristics are computed from the paper's own experiments. The citation to the authors' prior CAFormer paper [13] is not load-bearing because the training procedure is described in this paper rather than imported. The main circularity is in the Section VI-D interpretability result. The valuations fed to the auction and the 'Blue actions' counts used for correlation both originate from the same CardiffUni PPO policy/Q-function. A policy that acts according to Q-values will naturally concentrate actions on hosts with high Q-values, and an auction that allocates to high Q-valued bundles will therefore correlate with Blue activity by construction. That does not validate that the auction has independently discovered operational priorities; it inherits the policy's own preferences. This is a partial circularity in a headline supporting result. Separately, the paper promises evaluation of 'downstream impact on cyber defense effectiveness' (Section I) but Section VI reports no such experiment; allocations are never executed in CAGE Challenge 2 to measure episode reward, compromised hosts, or resilience against a no-auction baseline. This is a serious validity gap for the 'viability' claim, but it is a missing experiment rather than a circular derivation, so per the rules it does not itself raise the circularity score. Overall, the auction-design contribution is not circular, but one of the paper's central interpretive findings reduces, in part, to its own input source.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The paper's central claim rests on treating centralized RL Q-values as private host valuations and on hand-chosen hyperparameters (lambda, mu, gamma) plus a stochastic curvature perturbation that are never fitted or validated against any external outcome. The auction framework itself is inherited from prior work, so the contribution's validity hinges on whether the Q-value proxy and the curvature model are meaningful, which is asserted rather than demonstrated.

free parameters (5)
  • lambda (submodular curvature strength) = not reported
    Hyperparameter in v_i(S) = sum Q_i(a) - lambda*(|S|-1)^2*epsilon; hand-chosen, no value or sensitivity analysis given (Section IV-B).
  • mu (supermodular curvature strength) = not reported
    Hyperparameter for the synergistic valuation variant; no value or sensitivity analysis given (Section IV-B).
  • gamma (revenue/regret trade-off) = not reported
    Loss weight in L = (1-gamma)*log(1+L_revenue) + gamma*L_regret; hand-chosen, value not stated (Section V).
  • epsilon_i(S) random perturbation = U(0,1)
    Multiplicative noise on curvature; its scale relative to Q-values is unexamined, so allocations could be noise-dominated (Section IV-B).
  • min-max normalization epsilon = unnamed small constant
    Guard in X = (X - min)/(max - min + epsilon); harmless but part of the untested preprocessing chain (Section IV-A).
assumptions (5)
  • domain assumption Q-values of a centralized PPO policy approximate private per-host utilities over defensive action bundles
    Acknowledged in Section VII: 'centralized Q-values reflect the global return under a shared policy rather than private agent utilities.' This is the load-bearing premise connecting the auction machinery to cyber defense.
  • domain assumption CC2 Scenario 2 provides a faithful proxy for real enterprise cyber defense
    The environment is a simulation with scripted Red agents; the paper offers no argument that auction-planning results transfer to real networks (Sections IV, VII).
  • domain assumption Revenue maximization over reported Q-derived valuations is the right objective for cyber defense planning
    The revenue objective (Section V) stands in for operational costs; no evidence ties synthetic payments to real defense outcomes.
  • ad hoc to paper Quadratic curvature term approximates pairwise action interactions
    Section IV-B justifies (|S|-1)^2 by the O(|S|^2) growth of pairwise interactions; this is a heuristic surrogate chosen by the authors, not derived.
  • ad hoc to paper Min-max normalization preserves incentive properties
    Section IV-A asserts ordinal ranking and incentive properties survive uniform scaling; the claim that adversarial misreports constrained to [0,1]^d carry over from RegretNet [22] is inherited, not re-derived.
invented entities (1)
  • Hosts as self-interested combinatorial bidders with private valuations
    purpose: Enables the auction abstraction: each simulated machine reports values over bundles of defensive actions (Section III).
    Hosts are simulated endpoints, not strategic agents; the 'private valuations' are Q-values from one shared policy. The paper concedes this in Section VII. No real-world falsifiable handle is provided for the auction's allocations improving defense, and no downstream defense metric is measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Strategic Cyber Defense via Reinforcement Learning-Guided Combinatorial Auctions." pith.science (2026). https://pith.science/paper/4UVMQWVS

@misc{pith2026250910983,
  author       = {Pith},
  title        = {Pith review of: Strategic Cyber Defense via Reinforcement Learning-Guided Combinatorial Auctions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UVMQWVS}},
  note         = {Machine review of arXiv:2509.10983}
}
read the original abstract

Cyber defense operations increasingly require long-term strategic planning under uncertainty and resource constraints. We propose a new use of combinatorial auctions for allocating defensive action bundles in a realistic cyber environment, using host-specific valuations derived from reinforcement learning (RL) Q-values. These Q-values encode long-term expected utility, allowing upstream planning. We train CAFormer, a differentiable Transformer-based auction mechanism, to produce allocations that are approximately incentive-compatible under misreporting. Rather than benchmarking against existing agents, we explore the qualitative and strategic properties of the learned mechanisms. Compared to oracle and heuristic allocations, our method achieves competitive revenue while offering robustness to misreporting. In addition, we find that allocation patterns correlate with adversarial and defensive activity, suggesting implicit alignment with operational priorities. Our results demonstrate the viability of auction-based planning in cyber defense and highlight the interpretability benefits of RL-derived value structures.

Figures

Figures reproduced from arXiv: 2509.10983 by the authors.

Figure 1
Figure 1. Overview of our cyber planning framework. Hosts in the DARPA [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of Q-values for different host types. The [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Training curves showing average revenue and regret for CAFormer [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Allocation matrices under truthful vs. perturbed agent valuations. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 2 canonical work pages

  1. [1]

    Game-theoretic methods for robustness, security, and resilience of cyberphysical control systems: Games-in-games prin- ciple for optimal cross-layer resilient control systems,

    Q. Zhu and T. Basar, “Game-theoretic methods for robustness, security, and resilience of cyberphysical control systems: Games-in-games prin- ciple for optimal cross-layer resilient control systems,”IEEE Control Systems Magazine, vol. 35, no. 1, pp. 46–65, 2015

    2015

  2. [2]

    Don’t bury your head in warnings: A game-theoretic approach for intelligent allocation of cyber- security alerts,

    A. Schlenker, H. Xu, M. Guirguis, C. Kiekintveld, A. Sinha, M. Tambe, S. Sonya, D. Balderas, and N. Dunstatter, “Don’t bury your head in warnings: A game-theoretic approach for intelligent allocation of cyber- security alerts,” inProceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, 2017, pp. 381–387. [Online...

    doi:10.24963/ijcai.2017/54 2017

  3. [3]

    On autonomous agents in a cyber defence environment,

    M. Kiely, D. Bowman, M. Standen, and C. Moir, “On autonomous agents in a cyber defence environment,”arXiv preprint arXiv:2309.07388, 2023

    arXiv 2023

  4. [4]

    Comprehensive, multi-source cyber-security events data set,

    A. D. Kent, “Comprehensive, multi-source cyber-security events data set,” 05 2015. [Online]. Available: https://www.osti.gov/biblio/1179829

    arXiv 2015

  5. [5]

    An annotated review of past papers on attack graphs,

    R. Lippmann and K. Ingols, “An annotated review of past papers on attack graphs,”MIT Lincoln Laboratory Technical Report, 2005

    2005

  6. [6]

    A probabilistic computing approach of attack graph-based nodes in large-scale network,

    Y . Yun, X. Xi-shan, and Q. Zhi-chang, “A probabilistic computing approach of attack graph-based nodes in large-scale network,” Procedia Environmental Sciences, vol. 10, pp. 3–8, 2011, 2011 3rd International Conference on Environmental Science and Information Application Technology ESIAT 2011. [Online]. Available: https://www.sciencedirect.com/science/art...

    2011

  7. [7]

    An online auction mechanism for cloud computing resource allocation and pricing based on user evaluation and cost,

    J. Zhang, N. Xie, X. Zhang, and W. Li, “An online auction mechanism for cloud computing resource allocation and pricing based on user evaluation and cost,”Future Generation Computer Systems, vol. 89, pp. 286–299, 2018. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0167739X17317983

    2018

  8. [8]

    Alpcan and T

    T. Alpcan and T. Bas ¸ar,Network Security: A Decision and Game- Theoretic Approach. Cambridge University Press, 2010

    2010

Show all 28 references
  1. [9]

    Automated generation and analysis of attack graphs,

    O. Sheyner, J. Haines, S. Jha, R. Lippmann, and J. M. Wing, “Automated generation and analysis of attack graphs,” inProceedings 2002 IEEE Symposium on Security and Privacy. IEEE, 2002, pp. 273–284

    2002

  2. [10]

    Optimal network security hardening using attack graph games,

    K. Durkota, V . Lisy, B. Boˇsansky, and C. Kiekintveld, “Optimal network security hardening using attack graph games,” inProceedings of the 24th International Conference on Artificial Intelligence, ser. IJCAI’15. AAAI Press, 2015, p. 526–532

    2015

  3. [11]

    Deep reinforcement learning for cyber security,

    T. T. Nguyen and V . J. Reddi, “Deep reinforcement learning for cyber security,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 8, pp. 3779–3795, 2023

    2023

  4. [12]

    Deep learning algorithms for cyber security applications: A survey,

    G. Li, P. Sharma, L. Pan, S. Rajasegarar, C. Karmakar, and N. Patterson, “Deep learning algorithms for cyber security applications: A survey,” J. Comput. Secur., vol. 29, no. 5, p. 447–471, Jan. 2021. [Online]. Available: https://doi.org/10.3233/JCS-200095

    2021 doi

  5. [13]

    Advancing differentiable economics: A neural network framework for revenue-maximizing combinatorial auction mechanisms,

    M. Pham, V . Vaze, and P. Chin, “Advancing differentiable economics: A neural network framework for revenue-maximizing combinatorial auction mechanisms,” 2025. [Online]. Available: https://arxiv.org/abs/2501.19219

    2025 arXiv

  6. [14]

    Automated cyber defence: A review,

    S. Vyas, J. Hannay, A. Bolton, and P. Burnap, “Automated cyber defence: A review,”arXiv preprint arXiv:2303.04926, 2023

    2023 arXiv

  7. [15]

    Cyberbattlesim,

    M. D. R. Team., “Cyberbattlesim,” https://github.com/microsoft/cyberbattlesim, 2021

    2021

  8. [16]

    Cyborg: A gym for the development of autonomous cyber agents,

    M. Standen, M. Lucas, D. Bowman, T. J. Richer, J. Kim, and D. Marriott, “Cyborg: A gym for the development of autonomous cyber agents,” 2021

    2021

  9. [17]

    Network environment design for autonomous cyberdefense,

    A. Molina-Markham, C. Miniter, B. Powell, and A. Ridley, “Network environment design for autonomous cyberdefense,” arXiv preprint arXiv:2103.07583, 2021. [Online]. Available: https://arxiv.org/abs/2103.07583

    2021 arXiv

  10. [18]

    Eireland: Evaluating and interpreting reinforcement-learning-based network defenses,

    S. Cheung, J. Claypoole, P. Sharma, V . Yegneswaran, A. Gehani, S. Jha, J. Emanuello, and A. Ridley, “Eireland: Evaluating and interpreting reinforcement-learning-based network defenses,” https://sites.google.com/go.spawar.navy.mil/acd-2023/proceedings, 2023, accessed: 2025-06-23

    2023

  11. [19]

    Autonomous network defence using reinforcement learning,

    M. Foley, C. Hicks, K. Highnam, and V . Mavroudis, “Autonomous network defence using reinforcement learning,”Proceedings of the 2022 ACM Workshop on Artificial Intelligence and Security, pp. 79–90, 2022

    2022

  12. [20]

    Beyond cage: Investigating generalization of learned autonomous network defense policies,

    M. Wolk, A. Applebaum, C. Dennler, P. Dwyer, M. Moskowitz, H. Nguyen, N. Nichols, N. Park, P. Rachwalski, F. Rau, and A. Web- ster, “Beyond cage: Investigating generalization of learned autonomous network defense policies,”arXiv preprint arXiv:2211.15557, 2022

    2022 arXiv

  13. [21]

    An auction-based incentivized solution against ddos attacks,

    B. B. Gupta and A. Dahiya, “An auction-based incentivized solution against ddos attacks,”J. Inf. Technol. Res., vol. 14, no. 1, p. 1–19, Jan

  14. [22]

    Optimal auctions through deep learning,

    P. Duetting, Z. Feng, H. Narasimhan, D. Parkes, and S. Ravindranath, “Optimal auctions through deep learning,” inInternational Conference on Machine Learning (ICML), 2019, pp. 1706–1715

    2019

  15. [23]

    A permutation- equivariant neural network architecture for auction design,

    J. Rahme, S. Jelassi, J. Bruna, and S. M. Weinberg, “A permutation- equivariant neural network architecture for auction design,” 2021. [Online]. Available: https://arxiv.org/abs/2003.01497

    2021 arXiv

  16. [24]

    Optimal-er auctions through attention,

    V . Ivanov, R. Rahme, A. Bian, and D. C. Parkes, “Optimal-er auctions through attention,” inAdvances in Neural Information Processing Sys- tems (NeurIPS), 2022

    2022

  17. [25]

    A scalable neural network for dsic affine maximizer auction design,

    R. Duan, R. Rahme, and D. C. Parkes, “A scalable neural network for dsic affine maximizer auction design,”Journal of Artificial Intelligence Research (JAIR), vol. 79, pp. 543–576, 2024

    2024

  18. [26]

    Team cardiffuni’s cage-2 challenge submission,

    T. Cardiffuni, “Team cardiffuni’s cage-2 challenge submission,” https://github.com/johncardiff/-cyborg-cage-2, 2022, accessed: 2025-06- 23

    2022

  19. [27]

    Interpreting agent behaviors in reinforcement-learning- based cyber-battle simulation platforms,

    J. Claypoole, S. Cheung, A. Gehani, V . Yegneswaran, and A. Ridley, “Interpreting agent behaviors in reinforcement-learning- based cyber-battle simulation platforms,” 2025. [Online]. Available: https://arxiv.org/abs/2506.08192

    2025 arXiv

  20. [2021]

    Available: https://doi.org/10.4018/JITR.2021010101

    [Online]. Available: https://doi.org/10.4018/JITR.2021010101

    doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.