Pith. sign in

REVIEW 3 major objections 6 minor 30 references

HGFormer: A Hierarchical Graph Transformer Framework for Two-Stage Colonel Blotto Games via Reinforcement Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read HGformer, a hierarchical graph Transformer with a pointer-network Planner and a dual-scale Transfer agent, outperforms MILP-based baselines in two-stage Colonel Blotto while keeping inference near 0.2 seconds.

desk verdict A reimplementable hierarchical RL framework for graph Blotto that roughly matches MILP at ~0.1 s latency, with an overstated LFRT gain and evaluation against a single fixed Blue policy. read the letter →

arxiv 2506.08580 v1 pith:KMJHLBF3 submitted 2025-06-10 cs.AI

Yang Lv , Jinlong Lei , Peng Yi This is my paper
classification cs.AI
keywords ColonelBlottogamehierarchicalreinforcementlearninggraphtransformeradversarialresourceallocationtwo-stagedynamicpointernetworkattention
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 argues that a two-stage dynamic Colonel Blotto game on a graph—where one player first observes the opponent's deployment, then both players reallocate resources over repeated rounds—can be handled by a hierarchical reinforcement learning architecture instead of per-round exact optimization. The paper proposes HGformer, in which a Planner agent produces the initial allocation with a pointer-network decoder and a Transfer agent chooses round-by-round moves with a dual-scale encoder, both built on a graph Transformer that injects shortest-path structure and a global virtual node. It claims that a layered feedback training scheme, which feeds the Transfer agent's long-term returns back into the Planner, closes the coordination gap between the two stages. If the claim holds, large-scale adversarial allocation problems that make mixed-integer linear programming impractical can be solved in near-real time with comparable or better payoff.

What carries the argument

The machinery is the HGformer architecture plus its three-phase training. The Enhanced Graph Transformer Encoder (EGTE) augments multi-head self-attention with a learnable shortest-path-distance bias per node pair and adds a virtual node that aggregates global state, so each node embedding carries both topology and game dynamics. The Planner decoder is a pointer network with multi-head cross-attention: it picks one node at a time, updates the remaining budget, and uses the previously chosen node's embedding as context, which makes the initial allocation sequential and budget-aware. The Transfer decoder combines EGTE's global embeddings with a GATv2 local module and then produces proportional transfer flows by attention over neighbors, merging long-range strategy with local mobility. Layered Feedback Reinforced Training (LFRT) first pretrains the Planner with REINFORCE, then trains the Transfer agent with PPO, then refines the Planner with an augmented return $R_0 + \lambda U_r$ that includes the Transfer agent's realised cumulative payoff, which is the mechanism that coordinates the two decision stages.

What would settle it

Re-run the same end-to-end evaluation with Blue's dynamic policy replaced by an optimized or learned adaptive policy, or with graph instances from a different generative family with explicit edge weights and round limits; if HGformer's Red utility no longer exceeds the MILP-based baselines by more than the reported error bars, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own experiments, HGformer's central claim is that the learned hierarchical policies dominate both optimization and learning baselines on graphs of 10 to 70 nodes. At $N=70$, the full framework with feedback attains Red utility $395.71 \pm 19.35$ and transfer cost $70.15 \pm 9.35$, against $359.02 \pm 27.62$ and $87.79$ for the two-stage MILP baseline and $364.51 \pm 28.53$ and $86.35$ for MILP allocation combined with the learned Transfer agent, while keeping inference around $0.16$ seconds where MILP needs hundreds of seconds per round. The margin over MILP-based methods widens as the graph grows, and the gap is small or overlapping at $N=10$. The paper attributes this result to the combination of structural bias in the encoder, the sequential nature of the Planner, the local-global fusion in the Transfer agent, and the cross-stage feedback in training.

Load-bearing premise

The load-bearing premise is that every comparison is run against a single fixed rule-based Blue policy operating on self-generated random graphs whose topology, edge weights, and round limit are not disclosed, so the reported payoff margins could shrink or invert if Blue were adaptive or the environment family changed.

Editorial extensions

If this is right

  • HGformer sustains sub-second inference at 70 nodes where the full MILP baseline exceeds 500 seconds per episode, making real-time reallocation feasible.
  • The feedback mechanism (LFRT), compared with the same architecture without feedback, cuts transfer cost by about 10% at $N=50$ and $N=70$ and improves cumulative payoff on large graphs.
  • In the dynamic-transfer stage, the dual-scale encoder outperforms global-only and local-only ablations by 4–7% utility, so both long-range and neighborhood information contribute to the result.
  • The sequential pointer decoder beats one-shot allocation and vanilla Transformer/GCN encoders in the Planner ablation, indicating that inter-decision dependencies matter for initial deployment.
  • At small graph sizes the learned planner matches MILP-level utility, so the speed gain does not come at the cost of accuracy where the exact solver is tractable.

Reading between the lines

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

  • Because the reported margins are measured against a single rule-based Blue policy, a natural next test is whether HGformer's advantage survives an adaptive or optimized Blue; the paper lists this as future work, and the margins at intermediate graph sizes are small enough that the ordering could change.
  • The architecture's components are not specific to Colonel Blotto, so the same Planner–Transfer loop with SPD-biased Transformer encoding and feedback training could be applied to other sequential allocation problems on networks, such as security resource placement or disaster response.
  • The paper does not disclose the random-graph generation model (topology, edge weights, round limit), so an immediate extension is to benchmark on standard graph families such as scale-free or community-structured graphs; the claimed topology generalisation is currently supported only by the paper's own generated instances.
  • The lower transfer cost achieved with feedback suggests the training implicitly budgets movement; measuring the sparsity or total edge-distance traveled by the learned transfer matrices against MILP could reveal the mechanism.
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

3 major / 6 minor

Summary. The paper proposes HGformer, a hierarchical graph Transformer framework for a two-stage dynamic Colonel Blotto game on graphs. Red first allocates resources after observing Blue's initial deployment (Planner Agent, a pointer-network decoder), then repeatedly reallocates resources over rounds (Transfer Agent, a dual-scale EGTE+GAT decoder). The two levels are trained with a three-phase scheme (LFRT): REINFORCE pretraining of the Planner, PPO pretraining of the Transfer agent, then feedback of the Transfer agent's cumulative utility into the Planner's return. Experiments on synthetic graphs with N=10-70 compare HGformer against rule-based, simulated annealing, MILP, GPN-PPO, and several ablations, reporting Red utility, transfer cost, and inference latency. The paper reports superior payoff and much lower latency than MILP-based baselines, attributing the gains to the EGTE encoder, the hierarchical architecture, and LFRT.

Significance. If the empirical claims held, this would be a useful contribution to hierarchical reinforcement learning for graph-constrained adversarial resource allocation. The paper's strengths include a clear problem formulation, a detailed architectural description, and a broad ablation set spanning encoder choice, decoder choice, local/global features, and cross-stage feedback. The speed advantage over per-round MILP (milliseconds versus hours) is substantial and practically relevant. However, the central empirical claim is currently supported only against a single fixed rule-based Blue policy in an incompletely specified synthetic environment, and at least one headline quantitative claim in Section IV.D is contradicted by the paper's own table. These issues are load-bearing for the stated conclusion of 'superior overall performance in complex dynamic game scenarios.'

major comments (3)
  1. [Section IV.D, Table III] The statement that 'HGformer with LFRT surpasses its no-feedback counterpart by more than 8% in cumulative payoff' is not supported by the reported numbers. From Table III, the LFRT versus no-LFRT differences are 0.91 at N=40 (216.36±11.03 vs 215.45±15.96), 5.48 at N=50 (271.36±18.85 vs 265.88±19.62), 15.29 at N=60 (348.25±15.32 vs 332.96±21.63), and 6.71 at N=70 (395.71±19.35 vs 389.00±25.12). These are 0.4% to 4.6% relative to the no-feedback value, not more than 8%, and the standard deviations overlap at every size. The only >8% margins in Table III are against DMILP, not against the no-feedback variant. Please correct this claim and report a significance test or a proper effect-size statistic for the LFRT comparison.
  2. [Section IV.A.1 and IV.A.2] The test environment is underspecified and the opponent is fixed. The paper does not disclose the graph topology model (e.g., Erdős–Rényi, Barabási–Albert, grid), the edge-weight distribution, the round limit, or the tie-breaking rule; only node values and resource budgets are given. In addition, Blue is a single rule-based policy (proportional initial allocation and recapture of high-value neighbors), and Section V explicitly lists adaptive Blue as future work. The reported margins over DMILP and MILP-T are therefore estimates of Red's payoff against one predictable opponent in one unverifiable environment family, not a demonstration of 'superior overall performance in complex dynamic game scenarios.' Please disclose the full generator specification and evaluate against at least one adaptive or alternative Blue policy (e.g., a PPO-trained Blue or a second rule variant), reporting per-graph win rates or payoff distributions rather than only mean utility.
  3. [Tables II and III, MILP rows for N≥40] The text states that per-round MILP 'prevents full-episode completion' when N>30, yet Tables II and III report MILP cumulative utilities for N=40, 50, 60, and 70. It is unclear how these values were obtained: partial episodes, a time limit with a truncated return, a warm start, or some other protocol. Without this information the MILP baselines are not commensurable with the learning-based methods that complete full episodes. Relatedly, the 'upper-bound reference' label for DMILP in Section IV.D is misleading because per-round MILP optimizes immediate reward greedily and is not a full-horizon upper bound. Please specify the exact protocol used for the large-N MILP rows, or remove those rows and clearly state that full-episode MILP is intractable.
minor comments (6)
  1. [Section II.B, Eq. (4) and Eq. (6)] Eq. (4) writes a sum over (i,i) ∈ E, which should be (i,j) ∈ E, and Eq. (6) uses I(sb_i(T) ≥ sb_i(T)) where the right-hand side should be sr_i(T).
  2. [Algorithm 1, line 5] The function names ReluPlanner and ReluTransfer are unexplained; these appear to be typographical variants of 'rule-based' or 'Blue' policy names and should be defined or corrected.
  3. [Section III.D, Phase C] The notation Ur(θ_T) is ambiguous: the dynamic return depends on the Planner's allocation as well as the Transfer policy, so it should be written as a function of the full rollout, e.g., Ur(a0, θ_T) or Ur(θ_P, θ_T), with the dependence on the Planner parameters made explicit.
  4. [Figure 3 caption] The caption refers to a 'GTE' variant for the single-shot decoder, while the text and Table I call this variant GAE-P; please use a consistent name.
  5. [References] References [5] and [11] are the same paper, and references [12] and [13] are also the same paper; please consolidate the duplicates.
  6. [Table III caption] The table header uses the abbreviations HGform(θP) and HGform(θP) with tilde/hat accents that are easy to confuse. Please spell out the full names (HGformer without LFRT and HGformer with LFRT) in the caption or legend.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained and evaluation anchors are external.

full rationale

The paper's central claim is empirical: HGformer, combining the EGTE encoder, a pointer-network Planner, a dual-scale Transfer agent, and the LFRT training scheme, achieves higher payoff and lower transfer cost than MILP, simulated annealing, greedy heuristics, and GPN-PPO on a graph-based two-stage Colonel Blotto game. The evaluation anchors are external to the paper: per-round MILP, simulated annealing, greedy allocation, and the published GPN-PPO baseline [32], plus internal ablations (GNN-P, GTE-P, GAE-P, Local-only, Global-only). No load-bearing step reduces to a self-citation; the reference list contains no works by the present authors, and the cited components (Pointer Networks [26], Transformers [27], GATv2 [28], PPO [30]) are standard, externally developed methods. The only circularity-adjacent element is that the LFRT Phase C objective J'(theta_P) = E[R_P0 + lambda U_r(theta_hat_T)] includes the same aggregate utility U_r that Table III reports as the evaluation metric. This is ordinary hierarchical credit assignment rather than a definitional identity: the Planner must still learn an allocation that, in combination with the fixed Transfer policy, achieves high U_r on unseen random graphs, and the feedback weight lambda = 0.5 makes the training objective a weighted proxy rather than an exact copy of the reported utility. The modest margins over MILP-based baselines, overlapping error bars at some graph sizes, the single fixed rule-based Blue policy, and the undisclosed graph generator are external-validity and robustness concerns that the paper itself acknowledges in Section V as future work; they do not constitute circularity under the definition used here. Therefore no circular step can be exhibited, and the score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The ledger captures what the empirical claim pulls in beyond the paper's own contribution. The architecture adds no physical postulates, but the claims rest on the game rules (winner-take-all, linear transfer costs), on a single scripted Blue opponent, on a self-generated and partially undisclosed graph environment, on the sufficiency of a four-dimensional node feature, and on hand-set reward-shaping weights (alpha_s=0.1, lambda=0.5). The largest burden is the fixed-Blue assumption: every reported utility is a best response to one rule-based policy, not a verified margin against adaptive opponents.

free parameters (5)
  • reward-shaping weight alpha_s = 0.1
    Hand-set weight in the Transfer reward RT_t (Section III.D, Phase B) balancing control-score change against resource-efficiency terms; no sensitivity analysis is provided.
  • LFRT feedback weight lambda = 0.5
    Hand-set coefficient in Phase C coupling the Planner's return to the Transfer agent's average utility; no ablation over lambda is reported.
  • Planner per-node allocation rule = exactly Blue's node amount s^b_j(0)
    Eq. (22) subtracts s^b_j(0) from the remaining budget, 'typically set to the minimum required to overcome Blue's defense'; this fixes Red's commitment sizes and leaves only node ordering as the Planner's decision.
  • environment round limit and graph-generation parameters = not stated
    Section IV.A.1 gives node values from U[0,1], Rr=0.5Rb, and N=10-70, but the graph topology model, edge weights, number of game rounds, and termination condition are unspecified, making the exact environment unreproducible.
  • architecture hyperparameters = L=3, H=4, d=32, d_ff=64, lr 1e-3 and 5e-4
    Section IV.A.2 fixes encoder depth, attention heads, embedding dimension, FFN dimension, and learning rates; standard design choices not subjected to sensitivity analysis.
assumptions (6)
  • domain assumption Node-level winner-take-all conflict: the side with fewer resources at a node loses all resources there, with ties favoring Red (Section II.A).
    Defines the game being solved; all empirical claims are relative to this resolution rule.
  • domain assumption Linear transfer-cost model: moving resources along edge (i,j) costs w_ij times the amount moved (Eq. 3).
    The dissipation model determines the cost term in the utility; alternative cost models would change optimal strategies.
  • domain assumption Blue is a fixed rule-based policy with no learning or adaptation (Section IV.A.2).
    All reported Red utilities are against this single opponent; the generality of the main claim depends on it, as Section V concedes.
  • domain assumption The node feature triplet [v_i, s^r_i(t), s^b_i(t)] plus degree is a sufficient Markov state for the decision process (Eq. 7).
    The policy is a function of these features only; any information needed for optimal play must be representable in this vector.
  • domain assumption Self-generated random graphs with uniform node values represent 'complex dynamic game scenarios' (Section IV.A.1).
    The abstract's generality claim rests on this representativeness assumption; graph-generation parameters are undisclosed.
  • standard math Standard RL estimation assumptions: PPO clipped surrogate and GAE provide usable policy-gradient estimates (Section III.D).
    Background methodology from refs. [29] and [30], used without modification.
invented entities (2)
  • Virtual node v0, an SPD-distance-1 aggregator token
    purpose: Global aggregation token that participates in every attention layer and outputs the whole-graph embedding h^L_v0 used by both decoders (Section III.A.4, Algorithm 2).
    Standard CLS-style mechanism borrowed from BERT [25]; it is an architectural device with no external falsifiable handle, but it also makes no physical claim, so it is not a graviton-problem entity.
  • Learned global query vector h_va for GAT aggregation
    purpose: Aggregates a global view over the GATv2 local embeddings via weighted summation (Eq. 26), combined with h_prev and W_step as pointer-decoder context (Eqs. 14, 17).
    Trainable parameters whose values are learned from data; no independent evidence exists outside the fitted policies, which is expected for network weights rather than physical postulates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HGFormer: A Hierarchical Graph Transformer Framework for Two-Stage Colonel Blotto Games via Reinforcement Learning." pith.science (2026). https://pith.science/paper/KMJHLBF3

@misc{pith2026250608580,
  author       = {Pith},
  title        = {Pith review of: HGFormer: A Hierarchical Graph Transformer Framework for Two-Stage Colonel Blotto Games via Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KMJHLBF3}},
  note         = {Machine review of arXiv:2506.08580}
}
read the original abstract

Two-stage Colonel Blotto game represents a typical adversarial resource allocation problem, in which two opposing agents sequentially allocate resources in a network topology across two phases: an initial resource deployment followed by multiple rounds of dynamic reallocation adjustments. The sequential dependency between game stages and the complex constraints imposed by the graph topology make it difficult for traditional approaches to attain a globally optimal strategy. To address these challenges, we propose a hierarchical graph Transformer framework called HGformer. By incorporating an enhanced graph Transformer encoder with structural biases and a two-agent hierarchical decision model, our approach enables efficient policy generation in large-scale adversarial environments. Moreover, we design a layer-by-layer feedback reinforcement learning algorithm that feeds the long-term returns from lower-level decisions back into the optimization of the higher-level strategy, thus bridging the coordination gap between the two decision-making stages. Experimental results demonstrate that, compared to existing hierarchical decision-making or graph neural network methods, HGformer significantly improves resource allocation efficiency and adversarial payoff, achieving superior overall performance in complex dynamic game scenarios.

Figures

Figures reproduced from arXiv: 2506.08580 by the authors.

Figure 1
Figure 1. Illustration of the two-stage Colonel Blotto game. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. illustrates the detailed flow of the node resource allocation modules in the HGformer framework. From top to bottom, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Training convergence curves for Planner ablation experiments: (a) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [1]

    Event-based secure leader-following consensus control for multiagent systems with multiple cyber attacks,

    J. Liu, T. Yin, D. Yue, H. R. Karimi, and J. Cao, “Event-based secure leader-following consensus control for multiagent systems with multiple cyber attacks,” IEEE Transactions on Cybernetics , vol. 51, no. 1, pp. 162–173, 2020

    work page 2020

  2. [2]

    Multi-attribute game theoretic model for resource allocation in military attack-defense application,

    X. Zhang, B. Ge, and Y . Tan, “Multi-attribute game theoretic model for resource allocation in military attack-defense application,” Guofang Keji Daxue Xuebao/Journal of National University of Defense Technology , 2018

    work page 2018

  3. [3]

    A game theoretic model for resource allocation among countermeasures with multiple attributes,

    E. C. Paulson, I. Linkov, and J. M. Keisler, “A game theoretic model for resource allocation among countermeasures with multiple attributes,” European Journal of Operational Research, vol. 252, no. 2, pp. 610–622, 2016

    work page 2016

  4. [4]

    N-dimensional blotto game with heterogeneous battlefield values,

    C. Thomas, “N-dimensional blotto game with heterogeneous battlefield values,” Economic Theory, vol. 65, no. 3, pp. 509–544, 2018

    work page 2018

  5. [6]

    Online learning in budget-constrained dynamic colonel blotto games,

    V . Leon and S. R. Etesami, “Online learning in budget-constrained dynamic colonel blotto games,” Dynamic Games and Applications , vol. 14, no. 4, pp. 865–887, 2024

    work page 2024

  6. [7]

    Finite-time dynamic alloca- tion and control in multiagent coordination for target tracking,

    X. Zhao, Q. Zong, B. Tian, and M. You, “Finite-time dynamic alloca- tion and control in multiagent coordination for target tracking,” IEEE transactions on cybernetics , vol. 52, no. 3, pp. 1872–1880, 2020

    work page 2020

  7. [8]

    A three- stage colonel blotto game with applications to cyberphysical security,

    A. Gupta, G. Schwartz, C. Langbort, S. S. Sastry, and T. Ba ˇrar, “A three- stage colonel blotto game with applications to cyberphysical security,” in 2014 American Control Conference , pp. 3820–3825, IEEE, 2014

    work page 2014

  8. [9]

    Power allocation in multi- user cellular networks: Deep reinforcement learning approaches,

    F. Meng, P. Chen, L. Wu, and J. Cheng, “Power allocation in multi- user cellular networks: Deep reinforcement learning approaches,” IEEE Transactions on Wireless Communications , vol. 19, no. 10, pp. 6255– 6267, 2020

    work page 2020

Show all 30 references
  1. [10]

    Graph reinforcement learning with relational priors for predictive power allocation,

    J. Zhao and C. Yang, “Graph reinforcement learning with relational priors for predictive power allocation,” Science China Information Sciences, vol. 68, no. 2, pp. 1–18, 2025

    2025

  2. [11]

    Equilibrium ap- proximating and online learning for anti-jamming game of satellite communication power allocation,

    M. Zou, J. Chen, J. Luo, Z. Hu, and S. Chen, “Equilibrium ap- proximating and online learning for anti-jamming game of satellite communication power allocation,” Electronics, vol. 11, no. 21, p. 3526, 2022

    2022

  3. [13]

    Colonel blotto games in network systems: models, strategies, and applications,

    S. Guan, J. Wang, H. Yao, C. Jiang, Z. Han, and Y . Ren, “Colonel blotto games in network systems: models, strategies, and applications,” IEEE Transactions on Network Science and Engineering , vol. 7, no. 2, pp. 637–649, 2019

    2019

  4. [14]

    Resource allocation problems,

    N. Katoh, A. Shioura, and T. Ibaraki, “Resource allocation problems,” Handbook of combinatorial optimization , pp. 1–93, 2024

    2024

  5. [15]

    A dynamic resource allocation strategy for cloud-native applications leveraging markov properties,

    Z. Shen, Y . Ma, and J. Shen, “A dynamic resource allocation strategy for cloud-native applications leveraging markov properties,” International Journal of Advance in Applied Science Research , vol. 3, pp. 99–107, 2024

    2024

  6. [16]

    Cooperative particle swarm opti- mization with a bilevel resource allocation mechanism for large-scale dynamic optimization,

    X.-F. Liu, J. Zhang, and J. Wang, “Cooperative particle swarm opti- mization with a bilevel resource allocation mechanism for large-scale dynamic optimization,” IEEE Transactions on Cybernetics , vol. 53, no. 2, pp. 1000–1011, 2022

    2022

  7. [17]

    Discrete colonel blotto and general lotto games,

    S. Hart, “Discrete colonel blotto and general lotto games,” International Journal of Game Theory , vol. 36, no. 3–4, pp. 441–460, 2008

    2008

  8. [18]

    Generalizations of the colonel blotto and general lotto games,

    D. Kovenock and B. Roberson, “Generalizations of the colonel blotto and general lotto games,” in The Oxford Handbook of the Economics of Peace and Conflict, Oxford University Press, 2012

    2012

  9. [19]

    Double oracle algorithm for game-theoretic robot allocation on graphs,

    M. An and E. Zhou, “Double oracle algorithm for game-theoretic robot allocation on graphs,” in Proc. NeurIPS Workshop on Graphs and Optimization, 2023

    2023

  10. [20]

    Unsupervised resource alloca- tion with graph neural networks,

    M. Cranmer, P. Melchior, and B. Nord, “Unsupervised resource alloca- tion with graph neural networks,” in NeurIPS 2020 Workshop on Pre- registration in Machine Learning , pp. 272–284, PMLR, 2021

    2020

  11. [21]

    Semi-supervised graph structure learning via dual reinforcement of label and prior structure,

    R. Yuan, Y . Tang, Y . Wu, J. Niu, and W. Zhang, “Semi-supervised graph structure learning via dual reinforcement of label and prior structure,” IEEE Transactions on Cybernetics , 2024

    2024

  12. [22]

    A three- stage colonel blotto game with applications to cyber-physical security,

    A. Gupta, G. Schwartz, C. Langbort, S. Sastry, and T. Bas ¸ar, “A three- stage colonel blotto game with applications to cyber-physical security,” in Proceedings of the American Control Conference (ACC) , pp. 3820– 3825, 2014

    2014

  13. [23]

    A blotto game with incomplete information,

    T. Adamo and A. Matros, “A blotto game with incomplete information,” Economics Letters, vol. 105, no. 1, pp. 100–102, 2009

    2009

  14. [24]

    Reinforcement learning agents in colonel blotto,

    J. C. G. Noel, “Reinforcement learning agents in colonel blotto,” arXiv preprint arXiv:2204.02785, 2022

    2022 arXiv

  15. [25]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo...

    2019

  16. [26]

    Pointer networks,

    O. Vinyals, M. Fortunato, and N. Jaitly, “Pointer networks,” Advances in neural information processing systems , vol. 28, 2015

    2015

  17. [27]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  18. [28]

    How attentive are graph attention networks?,

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?,” arXiv preprint arXiv:2105.14491 , 2021

    2021 arXiv

  19. [29]

    Sample efficient reinforcement learning with reinforce,

    J. Zhang, J. Kim, B. O’Donoghue, and S. Boyd, “Sample efficient reinforcement learning with reinforce,” in Proceedings of the AAAI conference on artificial intelligence , vol. 35, pp. 10887–10895, 2021

    2021

  20. [30]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  21. [31]

    Traveling-salesman-problem algorithm based on simulated annealing and gene-expression programming,

    A.-H. Zhou, L.-P. Zhu, B. Hu, S. Deng, Y . Song, H. Qiu, and S. Pan, “Traveling-salesman-problem algorithm based on simulated annealing and gene-expression programming,” Information, vol. 10, no. 1, p. 7, 2018

    2018

  22. [32]

    Graph pointer neural networks,

    T. Yang, Y . Wang, Z. Yue, Y . Yang, Y . Tong, and J. Bai, “Graph pointer neural networks,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, pp. 8832–8839, 2022

    2022

Pith tools

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