Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Two-Player Zero-Sum Differential Games with One-Sided Information

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

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

Pith's one-line read This paper shows that two-player zero-sum differential games with one-sided information can be solved at a cost set by the number of payoff types, not the size of the action space.

desk verdict A serious but incomplete algorithm paper: the O(I^2) reformulation of the Bellman backup is genuinely useful, but the convergence claim rests on an unproven global-optimality assumption for nonconvex-nonconcave minimax subproblems. read the letter →

arxiv 2502.05314 v2 pith:TUIGOIAS submitted 2025-02-07 cs.GT

Mukesh Ghimire , Zhe Xu , Yi Ren This is my paper · ORCID
classification cs.GT MSC 91A2349N70
keywords differentialgamesone-sidedinformationzero-sumcontinuousactionspacesmixed-strategyequilibriumconvexificationdualgamehoming
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that two-player zero-sum differential games with one-sided information can be solved without discretizing a continuous action space, at a cost that depends on the number of payoff types I rather than the size of the action space |A|. The key structural facts are that the informed player's equilibrium mixing is I-atomic, the uninformed player's is (I+1)-atomic, and the two players' equilibrium strategies can be computed separately through primal and dual game formulations. If true, this makes a class of continuous-control attack-defense and sports scenarios tractable for equilibrium computation, in contrast to standard no-regret solvers whose complexity grows with |A|. The authors instantiate the resulting CAMS solver on a homing game and report that it approximates the analytic equilibrium where CFR+ and related baselines scale poorly.

What carries the argument

Three pieces carry the argument. The convexification operator $V_{\mathrm{ex}}$ replaces the value function by its envelope in the public belief $p$; this is what forces P1's strategy to be $I$-atomic and P2's strategy to be $(I+1)$-atomic, turning the $O(|\mathcal{A}|)$ Bellman backup into the $O(I^2)$ minimax problems labeled (P1) and (P2). The Isaacs' condition, $\min_u \max_v H = \max_v \min_u H$, guarantees that complete-information versions of these differential games have pure equilibria, so all nontrivial mixing in the incomplete-information game is devoted to belief manipulation rather than to action uncertainty. Finally, the primal and dual backward inductions, Eq. (4)-(5) for P1 and Eq. (7)-(8) for P2, are connected by the subgradient condition $\hat{p} \in \partial_p V(t_0, x_0, p)$, which justifies solving the two players separately. DS-GDA is the numerical engine that solves each nonconvex-nonconcave minimax subproblem in $O(\varepsilon^{-4})$ iterations.

What would settle it

Run CAMS on a longer-horizon Hexner homing game with an analytic equilibrium, increasing the number of collocation points N and decreasing the time step tau: if the expected action error stops shrinking as N grows or grows as the horizon lengthens, the assumed convergence of the value-model pipeline is false. Equivalently, at collocation points check whether the primal and dual value functions satisfy the Fenchel identity $V(t,x,p) = \max_{\hat{p}}(p \cdot \hat{p} - V^*(t,x,\hat{p}))$; a persistent mismatch would show that the two decoupled solvers are not computing the same game value.

Watch

Extended reading notes

Core claim

The central discovery is that the convexification structure of incomplete-information games, together with the Isaacs' condition, compresses the Bellman backup at every infostate to a minimax problem over O($I^{2}$) decision variables instead of O(|A|). At any belief p, P1 needs to randomize over at most I actions, one per possible payoff type, and P2 over at most I+1 dual actions; each mixed action is tied to a posterior belief, which is exactly the splitting mechanism of Aumann and Maschler. Because information is one-sided, P1's value can be computed by backward induction on the primal game while P2's is computed on a dual game obtained through the Fenchel conjugate, and the opponent in each formulation can be treated as playing a pure best response. This decoupling avoids the recurrent learning dynamics of joint strategy updates. The authors package these insights into the CAMS algorithm, solve the per-infostate nonconvex-nonconcave minimax problems with DS-GDA, and validate the approach on Hexner's homing game.

Load-bearing premise

The whole pipeline presumes that fitting the value model to finitely many collocation points at each time step yields an accurate approximation of the true value, and that these approximation errors do not build up as the solver moves backward through time; the paper assumes this convergence rather than proving it.

Editorial extensions

If this is right

  • Continuous and even infinite action spaces can be handled without discretization, because the per-infostate backup has $O(I^2)$ decision variables regardless of $|\mathcal{A}|$.
  • P1's and P2's equilibrium strategies can be solved as two separate computations, each against a pure best-responding opponent, removing the need for joint last-iterate learning dynamics.
  • Per-infostate accuracy enters polynomially as $O(\varepsilon^{-4})$ through DS-GDA, so improving the equilibrium approximation does not reintroduce dependence on the action-space size.
  • On the Hexner homing game, CAMS approximates the analytic equilibrium while CFR+, MMD, and DeepCFR require discretized action sets and their costs grow with the discretization size.

Reading between the lines

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

  • An implication the paper leaves implicit is that the $I$-atomic support bound should transfer to any one-sided-information game whose value is convexified in the public belief, such as pursuit-evasion with several possible targets, so the same $O(I^2)$ backup would apply with the type set reinterpreted as the target set.
  • A testable variant would replace the collocation-plus-regression value model with a representation carrying certified Lipschitz bounds, turning the assumed convergence of CAMS into a quantifiable error-propagation statement; nothing in the paper's formulation prevents such a certified implementation.
  • The primal/dual decoupling suggests that in practice P1's and P2's value functions could be learned by two independent regression tasks on the same state-time-belief grid, parallelizing equilibrium computation across the two players, a possibility the paper does not discuss.
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 studies two-player zero-sum differential games with one-sided information and continuous action spaces. It argues that the equilibrium value can be computed through per-infostate Bellman backups whose variables scale with the number of payoff types I rather than the action-space size |A|, and that the primal and dual formulations allow the two players' equilibrium strategies to be computed separately. The authors propose CAMS, a solver that combines DS-GDA for the per-infostate minimax problems with value-function fitting over collocation points, and they validate it on Hexner's homing game against CFR+, MMD, CFR-BR-Primal, and DeepCFR. The main claims are that CAMS achieves computational complexity independent of |A| and that it is the first algorithm with scalable convergence to equilibrium for this class of games.

Significance. If the convergence claim were established, the paper would address a real bottleneck: existing no-regret and extensive-form solvers scale with the size of a discretized action set, which is untenable for continuous-action differential games. The paper's use of known results from differential game theory (Cardaliaguet 2007, 2009; Aumann, Maschler, and Stearns 1995) is appropriate, and the empirical comparison against the analytical Hexner solution is a sensible benchmark. The paper also provides a public code link. However, the central theoretical gap, namely the absence of a convergence proof for the CAMS value iteration with function approximation and for the global optimality of the per-infostate minimax solves, means that the main advertised results are currently unsupported. The paper is best viewed as a promising algorithmic proposal with a suggestive empirical demonstration rather than a proven scalable-convergence result.

major comments (3)
  1. [Methods, Eqs. (P1) and (P2)] The central convergence claim is not established. The per-infostate problem (P1) is explicitly nonconvex-nonconcave, and the paper invokes DS-GDA (Zheng et al. 2023) with worst-case O(epsilon^-4) complexity. However, DS-GDA as cited provides convergence to stationary points, not to the global minimax value required by Eq. (5)/(4) for the convexified value. The Isaacs condition (3) only ensures a pure Nash equilibrium in the complete-information inner game; it does not make the split-lottery parameterization in (P1) convex-concave or exclude spurious stationary points. Because the fitted value model and the backward sweep are built from these per-infostate values, a local stationary point at one stage can propagate errors through all earlier stages. The paper supplies no global-optimality proof for P1/P2 and no experiments with multiple restarts to rule out initialization dependence. This gap is load-bearing for the abstract claim of scalable convergence.
  2. [Algorithm 1 and the Game solver section] There is no convergence proof for the value iteration with function approximation. The algorithm solves P1 at N collocation points, fits a model \hat{V}_t to the collected data, and iterates backward. No error bound relates N, the function class, or the fitting error to the error in \hat{V}_t, and no accumulation bound over the T/tau steps is given. The empirical section reports only a 1-stage and a 4-stage Hexner game (Fig. 3) and does not vary N or tau to demonstrate convergence to the analytical value. Therefore the Conclusion's statement that CAMS is the first method to provide a tractable solution, and the abstract's phrase scalable convergence, are unsupported.
  3. [Abstract and Methods (Reformulation)] The claimed computational complexity independent of |A| does not follow from the statements in the paper. Replacing an O(|A|) convex-concave problem with an O(I^2)-variable nonconvex-nonconcave problem does not by itself guarantee tractability: global solutions of nonconvex-nonconcave minimax problems generally require more than stationarity, and DS-GDA's O(epsilon^-4) bound is for finding stationary points. The decision variables u_k and v_k remain continuous and nonconvex even though the number of atoms is small. To support the complexity claim, the paper must either prove global optimality for P1/P2 under the Isaacs condition or substantially weaken the claim to a per-step stationarity guarantee.
minor comments (6)
  1. [Eq. (4)] The convexification operator V_ex (also written Vexp) is used without a formal definition; please define it and cite the exact theorem from Cardaliaguet (2009) that justifies the backward induction.
  2. [Algorithm 1] The algorithm's dataset D is initialized once and accumulates data from all time steps, but the text defines a per-stage dataset D_t; it is unclear whether the model is fit only to the current stage or to all stages. Please clarify the fitting procedure.
  3. [Introduction, Figure 1] The text has an orphaned fragment immediately before Figure 1, apparently a layout artifact; this should be fixed.
  4. [Behavioral strategy of P1] The phrase Cav u theorem should be typeset as Cav(u) theorem, the standard notation in the repeated-games literature.
  5. [References] Brown et al. 2020a and 2020b are duplicate entries with the same title and venue; they should be merged.
  6. [Empirical Validation] The experimental section omits details of the value-function model, its training loss, optimizer, hyperparameters, and the initialization of DS-GDA; including these is necessary for reproducibility even with the code link.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the core derivation chain rests on external theorems and algebraic reformulations, not on fitted inputs or self-citations.

full rationale

The paper's load-bearing derivation chain starts from externally sourced results: the atomicity/convexification of P1 and P2 strategies and the Cav u splitting mechanism are attributed to Aumann, Maschler, and Stearns (1995), De Meyer (1996), and Cardaliaguet (2007, 2009). The backward-induction equations (4) and (7) are presented as quoted from Cardaliaguet, and the per-infostate programs P1 and P2 are algebraic reparameterizations of equations (5) and (8). For example, the omitted constraint sum_k lambda_k p_k = p follows automatically from lambda_k = sum_i alpha_ki p[i], p_k[i] = alpha_ki p[i]/lambda_k, and sum_k alpha_ki = 1, so P1 is equivalent to the splitting equation rather than a renamed version of it. The claimed independence from |A| follows from the O(I^2)-sized parameterization, not from any fitted quantity, and the dual-game decoupling is imported from Cardaliaguet (2007), an independent external theorem. Empirically, the algorithm is evaluated against Hexner's analytical solution, an external ground truth, so the reported action errors are not predictions of fitted values. The only self-citations (Ghimire et al. 2024) appear in the introduction and related work and are not load-bearing; no central premise depends on them. The main gap is algorithmic convergence: DS-GDA (Zheng et al. 2023) gives stationary-point guarantees for the nonconvex-nonconcave program but not a guarantee of the global convexified value, and the fitted value model has no error-propagation bound. That is a correctness/completeness limitation, not a circular step: the paper does not define the value as whatever the solver returns, and it does not fit parameters to the benchmark it later predicts.

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

The paper introduces no new free parameters or entities. It restates standard assumptions and theorems from differential game theory. The algorithmic implementation does involve choices like the number of collocation points and DS-GDA parameters, but these are not part of the theoretical claim and are not reported in sufficient detail to count as free parameters.

assumptions (4)
  • domain assumption Isaacs' condition holds for the Hamiltonian (Eq. 3).
    The paper relies on this to guarantee pure NE in complete-information subgames and to justify the atomicity of mixed strategies. It is Assumption 4 in the preliminaries.
  • domain assumption The five Cardaliaguet assumptions hold: compact action sets, Lipschitz dynamics, bounded payoffs, full knowledge, and perfect recall.
    These are listed as Assumptions 1-5 and are used to invoke the value existence and atomicity results from Cardaliaguet (2009).
  • standard math The behavioral strategy of P1 is I-atomic and convexifies the value (Cav u theorem).
    This is a theorem from Aumann-Maschler-Stearns and De Meyer, used as the basis for reformulating the Bellman backup into the finite minimax problem (P1).
  • standard math The dual game value V* is equivalent to the primal value at subgradient points, and P2's equilibrium in the dual game transfers to the primal game.
    This is a theorem from Cardaliaguet (2007), used to derive the P2 reformulation (P2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Two-Player Zero-Sum Differential Games with One-Sided Information." pith.science (2026). https://pith.science/paper/TUIGOIAS

@misc{pith2026250205314,
  author       = {Pith},
  title        = {Pith review of: Two-Player Zero-Sum Differential Games with One-Sided Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TUIGOIAS}},
  note         = {Machine review of arXiv:2502.05314}
}
abstract

Unlike Poker where the action space $\mathcal{A}$ is discrete, differential games in the physical world often have continuous action spaces not amenable to discrete abstraction, rendering no-regret algorithms with $\mathcal{O}(|\mathcal{A}|)$ complexity not scalable. To address this challenge within the scope of two-player zero-sum (2p0s) games with one-sided information, we show that (1) a computational complexity independent of $|\mathcal{A}|$ can be achieved by exploiting the convexification property of incomplete-information games and the Isaacs' condition that commonly holds for dynamical systems, and that (2) the computation of the two equilibrium strategies can be decoupled under one-sidedness of information. Leveraging these insights, we develop an algorithm that successfully approximates the optimal strategy in a homing game. Code available in https://github.com/ghimiremukesh/cams/tree/workshop

Figures

Figures reproduced from arXiv: 2502.05314 by the authors.

Figure 1
Figure 1. SOTA algo￾rithms like CFR require ex￾panding over entire action space (left), whereas our al￾gorithm only requires ex￾panding over at most I ac￾tions for P1 (I+1 for P2) at each decision node (right). ng insights: (1) At any infostate, P1’s (resp. P2’s) behavioral strategy is I (resp. I + 1)-atomic and convexifies the primal (resp. dual) value with respect to the public belief ( [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 3
Figure 3. Comparisons b/w CAMS and baseline algorithms. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Trajectories using strategies from CAMS and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    L.; and Hazan, E

    Abernethy, J.; Bartlett, P. L.; and Hazan, E. 2011. Blackwell approachability and no-regret learning are equivalent. In Proceedings of the 24th Annual Conference on Learning Theory, 27--46. JMLR Workshop and Conference Proceedings

    work page 2011

  4. [4]

    J.; Maschler, M.; and Stearns, R

    Aumann, R. J.; Maschler, M.; and Stearns, R. E. 1995. Repeated games with incomplete information. MIT press

    work page 1995

  5. [5]

    Billings, D.; Burch, N.; Davidson, A.; Holte, R.; Schaeffer, J.; Schauenberg, T.; and Szafron, D. 2003. Approximating game-theoretic optimal strategies for full-scale poker. In IJCAI, volume 3, 661

    work page 2003

  6. [6]

    Blackwell, D. 1956. An analog of the minimax theorem for vector payoffs

    work page 1956

  7. [7]

    Brown, N.; Bakhtin, A.; Lerer, A.; and Gong, Q. 2020 a . Combining deep reinforcement learning and search for imperfect-information games. Advances in Neural Information Processing Systems, 33: 17057--17069

    work page 2020

  8. [8]

    Brown, N.; Bakhtin, A.; Lerer, A.; and Gong, Q. 2020 b . Combining deep reinforcement learning and search for imperfect-information games. Advances in Neural Information Processing Systems, 33: 17057--17069

    work page 2020

Show all 39 references
  1. [9]

    Brown, N.; Lerer, A.; Gross, S.; and Sandholm, T. 2019. Deep counterfactual regret minimization. In International conference on machine learning, 793--802. PMLR

    2019

  2. [10]

    Brown, N.; and Sandholm, T. 2019. Superhuman AI for multiplayer poker. Science, 365(6456): 885--890

    2019

  3. [11]

    Burch, N.; Johanson, M.; and Bowling, M. 2014. Solving imperfect information games using decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 28

    2014

  4. [12]

    Cardaliaguet, P. 2007. Differential games with asymmetric information. SIAM journal on Control and Optimization, 46(3): 816--838

    2007

  5. [13]

    Cardaliaguet, P. 2009. Numerical approximation and optimal strategies for differential games with lack of information on one side. Advances in Dynamic Games and Their Applications: Analytical and Numerical Developments, 1--18

    2009

  6. [14]

    Cen, S.; Wei, Y.; and Chi, Y. 2021. Fast policy extragradient methods for competitive games with entropy regularization. Advances in Neural Information Processing Systems, 34: 27952--27964

    2021

  7. [15]

    De Meyer, B. 1996. Repeated games, duality and the central limit theorem. Mathematics of Operations Research, 21(1): 237--251

    1996

  8. [16]

    FAIR†, M. F. A. R. D. T.; Bakhtin, A.; Brown, N.; Dinan, E.; Farina, G.; Flaherty, C.; Fried, D.; Goff, A.; Gray, J.; Hu, H.; et al. 2022. Human-level play in the game of Diplomacy by combining language models with strategic reasoning. Science, 378(6624): 1067--1074

    2022

  9. [17]

    Ghimire, M.; Zhang, L.; Xu, Z.; and Ren, Y. 2024. State-Constrained Zero-Sum Differential Games with One-Sided Information. In Salakhutdinov, R.; Kolter, Z.; Heller, K.; Weller, A.; Oliver, N.; Scarlett, J.; and Berkenkamp, F., eds., Proceedings of the 41st International Confe...

    2024

  10. [18]

    Gilpin, A.; Hoda, S.; Pena, J.; and Sandholm, T. 2007. Gradient-based algorithms for finding Nash equilibria in extensive form games. In Internet and Network Economics: Third International Workshop, WINE 2007, San Diego, CA, USA, December 12-14, 2007. Proceedings 3, 57--69. Springer

    2007

  11. [19]

    Gilpin, A.; and Sandholm, T. 2006. Finding equilibria in large sequential games of imperfect information. In Proceedings of the 7th ACM conference on Electronic commerce, 160--169

    2006

  12. [20]

    Bayesian

    Harsanyi, J. C. 1967. Games with incomplete information played by “Bayesian” players, I--III Part I. The basic model. Management science, 14(3): 159--182

    1967

  13. [21]

    Hexner, G. 1979. A differential game of incomplete information. Journal of Optimization Theory and Applications, 28: 213--232

    1979

  14. [22]

    Johanson, M.; Bard, N.; Burch, N.; and Bowling, M. 2012. Finding optimal abstract strategies in extensive-form games. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 26, 1371--1379

    2012

  15. [23]

    Koller, D.; and Megiddo, N. 1992. The complexity of two-person zero-sum games in extensive form. Games and economic behavior, 4(4): 528--552

    1992

  16. [24]

    Lanctot, M.; Waugh, K.; Zinkevich, M.; and Bowling, M. 2009. Monte Carlo sampling for regret minimization in extensive games. Advances in neural information processing systems, 22

    2009

  17. [25]

    Lanctot, M.; Zambaldi, V.; Gruslys, A.; Lazaridou, A.; Tuyls, K.; P \'e rolat, J.; Silver, D.; and Graepel, T. 2017. A unified game-theoretic approach to multiagent reinforcement learning. Advances in neural information processing systems, 30

    2017

  18. [26]

    McMahan, B. 2011. Follow-the-Regularized-Leader and Mirror Descent: Equivalence Theorems and L1 Regularization. In Gordon, G.; Dunson, D.; and Dudík, M., eds., Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 of Procee...

    2011

  19. [27]

    Morav c \' k, M.; Schmid, M.; Burch, N.; Lis \`y , V.; Morrill, D.; Bard, N.; Davis, T.; Waugh, K.; Johanson, M.; and Bowling, M. 2017. Deepstack: Expert-level artificial intelligence in heads-up no-limit poker. Science, 356(6337): 508--513

    2017

  20. [28]

    T.; Burch, N.; Anthony, T.; et al

    Perolat, J.; De Vylder, B.; Hennes, D.; Tarassov, E.; Strub, F.; de Boer, V.; Muller, P.; Connor, J. T.; Burch, N.; Anthony, T.; et al. 2022. Mastering the game of Stratego with model-free multiagent reinforcement learning. Science, 378(6623): 990--996

    2022

  21. [29]

    Perolat, J.; Munos, R.; Lespiau, J.-B.; Omidshafiei, S.; Rowland, M.; Ortega, P.; Burch, N.; Anthony, T.; Balduzzi, D.; De Vylder, B.; et al. 2021. From poincar \'e recurrence to convergence in imperfect information games: Finding equilibrium via regularization. In Internation...

    2021

  22. [30]

    Sandholm, T. 2010. The state of solving large incomplete-information games, and application to poker. Ai Magazine, 31(4): 13--32

    2010

  23. [31]

    Z.; et al

    Schmid, M.; Morav c \' k, M.; Burch, N.; Kadlec, R.; Davidson, J.; Waugh, K.; Bard, N.; Timbers, F.; Lanctot, M.; Holland, G. Z.; et al. 2023. Student of Games: A unified learning algorithm for both perfect and imperfect information games. Science Advances, 9(46): eadg3256

    2023

  24. [32]

    Silver, D.; Hubert, T.; Schrittwieser, J.; Antonoglou, I.; Lai, M.; Guez, A.; Lanctot, M.; Sifre, L.; Kumaran, D.; Graepel, T.; et al. 2017 a . Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815

    2017 arXiv

  25. [33]

    Silver, D.; Schrittwieser, J.; Simonyan, K.; Antonoglou, I.; Huang, A.; Guez, A.; Hubert, T.; Baker, L.; Lai, M.; Bolton, A.; et al. 2017 b . Mastering the game of go without human knowledge. nature, 550(7676): 354--359

    2017

  26. [34]

    Z.; Loizou, N.; Lanctot, M.; Mitliagkas, I.; Brown, N.; and Kroer, C

    Sokota, S.; D'Orazio, R.; Kolter, J. Z.; Loizou, N.; Lanctot, M.; Mitliagkas, I.; Brown, N.; and Kroer, C. 2022. A unified approach to reinforcement learning, quantal response equilibria, and two-player zero-sum games. arXiv preprint arXiv:2206.05825

    2022 arXiv

  27. [35]

    Tammelin, O. 2014. Solving large imperfect information games using CFR+. arXiv preprint arXiv:1407.5042

    2014 arXiv

  28. [36]

    Vieillard, N.; Kozuno, T.; Scherrer, B.; Pietquin, O.; Munos, R.; and Geist, M. 2020. Leverage the average: an analysis of kl regularization in reinforcement learning. Advances in Neural Information Processing Systems, 33: 12163--12174

    2020

  29. [37]

    K.; Piccinini, F.; et al

    Wang, Z.; Veli c kovi \'c , P.; Hennes, D.; Toma s ev, N.; Prince, L.; Kaisers, M.; Bachrach, Y.; Elie, R.; Wenliang, L. K.; Piccinini, F.; et al. 2024. TacticAI: an AI assistant for football tactics. Nature communications, 15(1): 1906

    2024

  30. [38]

    M.-C.; Blanchet, J.; and Li, J

    Zheng, T.; Zhu, L.; So, A. M.-C.; Blanchet, J.; and Li, J. 2023. Universal gradient descent ascent method for nonconvex-nonconcave minimax optimization. Advances in Neural Information Processing Systems, 36: 54075--54110

    2023

  31. [39]

    Zinkevich, M.; Johanson, M.; Bowling, M.; and Piccione, C. 2007. Regret minimization in games with incomplete information. Advances in neural information processing systems, 20

    2007

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.