Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Online Competitive Information Gathering for Partially Observable Trajectory Games

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that online trajectory plans can actively gather information by solving a finite-history partially observable stochastic game with particle-based gradient play.

desk verdict A useful online planner for active information gathering in trajectory games, but the core reduction is not proved for the implemented policy class and the statistics are overstated. read the letter →

arxiv 2506.01927 v1 pith:6XR3SOVR submitted 2025-06-02 cs.GT cs.AIcs.MAcs.RO

Mel Krusniak , Hang Xu , Parker Palermo , Forrest Laine This is my paper
classification cs.GTcs.AIcs.MAcs.RO
keywords partiallyobservablestochasticgamesmodelpredictivegameplayactiveinformationgatheringgradientparticlebeliefrepresentationpursuit-evasionimperfectmulti-agentplanning
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 tackles the problem of planning trajectories for several interacting robots that do not fully observe one another, where a player can profit by deliberately seeking information about an opponent. It proposes a finite-history, finite-horizon partially observable stochastic game, in which each policy maps recent observation histories to actions, and solves it online with particle-based estimates of the joint distribution over states and observation histories plus stochastic gradient play. The central empirical claim is that this active information-gathering variant lowers a player's cost relative to a passive variant that ignores future observations, in continuous pursuit-evasion and warehouse-pickup scenarios; the one tested exception is the evader in the hide-and-seek variant. If correct, the method gives a route to information-aware game-theoretic planning from scratch at run time, without offline training or layered beliefs about other agents' beliefs.

What carries the argument

The central object is the finite-history/horizon partially observable stochastic game in Eq. 1, where each player minimizes $\mathbb{E}[c^{(i)}(x)]$ under the joint measure $p(x,z\mid \bar{x}_{\bar t}, z_{[\bar t]}) \, \bar q_{\bar t}(\bar{x}_{\bar t}, z_{[\bar t]})$ and policies map observation histories to actions. The machinery that makes it tractable is particle-based stochastic gradient play: a large set of weighted particles $(X,Z,w)$ approximates the joint distribution, a small batch is rolled out over $T_{\mathrm{future}}$ steps to produce Monte Carlo cost estimates, and automatic differentiation drives gradient-based updates on the policy networks. A distribution update steps particles forward with the equilibrium policies and reweights a fraction $\gamma$ of them by the true observation likelihood, balancing closed-loop responsiveness against accurate opponent modeling.

What would settle it

For a small finite POSG (for example, a two-state, two-action tag game), compute the equilibrium of the per-history objective (Eq. 3) and of the unconditioned joint objective (Eq. 1) exactly and compare the resulting policies; any difference, or any configuration in which the shared-network active player does not beat a passive player, would falsify the claimed reduction.

Watch

Extended reading notes

Core claim

The paper's central claim is that rational trajectory plans in imperfect-information games can be computed online by solving a game in which every player minimizes expected cost over the unconditioned joint distribution of states and observation histories (Eq. 1), rather than conditioning on its own private history (Eq. 3). The reduction from the conditioned objectives to the joint objective rests on the argument that per-history positive scaling factors do not change the policy parameters at equilibrium. The solver represents the joint belief as weighted particles, rolls out each particle forward with the current policies, and updates policy parameters by gradient descent on the estimated cost; experiments show that this active planning lowers cost against passive play in all tested configurations except the hide-and-seek evader, which the authors attribute to the limited history length.

Load-bearing premise

The load-bearing premise is that dropping the conditioning on each player's private observation history yields the correct game even though the implemented policy is a single shared neural network whose parameters couple all histories; if that reduction fails for this policy class, the game actually optimized is not the belief-conditioned game the authors intend.

Editorial extensions

If this is right

  • In pursuit-evasion and warehouse-pickup tests, planning that accounts for future observations lowers cost relative to open-loop planning, with several differences statistically significant at $p<0.05$.
  • Because policies are history-to-action maps, the planner produces recognizable active information gathering, such as an evader briefly turning to check one starting location before committing.
  • The framework extends to $N>2$ general-sum games (chain tag) and to obstacle-heavy environments (hide and seek), though the hide-and-seek evader shows the finite-history assumption can be a liability.
  • Running several equilibrium solves and keeping particles under each candidate policy improves agreement between agents' models, reducing surprisal about the opponent's true position.
  • Measured per-gradient-step times of roughly 0.06–0.22 s in the tested scenarios indicate that real-time execution is plausible with code and hardware improvements.

Reading between the lines

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

  • The Eq. 3-to-Eq. 1 reduction is proved for per-history independent policy parameters, but the implementation uses one shared network; whether the reduction holds for this policy class is an open testable question, and the measured gains may partly come from the $\gamma$-reweighting rather than from equilibrium play.
  • The method's reliance on pure strategies and gradient play suggests a natural extension to deliberately mixed strategies, which the paper notes could add convergence guarantees and richer deceptive behavior.
  • The hide-and-seek evader failure points to a concrete remedy: increasing $T_{\mathrm{past}}$ or using a recurrent policy that compresses longer histories should restore the advantage in memory-dependent games.
  • Treating $\gamma$ as a tunable knob rather than a fixed constant could let an agent interpolate between faithful opponent modeling and closed-loop use of its own observations, which may matter for deployment with unreliable particle coverage.
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 / 5 minor

Summary. The paper proposes a finite-history/finite-horizon formulation of partially observable stochastic games for online trajectory planning, and an approximation method based on particle-filtered joint distributions and stochastic gradient play. The authors test the method on continuous pursuit-evasion and warehouse-pickup scenarios, comparing 'active' information-gathering policies against 'passive' versions of the same solver, and report improvements in most configurations. The central theoretical step is the reduction in Section III-C from per-history conditional objectives (Eq. 3) to an unconditioned joint objective (Eq. 1), which the paper justifies by arguing that per-history positive scaling factors do not affect equilibrium policy parameters. The experimental claims rest on Table I and Section VI-A, where active information gathering is said to reduce costs significantly in several cases.

Significance. If the formulation and the Eq. 3-to-Eq. 1 reduction were valid, the paper would provide a practically oriented bridge between extensive-form/POSG reasoning and online model-predictive game play, and the particle-based solver with a shared neural policy would be a useful step toward continuous-space competitive information gathering. The paper is clearly written, the algorithms are specified in enough detail to be reimplemented, and the active-versus-passive ablation is a reasonable first test of information-gathering behavior. However, the theoretical reduction is not established for the neural policy class actually used, and the statistical evidence for the headline improvement is weaker than the text claims. As a heuristic planner with a corrected theoretical framing and honest statistics, the method could still be of interest, but the current manuscript does not support its game-theoretic claims.

major comments (3)
  1. [III-C, Eq. (3) to Eq. (1)] The reduction is invalid for the policy class used in the implementation. The text states that positive factors f(zhat^(i)) do not affect player i's policy parameters at equilibrium 'because player i selects unique actions for all zhat^(i)'. This is correct only if each observation history has its own independently parameterized policy. Section IV-D states that policies are feedforward neural networks with parameters shared across all histories. With a shared theta, the Eq. (1) objective is a weighted average sum_h p(h) J_h(theta) over histories, and the weights p(h) are not irrelevant constants: they determine how much each information set contributes to the gradient of the shared parameters. A stationary point of this weighted average need not be a Nash equilibrium of the per-history game in Eq. (3), and the action selected for the agent's realized history can be suboptimal because it was optimized against all histories simultaneously. Algorithm 3's gamma-conditioning, which mixes true observations into a fraction of particles, changes the optimized objective even further. The paper provides no proof that the reduction survives for the neural policy class, so the claim that the solver computes rational plans in the intended POSG is unsupported.
  2. [VI-A, Table I] The claim that several cost differences, 'particularly in TAG', are statistically significant at p<0.05 is not supported by the reported standard errors. In TAG, the active-versus-passive pursuer comparison is 9.97 ± 1.19 versus 12.58 ± 1.44; the difference of 2.61 has approximate standard error sqrt(1.19^2 + 1.44^2) = 1.87, giving p ≈ 0.16. The active-evader comparison (13.97 ± 1.92 versus 12.58 ± 1.44) has a difference of 1.39 with approximate standard error sqrt(1.92^2 + 1.44^2) = 2.40, giving p ≈ 0.56. Even the WAREHOUSE comparison is only one cell, and no test procedure, sample-size adjustment, or multiple-comparison correction is described. The evidence for active information gathering is therefore much weaker than the prose suggests.
  3. [VI-A, WAREHOUSE and Section IV-D] The evaluation compares only the proposed active solver against a passive version of the same solver, which is an appropriate ablation for isolating the effect of information gathering. However, the abstract's phrase 'outperforming passive competitors' and the Section VI-A discussion should be qualified: these are not competitors from the literature but ablations of the same method. In addition, the active/passive distinction for the warehouse P1 agent is not meaningful because P1 has perfect self-observations and no opponent observations; the table's 'Active P1' column is confusing, and the n/a cells should be explained or omitted.
minor comments (5)
  1. [III-B and III-C] The integration domains in Eq. (1) and Eq. (3) are inconsistent: Eq. (1) uses z ∈ Z^{Tpast+1+Tfuture} while Eq. (3) uses z ∈ Z^{1+Tfuture}; the latter appears to omit the past-history length Tpast.
  2. [Algorithm 1] The pseudocode samples zt+1 from the observation model but then writes at1 ← πθ(o[t+1]); the symbol o is not defined and should presumably be the sampled observation zt+1.
  3. [VI-A] The experiments terminate Algorithm 2 after 100 gradient iterations rather than checking convergence, and the text notes that the active version typically takes more iterations to converge. Please report whether the passive version had converged at 100 iterations and provide learning curves or final-equilibrium costs for both modes, since the fixed-iteration comparison may reflect convergence speed rather than steady-state performance.
  4. [VI-A, Table I] The table layout is hard to read: the 'Passive P1 / Active P1' header spans only the warehouse block, the active/active warehouse cell is n/a without explanation, and the standard errors are presented as ± values without specifying the test used to obtain p-values.
  5. [VII-B] The explanation for the hide-and-seek evader's cost increase, that 'optimal evader strategy is too complex to be captured with only Tpast past observations', is plausible but not tested; an ablation varying Tpast in that scenario would make the claim concrete.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular reduction found; Eq. 1 is not defined from the measured outcomes, and the two self-citations are not load-bearing. The main risk is an unmet precondition in the Eq. 3-to-Eq. 1 reduction for shared neural-network policies, which is a derivation gap rather than circularity.

full rationale

The paper's claimed derivation chain is Section III-C: per-history conditional objectives (Eq. 3) are replaced by the unconditioned joint objective (Eq. 1) because, for a fixed observation history zhat^(i), any positive factor f(zhat^(i)) scales only the terms controlled by that history's policy parameters and so cannot change that block's argmin. This is a legitimate equivalence under the stated block-independent parameterization, not a definitional identity between the two equations. The implementation in Section IV-D instead uses a single feedforward neural network shared across histories, so the f(zhat^(i)) factors do change the weighted gradient of the shared theta. That is an unsupported step in the derivation, and I flag it as an omitted proof, but it is a correctness and approximation gap, not a reduction of the output to the input: Eq. 1 is not defined in terms of Eq. 3, and no experimental result is obtained by relabeling a fitted value. The head-to-head active and passive evaluation in Section VI-A minimizes the same task costs for both variants; neither planner's parameters are fit to the measured comparison, and the passive policy class is a controlled ablation, not a fitted baseline. Two references overlap with the authors ([13] and [18]), but [13] is only a related-work contrast and [18] is used to justify the pure-strategy assumption; the central algorithm and its evaluation do not depend on the truth of either citation. The Section VII-B limitations explicitly concede the finite-history assumption and the assumptions of Section III-A, which is consistent with a self-contained approximation method rather than a circular one.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

These are the main quantities and assumptions the reported results depend on. The free parameters are tuning and approximation choices, not fit to held-out data, but they affect whether the method works at all. The most serious unproved step is the reduction from Eq. 3 to Eq. 1; the other axioms are stated or admitted limitations.

free parameters (6)
  • gamma = 0.1
    Fraction of particles conditioned on true observations; chosen by hand to trade off closed-loop accuracy against opponent-model correctness. Section IV-B, Algorithm 3.
  • T_past and T_future = 6 or 7
    Length of the finite observation-history window and planning horizon. The method's behavior depends on this window being informative, which the authors note fails in hide-and-seek. Sections III-B and VI-A.
  • K_all and K_batch = K_all=1000, K_batch=10
    Particle counts for the joint distribution and the Monte Carlo gradient batch. Table IV shows K_batch materially changes the resulting cost, so these are not innocuous implementation details.
  • N_eq = varied in experiments (1-4)
    Number of equilibrium solves used to maintain multiple candidate policies for opponent modeling. Section IV-C and Section VI-C.
  • Warehouse cost and noise coefficients = alpha=4, beta=20, eta1=eta2=4
    Scenario-specific reward shaping and observation-noise scaling used in the warehouse experiment. Section V-B, but not core to the method.
  • Pursuit observation noise parameters = sigma_base, C_scale, f (values not reported)
    Shape the sensor model in all tag variants; values are not stated in the text, which also hurts reproducibility.
assumptions (5)
  • ad hoc to paper Eq. 3 reduces to Eq. 1 because per-history positive scaling factors do not change equilibrium
    Section III-C asserts this without proof for the shared neural-network policies used in Section IV-D, where histories are coupled through shared parameters.
  • domain assumption Pure-strategy Nash equilibria exist and stochastic gradient play converges to one
    Assumption 3 and Sections IV-A and VI-B rely on empirical convergence; cited reference [15] notes gradient play may not converge.
  • domain assumption A finite observation-history window of length T_past is sufficiently informative
    Section VII-B admits this fails in the hide-and-seek evader case, making it load-bearing for the method's scope.
  • domain assumption The unconditioned joint particle distribution q_tbar represents all players' beliefs without a belief hierarchy
    Sections III-C and IV-B condition only a gamma fraction of particles with true observations, an approximation that can cause opponent models to diverge.
  • domain assumption All costs, transitions, and observations are differentiable and public
    Assumptions 1 and 2 in Section III-A; required for gradient play and common-knowledge planning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online Competitive Information Gathering for Partially Observable Trajectory Games." pith.science (2026). https://pith.science/paper/6XR3SOVR

@misc{pith2026250601927,
  author       = {Pith},
  title        = {Pith review of: Online Competitive Information Gathering for Partially Observable Trajectory Games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6XR3SOVR}},
  note         = {Machine review of arXiv:2506.01927}
}
abstract

Game-theoretic agents must make plans that optimally gather information about their opponents. These problems are modeled by partially observable stochastic games (POSGs), but planning in fully continuous POSGs is intractable without heavy offline computation or assumptions on the order of belief maintained by each player. We formulate a finite history/horizon refinement of POSGs which admits competitive information gathering behavior in trajectory space, and through a series of approximations, we present an online method for computing rational trajectory plans in these games which leverages particle-based estimations of the joint state space and performs stochastic gradient play. We also provide the necessary adjustments required to deploy this method on individual agents. The method is tested in continuous pursuit-evasion and warehouse-pickup scenarios (alongside extensions to $N > 2$ players and to more complex environments with visual and physical obstacles), demonstrating evidence of active information gathering and outperforming passive competitors.

Figures

Figures reproduced from arXiv: 2506.01927 by the authors.

Figure 1
Figure 1. Illustration of information gathering modalities in a pursuit-evasion [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of an equilibrium selection failure in a pursuit-evasion [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Trajectories in the three simulated UAV tag environments. Pursuer is blue; evader is red. Plans are solid lines and histories are dotted. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Active information gathering behavior in a 4-step warehouse pickup [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: P2’s cost as the first step of WAREHOUSE is solved. Ribbon indicates standard error. We also examined empirical scaling with respect to the method parameters. Table III shows scaling behavior w.r.t Tfuture over ten trials. The relation between Tfuture and time-to￾conve…
Figure 6
Figure 6. Figure 6: Distance between players in field of view tag for different [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Negative log likelihood of opponent position for settings of [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [1]

    MIT Press, 2024

    Stefano V Albrecht, Filippos Christianos, and Lukas Sch¨afer.Multi-agent reinforcement learning: Founda- tions and modern approaches. MIT Press, 2024

    work page 2024

  2. [2]

    Mastering the game of no-press diplomacy via human-regularized reinforcement learning and planning.arXiv preprint arXiv:2210.05492, 2022

    Anton Bakhtin, David J Wu, Adam Lerer, Jonathan Gray, Athul Paul Jacob, Gabriele Farina, Alexander H Miller, and Noam Brown. Mastering the game of no-press diplomacy via human-regularized reinforcement learning and planning.arXiv preprint arXiv:2210.05492, 2022

    arXiv

  3. [3]

    Springer, 2018

    Tamer Bas ¸ar and Georges Zaccour.Handbook of dynamic game theory. Springer, 2018

    work page 2018

  4. [4]

    Bridging the Gap be- tween Partially Observable Stochastic Games and Sparse POMDP Methods, May 2024

    Tyler Becker and Zachary Sunberg. Bridging the Gap be- tween Partially Observable Stochastic Games and Sparse POMDP Methods, May 2024. URL http://arxiv.org/abs/ 2405.18703. arXiv:2405.18703 [cs]

    arXiv 2024

  5. [5]

    Julia: A fresh approach to numerical computing.SIAM review, 59(1):65–98, 2017

    Jeff Bezanson, Alan Edelman, Stefan Karpinski, and Viral B Shah. Julia: A fresh approach to numerical computing.SIAM review, 59(1):65–98, 2017

    work page 2017

  6. [6]

    Deep counterfactual regret minimization

    Noam Brown, Adam Lerer, Sam Gross, and Tuomas Sandholm. Deep counterfactual regret minimization. InInternational conference on machine learning, pages 793–802. PMLR, 2019

    2019

  7. [7]

    ALGAMES: A Fast Solver for Constrained Dynamic Games

    Simon Le Cleac’h, Mac Schwager, and Zachary Manch- ester. Algames: A fast solver for constrained dynamic games.arXiv preprint arXiv:1910.09713, 2019

    work page Pith review arXiv 1910

  8. [8]

    Interac- tive pomdps: Properties and preliminary results

    Piotr J Gmytrasiewicz and Prashant Doshi. Interac- tive pomdps: Properties and preliminary results. In International Conference on Autonomous Agents: Pro- ceedings of the Third International Joint Conference on Autonomous Agents and Multiagent Systems-, volume 3, pages 1374–1375, 2004

    work page 2004

Show all 28 references
  1. [9]

    Learning others’ intentional models in multi-agent settings using interac- tive pomdps.Advances in Neural Information Processing Systems, 31, 2018

    Yanlin Han and Piotr Gmytrasiewicz. Learning others’ intentional models in multi-agent settings using interac- tive pomdps.Advances in Neural Information Processing Systems, 31, 2018

    2018

  2. [10]

    Deep reinforcement learning from self-play in imperfect-information games

    Johannes Heinrich and David Silver. Deep reinforcement learning from self-play in imperfect-information games. arXiv preprint arXiv:1603.01121, 2016

    2016 arXiv

  3. [11]

    Flux: Elegant machine learning with julia

    Mike Innes. Flux: Elegant machine learning with julia. Journal of Open Source Software, 3(25):602, 2018

    2018

  4. [12]

    Rethinking formal models of partially observable multiagent decision making.Artifi- cial Intelligence, 303:103645, 2022

    V ojtˇech Kova ˇr´ık, Martin Schmid, Neil Burch, Michael Bowling, and Viliam Lis `y. Rethinking formal models of partially observable multiagent decision making.Artifi- cial Intelligence, 303:103645, 2022

    2022

  5. [13]

    Multi-hypothesis interactions in game-theoretic motion planning

    Forrest Laine, David Fridovich-Keil, Chih-Yuan Chiu, and Claire Tomlin. Multi-hypothesis interactions in game-theoretic motion planning. In2021 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 8016–8023. IEEE, 2021

    2021

  6. [14]

    A unified game-theoretic approach to multiagent reinforcement learning.Advances in neural information processing systems, 30, 2017

    Marc Lanctot, Vinicius Zambaldi, Audrunas Gruslys, Angeliki Lazaridou, Karl Tuyls, Julien P ´erolat, David Silver, and Thore Graepel. A unified game-theoretic approach to multiagent reinforcement learning.Advances in neural information processing systems, 30, 2017

    2017

  7. [15]

    On gradient-based learning in continuous games.SIAM Journal on Mathematics of Data Science, 2(1):103–131, 2020

    Eric Mazumdar, Lillian J Ratliff, and S Shankar Sastry. On gradient-based learning in continuous games.SIAM Journal on Mathematics of Data Science, 2(1):103–131, 2020

    2020

  8. [16]

    XDO: A Double Oracle Algorithm for Extensive-Form Games

    Stephen McAleer, JB Lanier, Kevin A Wang, Pierre Baldi, and Roy Fox. XDO: A Double Oracle Algorithm for Extensive-Form Games. InAdvances in Neural Information Processing Systems, volume 34, pages 23128–23139. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/pa...

    2021

  9. [17]

    Mastering the game of stratego with model-free multi- agent reinforcement learning.Science, 378(6623):990– 996, 2022

    Julien Perolat, Bart De Vylder, Daniel Hennes, Eugene Tarassov, Florian Strub, Vincent de Boer, Paul Muller, Jerome T Connor, Neil Burch, Thomas Anthony, et al. Mastering the game of stratego with model-free multi- agent reinforcement learning.Science, 378(6623):990– 996, 2022

    2022

  10. [18]

    Learn- ing mixed strategies in trajectory games

    Lasse Peters, David Fridovich-Keil, Laura Ferranti, Cyrill Stachniss, Javier Alonso-Mora, and Forrest Laine. Learn- ing mixed strategies in trajectory games. In18th Robotics: Science and Systems, RSS 2022. MIT Press Journals, 2022

    2022

  11. [19]

    Belief space planning assuming maximum likelihood observations

    Robert Platt Jr, Russ Tedrake, Leslie Pack Kaelbling, and Tomas Lozano-Perez. Belief space planning assuming maximum likelihood observations. InRobotics: Science and systems, volume 2, 2010

    2010

  12. [20]

    Information gathering actions over human internal state

    Dorsa Sadigh, S Shankar Sastry, Sanjit A Seshia, and Anca Dragan. Information gathering actions over human internal state. In2016 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), pages 66–73. IEEE, 2016

    2016

  13. [21]

    Theory of mind for a humanoid robot

    Brian Scassellati. Theory of mind for a humanoid robot. Autonomous Robots, 12:13–24, 2002

    2002

  14. [22]

    Stochastic dynamic games in belief space

    Wilko Schwarting, Alyssa Pierson, Sertac Karaman, and Daniela Rus. Stochastic dynamic games in belief space. IEEE Transactions on Robotics, 37(6):2157–2172, 2021

    2021

  15. [23]

    Online planning for interactive-pomdps using nested monte carlo tree search

    Jonathon Schwartz, Ruijia Zhou, and Hanna Kurniawati. Online planning for interactive-pomdps using nested monte carlo tree search. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8770–8777. IEEE, 2022

    2022

  16. [24]

    Solving heads-up limit texas hold’em

    Oskari Tammelin, Neil Burch, Michael Johanson, and Michael Bowling. Solving heads-up limit texas hold’em. InTwenty-fourth international joint conference on artifi- cial intelligence, 2015

    2015

  17. [25]

    Grandmaster level in starcraft ii using multi-agent reinforcement learning.nature, 575(7782): 350–354, 2019

    Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Micha¨el Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning.nature, 575(7782): 350–354, 2019

    2019

  18. [26]

    On the computational complexity of stochastic controller optimization in pomdps.ACM Transactions on Compu- tation Theory (TOCT), 4(4):1–8, 2012

    Nikos Vlassis, Michael L Littman, and David Barber. On the computational complexity of stochastic controller optimization in pomdps.ACM Transactions on Compu- tation Theory (TOCT), 4(4):1–8, 2012

    2012

  19. [27]

    Game theoretic motion planning for multi-robot racing

    Zijian Wang, Riccardo Spica, and Mac Schwager. Game theoretic motion planning for multi-robot racing. In Distributed Autonomous Robotic Systems: The 14th In- ternational Symposium, pages 225–238. Springer, 2019

    2019

  20. [28]

    An overview of multi- agent reinforcement learning from game theoretical per- spective.arXiv preprint arXiv:2011.00583, 2020

    Yaodong Yang and Jun Wang. An overview of multi- agent reinforcement learning from game theoretical per- spective.arXiv preprint arXiv:2011.00583, 2020

    2011 arXiv

Pith tools

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