Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

GUARD: Constructing Realistic Two-Player Matrix and Security Games for Benchmarking Game-Theoretic Algorithms

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 Random game benchmarks are degenerate: random uniform games are nearly solved by a pure strategy, so this paper builds realistic security-game instances from open data.

desk verdict Solid random-game degeneracy theorems and a useful open benchmark framework, but 'realistic' is an overclaim until the hand-set utility constants are validated. read the letter →

arxiv 2505.14547 v2 pith:H3P45J5T submitted 2025-05-20 cs.GT

Noah Krever , Jakub Černý , Moïse Blanchard , Christian Kroer This is my paper · ORCID
classification cs.GT MSC 91A0591A1091A80
keywords game-theoreticbenchmarksStackelbergequilibriumsecuritygamesrandomnormal-formNashdata-drivengamegenerationbenchmarkdegeneracy
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

Game-theoretic algorithms are usually tested on random payoffs, and this paper argues that such benchmarks are degenerate. It proves that in two-player normal-form games with i.i.d. uniform payoffs, a pure strategy already achieves an expected Stackelberg value of $1 - 1/(n+1)$, while an explicitly constructed $O(\log n)$-sparse strategy matches the optimal value up to constant factors; in random security games, a single defender resource used with probability $O(1/\sqrt{T})$ already yields near-optimal utility. Random instances therefore hand an unrealistic advantage to the defender and make algorithm comparisons uninformative. The paper introduces GUARD, a framework that generates matrix and security games whose target values derive from open-access animal movement, demographic, and infrastructure data, and shows empirically that these instances produce larger equilibrium supports, lower defender utilities, and more demanding convergence behavior than random baselines.

What carries the argument

The argument rests on two constructions. The first is the i.i.d. uniform payoff model: with $A,B$ uniform in $[0,1]$, the attacker's best response to a pure strategy is the maximum of $n$ i.i.d. uniforms, which makes the pure-strategy Stackelberg value exactly a $\mathrm{Beta}(n,1)$ random variable; the sparse-strategy lower bound is built by concentrating weight on a high-value entry and then using Chernoff and Hoeffding bounds to keep the attacker's best response pinned to the column of that entry. The second is the schedule-ordering construction for security games: order the schedules by the attacker's best uncovered value, define coverage probabilities that lower the attacker's payoff on all earlier schedules below the best later one, and show the defender's value equals the best uncovered utility among the surviving schedules; this yields the $O(1/\sqrt{RT})$ bound. In GUARD itself, the machinery is a data pipeline that converts GPS collar records, census block populations, and map infrastructure features into graph nodes and target utilities, then expands them into normal-form or schedule-form games for equilibrium solvers.

What would settle it

Take random normal-form games with correlated payoffs (e.g., $A_{ij}=u_i+v_j+\epsilon_{ij}$ with shared row and column components and small noise) and compute the SSE value; if the best pure strategy falls well below $1-O(1/n)$ or optimal support grows with $n$, the degeneracy is an artifact of i.i.d. uniforms, not of randomness per se. Similarly, if measured step costs or infrastructure replacement values were substituted for the hand-set constants and randomized baselines still collapsed to pure strategies, the empirical contrast would be confirmed; if they did not, the realism claim would be weakened.

Watch

Extended reading notes

Core claim

The central claim is that randomness in benchmark generation, not the underlying game class, creates the easy problem. For normal-form games with payoffs drawn i.i.d. uniformly, Theorem 1 shows the best pure Stackelberg strategy has value distributed as $\mathrm{Beta}(n,1)$, so its expectation is $1 - 1/(n+1)$ and it falls below $1 - C/n$ with probability at most $e^{-C}$; the same theorem gives universal constants $c_0,c_1,c_2$ such that a $c_0\log n$-sparse strategy achieves the full Stackelberg value up to constant factors. For security games on partitioned schedules, Theorem 2 bounds the expected defender value below by $-c(\sqrt{\alpha/(RT)} + 1/k)$, where $\alpha$ is the ratio of largest to smallest schedule and $R$ is the number of resources, so even with $R \ll k$ a schedule used with vanishing probability is near-optimal. The paper concludes that uniform random games systematically inflate defender utility, and it validates this by comparing realistic GUARD instances with randomized baselines: real instances keep larger supports and much lower defender utilities, matching the theoretical prediction.

Load-bearing premise

The 'realistic' quality of the generated games rests on hand-specified utility constants such as infrastructure weights, step costs, and escape-proximity factors that are never checked against observed decisions, so if those constants are off, the empirical contrast is between two arbitrary constructions rather than realistic versus unrealistic benchmarks.

Editorial extensions

If this is right

  • For random uniform normal-form games, the Stackelberg value is near-solved in expectation without any support enumeration, so benchmark comparisons that report large gains over pure strategies on random instances are measuring noise rather than algorithm quality.
  • Randomized security-game instances with uniform target values collapse defender strategies to near-pure or single-resource play, which means defender utility numbers from such baselines overstate what a defender can expect in real deployments.
  • Data-derived GUARD instances preserve larger equilibrium supports and lower defender utilities, so algorithms that look strong on random baselines can be re-ranked when evaluated on these more demanding instances.
  • Because GUARD instances can be exported to standard game formats, they offer a reproducible way to benchmark Stackelberg and Nash solvers on security-inspired scenarios without access to the proprietary datasets that have limited past evaluation.

Reading between the lines

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

  • A natural extension is to characterize degeneracy under correlated payoff models: if real utility matrices share row or column components, the extreme-value behavior that drives Theorem 1 may vanish, and random benchmarks from those distributions would not be degenerate.
  • The framework's realism is carried by hand-set constants rather than measured payoffs, so the empirical contrast is best read as evidence about the design space, not about actual wildlife or infrastructure decisions; calibrating those constants against observed choices would make the realism claim testable.
  • A practical selection rule follows from the theorems: benchmark suites should report the pure-strategy Stackelberg value and support-size distributions of each instance, letting users discard instances where a pure strategy already reaches near-optimal value.
  • The same degeneracy logic suggests that randomizing only target values inside a realistic map, as the RT baselines do, is sufficient to produce inflated defender utility; future work could vary schedule structure and target distribution independently to identify which component drives hardness.
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 introduces GUARD, an open-source framework that constructs two-player matrix and security game instances from open-access geospatial data (Movebank animal tracking, OpenStreetMap infrastructure, and US Census block populations) for benchmarking game-theoretic algorithms. The theoretical component proves two degeneracy results for random games: for i.i.d. uniform normal-form games, a pure strategy achieves expected Stackelberg value 1 - 1/(n+1) and an O(log n)-sparse strategy matches optimal value up to constant factors (Theorem 1); for random uniform security games with partitioned schedules, a single defender resource used with vanishing probability already yields near-optimal defender utility (Theorem 2). The empirical component compares GUARD's preset GSG and ISG instances against randomized baselines on support size, runtime, iterative convergence, and Stackelberg equilibrium outcomes, concluding that random benchmarks give defenders an unrealistic advantage and that GUARD instances produce richer strategy profiles.

Significance. If the realism claim is established, GUARD would fill a real gap in benchmarking infrastructure for algorithmic game theory, since security-game target-value data have been largely inaccessible. The theoretical part is a genuine contribution: Theorems 1 and 2 are stated under explicit random models and supported by detailed proofs in Appendices A.1 and A.2, and the paper ships open-source code with reproducible experiments and randomized baselines with error bars. The framework's export to standard formats (OpenSpiel, Gambit) and its customizable game-class hierarchy are practical strengths. However, the paper's second central claim, that GUARD instances are 'realistic', is not established: key utility functions and constants are manually chosen rather than validated, and the real-vs-random empirical contrast rests on single real instances without sensitivity or replication analysis. The paper is likely to be a useful contribution after the realism claim is tempered or supported by additional validation.

major comments (3)
  1. [Section 4.2.1, Appendices D.4/D.7, Table 1] The paper's central claim that GUARD generates 'realistic' game instances is not supported by evidence about the utility constants that drive the games. Appendix D.7 describes INFRA_WEIGHTS as 'manually assigned based on qualitative assessments'; Appendix D.4 gives hand-estimated elephant dollar values, ranger step costs, and coverage penalty ratios; Section 4.2.1 defines escape-proximity alpha and the raw_score formula with user-specified exponents. None of these quantities is fitted to or validated against real decision data, and no sensitivity analysis is reported for them. Since the empirical contrast in Table 1 and Figures 3-4 depends directly on these constants, the experiments currently establish a difference between structured synthetic instances and random synthetic instances, not between realistic and unrealistic benchmarks. The authors should either validate the utility constants against domain data, report sensitivity of the qualitative conclusions over plausible parameter ranges, or explicitly reframe the contribution as 'data-informed structured' benchmarks rather than 'realistic' ones.
  2. [Section 3, Theorem 2] The statement of Theorem 2 is internally inconsistent. It assumes '1 ≤ k < R' for the number of schedules k and resources R, then immediately states 'When R ≥ k, we have V(x*) = 0', a case that never occurs under the stated assumption. The subsequent phrase 'even if R ∈ (0, 1]' also treats the integer number of resources as a real-valued parameter. Because Theorem 2 is the load-bearing result for the claim that random security games are degenerate for the defender, the parameter regimes and the meaning of V(x*) (defender utility under attacker best response) need to be stated precisely and consistently with the proof in Appendix A.2.
  3. [Section 5, Table 1, Figures 3-4] The empirical claims that real instances 'tend to exhibit higher complexity' and 'consistently exhibit faster convergence' are based on a single realization per real configuration, while the randomized baselines are averaged over 10 seeds with standard errors. For example, in Table 1 the GSG Simple row reports real support S=6 versus randomized-target support 5.0±0.52, a difference within the noise of the baseline. Without replication across data subsets, K-means seeds, or alternative target-construction settings, the qualitative real-vs-random differences may be specific to the particular preset instances rather than a general property of GUARD. The authors should add error bars or sensitivity analyses for the real instances, or narrow the claims accordingly.
minor comments (5)
  1. [Appendix C, first sentence] The sentence 'This Appendix details the the the GUARD game class hierarchy' contains a triple repetition of 'the' and should be corrected.
  2. [Section 4.4] The word 'aforemention' in 'the aforementionend Lobéké National Park elephants' is a typo for 'aforementioned'.
  3. [Appendix A.1, Lemma 2] The sentence 'There exists a universal constants c2 > 0' should read 'There exists a universal constant c2 > 0'.
  4. [Section 2] In the definition of best response, 'argmax_{y∈[m]} x^T B y' uses y both as an index and as a pure strategy; using column indices (e.g., j) consistently would improve readability.
  5. [References] Several references are ephemeral web pages (e.g., refs. [10], [14], [15], [31], [35]); the authors should consider archiving them or citing peer-reviewed sources where available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the degeneracy theorems are proven from explicit i.i.d. random models, and the empirical random baselines are independently generated; the hand-set utility constants affect realism fidelity, not derivation circularity.

full rationale

The paper's central theoretical claims are derived from first principles under explicit random models. Theorem 1 is proven constructively from i.i.d. uniform entries in A and B: the pure-strategy value is shown to be the maximum of n i.i.d. uniforms, and the O(log n)-sparse lower bound is established by an explicit sparse strategy construction analyzed with Chernoff/Hoeffding bounds. Theorem 2 constructs an explicit defender coverage distribution and bounds its expected value using only the assumed i.i.d. uniform uncovered utilities. Neither theorem uses GUARD-generated data as input, and neither defines its target quantity in terms of its own output. The empirical section compares GUARD instances against randomized baselines generated independently: 'randomized instances generated by uniformly sampling matrix entries within the observed range of real payoff values' (Section 5), and Table 1 randomizes target values, matrices, and schedules separately without fitting to the degeneracy contrast. The hand-specified utility constants (INFRA_WEIGHTS, elephant dollar values, step costs, coverage scaling factors) are choices about realism and are explicitly acknowledged in Appendix D.7 as 'manually assigned based on qualitative assessments,' but they are not fitted to reproduce the observed sparsity or defender-utility differences, so no fitted input is being relabeled as a prediction. The only self-citation in a load-bearing proof position is Blanchard and Voráček [4] for a standard large-deviations identity used inside Lemma 2; this is a technical lemma, not a smuggled ansatz or an imported uniqueness theorem, and the proof's main events are established from first principles. The Limitations section candidly notes that 'the framework depends on raw datasets that may imperfectly capture real-world dynamics,' which tempers the realism claim but does not make the derivation circular. Overall, the derivation chain is self-contained with respect to the paper's stated random-game models, and the empirical contrast, while potentially sensitive to hand-set parameters, does not reduce by construction to any fitted input.

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

The theoretical claims rest on standard random models and a partition-schedule assumption, which are stated explicitly. The framework's realism claim, in contrast, depends on a substantial set of hand-selected value functions and constants that are not fitted or validated, so the ledger is dominated by ad hoc utility-modeling choices.

free parameters (5)
  • infrastructure type weights (INFRA_WEIGHTS) = e.g., plant: 1.5, hospital: 1.5, school: 1.25, pole: 0.85
    Manually assigned criticality weights that determine ISG target scores via raw_score = W * (log(P+1))^alpha; they directly shape the 'realistic' target values in infrastructure games (Appendix D.7).
  • attacker/defender animal values = 2350 USD attacker, 22966 USD defender
    Hand-estimated black-market tusk value and ecotourism value that scale GSG payoffs; used in general-sum SSE experiments (Appendix D.4).
  • defender step costs = 1.17 per km (GSG), 1.00 per block (ISG)
    Estimated patrol labor/fuel/equipment costs that alter general-sum Stackelberg equilibria; derived from cited sources but ultimately hand-chosen (Appendix D.4).
  • coverage penalty factors = 5 (GSG), 3 (ISG)
    Schedule-form attacker and defender coverage scalers set to match cited apprehension and crime reduction rates; these affect the generated utility matrices and all empirical comparisons (Section 5).
  • escape proximity scaling alpha = 1.0 (GSG), 0.5 (ISG)
    Controls how strongly distance to the escape line or point multiplies attacker utility; chosen by the authors and not fitted (Section 4.2.1, Appendix D.4).
assumptions (5)
  • domain assumption Random game model: A and B (or uu_d, uu_a) are i.i.d. uniform
    Theorems 1 and 2 establish degeneracy only for uniformly distributed payoffs; the authors state they expect robustness to other distributions but do not prove it (Section 3).
  • domain assumption Standard security game structure: attacker selects targets, defender allocates resources, utilities depend only on coverage
    Both the theoretical results and the framework rely on the standard Stackelberg security game formulation (Section 2, Section 4.2).
  • domain assumption Schedules partition the target set
    Theorem 2 explicitly assumes the k schedules form a partition; the paper says results should be relatively robust to this, but the proof depends on it (Section 3, Appendix A.2).
  • ad hoc to paper Animal movement density or cluster size is a valid proxy for poaching target value
    GSG scoring converts GPS observation density or K-means cluster size into target scores via cluster_size * (num_animals/num_total_observations); this ecological equivalence is assumed, not validated (Section 4.2.1).
  • ad hoc to paper Census population and infrastructure type weights determine target value
    ISG scores are raw_score = W * (log(P+1))^alpha with hand-chosen weights; no external data validates this as a model of real infrastructure value (Section 4.2.1, Appendix D.7).

how reviews work

0 comments
Cite this review

Pith. "Pith review of GUARD: Constructing Realistic Two-Player Matrix and Security Games for Benchmarking Game-Theoretic Algorithms." pith.science (2026). https://pith.science/paper/H3P45J5T

@misc{pith2026250514547,
  author       = {Pith},
  title        = {Pith review of: GUARD: Constructing Realistic Two-Player Matrix and Security Games for Benchmarking Game-Theoretic Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H3P45J5T}},
  note         = {Machine review of arXiv:2505.14547}
}
read the original abstract

Game-theoretic algorithms are commonly benchmarked on recreational games, classical constructs from economic theory such as congestion and dispersion games, or entirely random game instances. While the past two decades have seen the rise of security games -- grounded in real-world scenarios like patrolling and infrastructure protection -- their practical evaluation has been hindered by limited access to the datasets used to generate them. In particular, although the structural components of these games (e.g., patrol paths derived from maps) can be replicated, the critical data defining target values -- central to utility modeling -- remain inaccessible. In this paper, we introduce a flexible framework that leverages open-access datasets to generate realistic matrix and security game instances. These include animal movement data for modeling anti-poaching scenarios and demographic and infrastructure data for infrastructure protection. Our framework allows users to customize utility functions and game parameters, while also offering a suite of preconfigured instances. We provide theoretical results highlighting the degeneracy and limitations of benchmarking on random games, and empirically compare our generated games against random baselines across a variety of standard algorithms for computing Nash and Stackelberg equilibria, including linear programming, incremental strategy generation, and self-play with no-regret learners.

Figures

Figures reproduced from arXiv: 2505.14547 by the authors.

Figure 1
Figure 1. The structure of the GUARD framework. error term is −c p α RT , with a smaller correction when schedules are few. Consequently, for random uniform security games, Player 1 achieves near-optimal utility with very few resources: if α = O(1), then E[V (x ∗ )] ≳ −O 1/ √ RT , so a single resource already yields expected utility −O(1/ √ T). Moreover, one can achieve near-maximum utility by using that resource only with p… view at source ↗
Figure 2
Figure 2. (Left) Down-sampled elephant movements in Lobéké National Park and the corresponding [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Comparison of GSG utility and runtime sparsity results for NFG and SFG settings with [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Convergence of iterative algorithms for different game types, formulations, and real vs. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Contested Route Planning

    cs.GT 2025-08 reject novelty 4.0 of 10

    Contested route planning is cast as a zero-sum game and solved approximately by double oracle, with NP-hardness of exact play proved, but a flawed Red best-response reduction undermines the equilibrium claims.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages · cited by 1 Pith paper

  1. [1]

    Commitment to sparse strategies in two-player games

    Salam Afiouni, Jakub ˇCerný, Chun Kai Ling, and Christian Kroer. Commitment to sparse strategies in two-player games. In AAAI, 2025

    work page 2025

  2. [2]

    Human-level play in the game of diplomacy by combining language models with strategic reasoning

    Anton Bakhtin, Noam Brown, Emily Dinan, Gabriele Farina, Colin Flaherty, Daniel Fried, Andrew Goff, Jonathan Gray, Hengyuan Hu, et al. Human-level play in the game of diplomacy by combining language models with strategic reasoning. Science, 378(6624):1067–1074, 2022

    2022

  3. [3]

    Nash equilibria in random games

    Imre Bárány, Santosh Vempala, and Adrian Vetta. Nash equilibria in random games. Random Structures & Algorithms, 31(4):391–405, 2007

    work page 2007

  4. [4]

    Tight bounds for local glivenko-cantelli

    Moïse Blanchard and Václav V oráˇcek. Tight bounds for local glivenko-cantelli. In International Conference on Algorithmic Learning Theory, pages 179–220. PMLR, 2024

    work page 2024

  5. [5]

    Heads-up limit hold’em poker is solved

    Michael Bowling, Neil Burch, Michael Johanson, and Oskari Tammelin. Heads-up limit hold’em poker is solved. Science, 347(6218):145–149, 2015

    work page 2015

  6. [6]

    George W. Brown. Iterative solution of games by fictitious play. In T. C. Koopmans, editor, Activity Analysis of Production and Allocation. Wiley, New York, 1951

    work page 1951

  7. [7]

    One size does not fit all: A game-theoretic approach for dynamically and effectively screening for threats

    Matthew Brown, Arunesh Sinha, Aaron Schlenker, and Milind Tambe. One size does not fit all: A game-theoretic approach for dynamically and effectively screening for threats. InProceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016

    work page 2016

  8. [8]

    Superhuman AI for heads-up no-limit poker: Libratus beats top professionals

    Noam Brown and Tuomas Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science, 359(6374):418–424, 2018

    work page 2018

Show all 51 references
  1. [9]

    J. C. Bruch, J. Matthiopoulos, D. L. Miller, A. Singh, and T. P. Young. Spatial context matters in habitat selection: Understanding behavioural drivers of population distribution in a recovering large mammal. Journal of Applied Ecology, 58(3):503–514, 2021

    2021

  2. [10]

    Cameroon plans 94.2% reduction in fuel price subsidies in 2025

    Cameroon News Agency. Cameroon plans 94.2% reduction in fuel price subsidies in 2025. https://cameroonnewsagency.com/ cameroon-plans-94-2-reduction-in-fuel-price-subsidies-in-2025/ , 2024. Accessed: 2025-05-16

    2025

  3. [11]

    Contested logistics: A game theoretic approach

    Jakub ˇCerný, Chun Kai Ling, Darshan Chakrabarti, Jingwen Zhang, Gabriele Farina, Christian Kroer, and Garud Iyengar. Contested logistics: A game theoretic approach. In International Conference on Decision and Game Theory for Security. Springer, 2024

    2024

  4. [12]

    Settling the complexity of two-player nash equilibrium

    Xi Chen and Xiaotie Deng. Settling the complexity of two-player nash equilibrium. In FOCS, volume 6, pages 261–272, 2006

    2006

  5. [13]

    Computing the optimal strategy to commit to

    Vincent Conitzer and Tuomas Sandholm. Computing the optimal strategy to commit to. In Proceedings of the 7th ACM conference on Electronic commerce, pages 82–90, 2006

    2006

  6. [14]

    Jeep wrangler mpg ratings.https://www

    DARCARS Chrysler Jeep Dodge of Marlow Heights. Jeep wrangler mpg ratings.https://www. darcarschryslerjeepdodgeofmarlowheights.com/jeep-wrangler-mpg.html , 2024. Accessed: 2025-05-16

    2024

  7. [15]

    The fuel efficiency of law enforcement vehicles

    Shea Denning. The fuel efficiency of law enforcement vehicles. https://nccriminallaw. sog.unc.edu/the-fuel-efficiency-of-law-enforcement-vehicles/ , 2020. Ac- cessed: 2025-05-16

    2020

  8. [16]

    Probability of a pure equilibrium point in n-person games

    Melvin Dresher. Probability of a pure equilibrium point in n-person games. Journal of Combinatorial Theory, 8(1):134–145, 1970

    1970

  9. [17]

    Tourism and the green militarisation of anti-poaching: New partnerships and technologies

    Rosaleen Duffy, Francis Massé, Daniel Smidt, Esther Marijnen, Judith Verweijen, and Elizabeth Lunstrum. Tourism and the green militarisation of anti-poaching: New partnerships and technologies. World Development Sustainability, 1:100027, 2022

    2022

  10. [18]

    Deploying paws: Field optimization of the protection assistant for wildlife security

    Fei Fang, Thanh Nguyen, Rob Pickles, Wai Lam, Gopalasamy Clements, Bo An, Amandeep Singh, Milind Tambe, and Andrew Lemieux. Deploying paws: Field optimization of the protection assistant for wildlife security. In Proceedings of the AAAI Conference on Artificial Intelligence, v...

    2016

  11. [19]

    Green security games: Apply game theory to addressing green security challenges

    Fei Fang and Thanh H Nguyen. Green security games: Apply game theory to addressing green security challenges. ACM SIGecom Exchanges, 15(1):78–83, 2016

    2016

  12. [20]

    When security games go green: Designing de- fender strategies to prevent poaching and illegal fishing

    Fei Fang, Peter Stone, and Milind Tambe. When security games go green: Designing de- fender strategies to prevent poaching and illegal fishing. In Twenty-Fourth International Joint Conference on Artificial Intelligence, 2015

    2015

  13. [21]

    Faster game solving via predictive blackwell approachability: Connecting regret matching and mirror descent

    Gabriele Farina, Christian Kroer, and Tuomas Sandholm. Faster game solving via predictive blackwell approachability: Connecting regret matching and mirror descent. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 5363–5371, 2021

    2021

  14. [22]

    Wayne M. Getz, W. Kilian, R. Zidon, and M. Tsalyuk. Data from: African elephants in etosha national park (data from tsalyuk et al. 2018). Movebank Data Repository, 2018. Movebank Study ID: 605129389

    2018

  15. [23]

    The probability of an equilibrium point

    K Goldberg, AJ Goldman, and M Newman. The probability of an equilibrium point. Journal of Research of the National Bureau of Standards, 72(2):93–101, 1968

    1968

  16. [24]

    A simple adaptive procedure leading to correlated equilib- rium

    Sergiu Hart and Andreu Mas-Colell. A simple adaptive procedure leading to correlated equilib- rium. Econometrica, 68(5):1127–1150, 2000

    2000

  17. [25]

    Software assistants for randomized patrol planning for the lax airport police and the federal air marshal service

    Manish Jain, Jason Tsai, James Pita, Christopher Kiekintveld, Shyamsunder Rathi, Milind Tambe, and Fernando Ordónez. Software assistants for randomized patrol planning for the lax airport police and the federal air marshal service. Interfaces, 40(4):267–290, 2010

    2010

  18. [26]

    The effect of privately provided police services on crime: Evidence from the johns hopkins university police depart- ment

    Johns Hopkins University Office of Public Safety. The effect of privately provided police services on crime: Evidence from the johns hopkins university police depart- ment. https://publicsafety.jhu.edu/assets/uploads/sites/9/2024/12/effect_ of_privately_provided_police_service...

    2024

  19. [27]

    Davidson, Marit Berger, Gil Bohrer, Wolfgang Fiedler, Andrea Flack, Jens Hirt, Christian Hahn, Daniel Gauggel, Brian Russell, and et al

    Roland Kays, Sarah C. Davidson, Marit Berger, Gil Bohrer, Wolfgang Fiedler, Andrea Flack, Jens Hirt, Christian Hahn, Daniel Gauggel, Brian Russell, and et al. The movebank system for studying global animal movement and demography. Methods in Ecology and Evolution, 13(2):419–431, 2022

    2022

  20. [28]

    simple schedules

    Christopher Kiekintveld, Manish Jain, Jonathan Tsai, Jay Pita, Fernando Ordóñez, and Milind Tambe. Computing optimal randomized resource allocations for massive security games. In Proceedings of the Eighth International Conference on Autonomous Agents and Multiagent Systems (A...

    2009

  21. [29]

    Complexity of computing optimal stackelberg strategies in security resource allocation games

    Dmytro Korzhyk, Vincent Conitzer, and Ronald Parr. Complexity of computing optimal stackelberg strategies in security resource allocation games. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 24, pages 805–810, 2010

    2010

  22. [30]

    OpenSpiel: A framework for reinforcement learning in games

    Marc Lanctot, Edward Lockhart, Jean-Baptiste Lespiau, Vinicius Zambaldi, Satyaki Upadhyay, Julien Pérolat, Sriram Srinivasan, Finbarr Timbers, Karl Tuyls, Shayegan Omidshafiei, Daniel Hennes, Dustin Morrill, Paul Muller, Timo Ewalds, Ryan Faulkner, János Kramár, Bart De Vylder...

    1908 arXiv

  23. [31]

    2016–2017 final report

    Los Angeles County Civil Grand Jury. 2016–2017 final report. https://grandjury.co.la. ca.us/pdf/LOSANGELESCOUNTY2016-2017CIVILGRANDJURYFINALREPORT.pdf, 2017. Accessed: 2024-05-14

    2016

  24. [32]

    Elephant research – lobéké national park (cameroon) [collars 14118, 14120, 39839, 39840, 46179, 47574]

    Mark MacAllister and Michael Loomis. Elephant research – lobéké national park (cameroon) [collars 14118, 14120, 39839, 39840, 46179, 47574]. https://www.movebank.org, 2025. Movebank Study IDs: 2789883238, 2789896888, 2784843611, 2784424144, 2784086541, 2782906165, 2784539839, ...

    2025

  25. [33]

    Planning in the presence of cost functions controlled by an adversary

    H Brendan McMahan, Geoffrey J Gordon, and Avrim Blum. Planning in the presence of cost functions controlled by an adversary. In Proceedings of the 20th International Conference on Machine Learning (ICML-03), pages 536–543, 2003

    2003

  26. [34]

    Weaver, Marie De Longcamp, and Pierre Du Preez

    Robin Naidoo, Larkin C. Weaver, Marie De Longcamp, and Pierre Du Preez. Evaluating the effectiveness of conservation interventions at reducing human-wildlife conflict. Journal of Applied Ecology, 56(7):1498–1507, 2019

    2019

  27. [35]

    war on police

    Justin Nix, Scott E. Wolfe, and Bradley A. Campbell. Command-level police officers’ per- ceptions of the “war on police” and de-policing. https://www.ncbi.nlm.nih.gov/pmc/ articles/PMC5718484/, 2017. Accessed: 2024-05-14

    2017

  28. [36]

    Run the gamut: A comprehensive approach to evaluating game-theoretic algorithms

    Eugene Nudelman, Jennifer Wortman, Yoav Shoham, and Kevin Leyton-Brown. Run the gamut: A comprehensive approach to evaluating game-theoretic algorithms. In AAMAS, volume 4, pages 880–887, 2004

    2004

  29. [37]

    Overpass turbo: A web-based data filtering and visualiza- tion tool for openstreetmap

    Overpass Turbo Development Team. Overpass turbo: A web-based data filtering and visualiza- tion tool for openstreetmap. https://overpass-turbo.eu, 2025. Accessed: 2025-05-12

    2025

  30. [38]

    Mastering the game of stratego with model-free multiagent reinforcement learning

    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 multiagent reinforcement learning. Science, 378(6623):990–996, 2022

    2022

  31. [39]

    Deployed armor protection: the application of a game theoretic model for security at the los angeles international airport

    James Pita, Manish Jain, Janusz Marecki, Fernando Ordóñez, Christopher Portway, Milind Tambe, Craig Western, Praveen Paruchuri, and Sarit Kraus. Deployed armor protection: the application of a game theoretic model for security at the los angeles international airport. In Proce...

    2008

  32. [40]

    Rahul Savani and Theodore L. Turocy. Gambit: The package for computation in game theory, version 16.3.0. https://www.gambit-project.org, 2025

    2025

  33. [41]

    All about elephants: Characteristics, 2024

    SeaWorld Parks & Entertainment. All about elephants: Characteristics, 2024. Accessed: 2025-05-16

    2024

  34. [42]

    Multiagent systems: Algorithmic, game-theoretic, and logical foundations

    Yoav Shoham and Kevin Leyton-Brown. Multiagent systems: Algorithmic, game-theoretic, and logical foundations. Cambridge University Press, 2008

    2008

  35. [43]

    Stackelberg security games: Looking beyond a decade of success

    Arunesh Sinha, Fei Fang, Bo An, Christopher Kiekintveld, and Milind Tambe. Stackelberg security games: Looking beyond a decade of success. IJCAI, 2018

    2018

  36. [44]

    Security and game theory: algorithms, deployed systems, lessons learned

    Milind Tambe. Security and game theory: algorithms, deployed systems, lessons learned . Cambridge university press, 2011

    2011

  37. [45]

    Solving large imperfect information games using cfr+

    Oskari Tammelin. Solving large imperfect information games using cfr+. arXiv preprint arXiv:1407.5042, 2014

    2014 arXiv

  38. [46]

    Corrupting conservation: Assess- ing how corruption impacts ranger work

    Targeting Natural Resource Corruption (TNRC). Corrupting conservation: Assess- ing how corruption impacts ranger work. https://www.worldwildlife.org/pages/ tnrc-topic-brief-corrupting-conservation-assessing-how-corruption-impacts-ranger-work ,

  39. [47]

    Census Bureau

    U.S. Census Bureau. TIGER/Line Shapefiles, 2020. https://www.census.gov/ geographies/mapping-files/time-series/geo/tiger-line-file.html , 2020. Ac- cessed: 2025-05-12

    2020

  40. [48]

    Leadership with commitment to mixed strategies

    Bernhard V on Stengel and Shmuel Zamir. Leadership with commitment to mixed strategies. Technical report, Citeseer, 2004

    2004

  41. [49]

    Davidson, and Roland Kays

    Martin Wikelski, Sarah C. Davidson, and Roland Kays. Movebank: archive, analysis and sharing of animal movement data. https://www.movebank.org, 2025. Hosted by the Max Planck Institute of Animal Behavior. Accessed: 2025-05-12

    2025

  42. [50]

    centroid

    Shuxin Zhuang, Shuxin Li, Tianji Yang, Muheng Li, Xianjie Shi, Bo An, and Youzhi Zhang. Solving urban network security games: Learning platform, benchmark, and challenge for ai research. arXiv preprint arXiv:2501.17559, 2025. 12 A Proofs A.1 Proof of Theorem 1 First, we prove ...

    2025 arXiv

  43. [2020]

    Accessed: 2025-05-16

    2025

Pith tools

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