Pith. sign in

REVIEW 3 major objections 5 minor 59 references

Pareto-efficient coordination in a multi-agent system is exactly characterized by the feasibility of a finite set of linear inequalities, and this characterization supports a Type-I-error-optimal detector and distributionally robust utility

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 19:56 UTC pith:KWJLYYN2

load-bearing objection The paper's load-bearing Theorem 1 appears to miss the shared-constraint coupling, and the DRO objective doesn't match the reported reconstruction-error metric, so the headline claims should be treated with caution. the 3 major comments →

arxiv 2509.08956 v1 pith:KWJLYYN2 submitted 2025-09-10 eess.SY cs.SYeess.SP

Multi-Agent Inverse Reinforcement Learning for Identifying Pareto-Efficient Coordination -- A Distributionally Robust Approach

Luke Snow , Vikram Krishnamurthy This is my paper
classification eess.SY cs.SYeess.SP
keywords multi-agent inverse reinforcement learningPareto efficiencyrevealed preferenceAfriat's theoremstatistical detectiondistributionally robust optimizationWasserstein ambiguity setUAV coordination
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central project is to turn the question 'is this multi-agent system coordinating Pareto-efficiently?' from a guess about unknown preferences into a finite testable condition. The main theoretical claim is that a dataset of probes and agent responses is consistent with Pareto-efficient coordination if and only if a system of per-agent linear inequalities (a multi-agent generalization of Afriat's revealed-preference theorem) is feasible. On that foundation the authors build a statistical detector that, under the null hypothesis of coordination, bounds the probability of false alarm by a prespecified level and is optimal among detectors in its class. For the harder reconstruction problem, they propose a Wasserstein distributionally robust estimator that minimizes worst-case expected violation of the same inequalities, so utility estimates remain meaningful even when measurements are noisy or adversarial. The reason to care: an analyst observing radar returns or other shared-resource systems could use these tools to detect coordinated behavior and forecast responses without knowing the agents' utilities in advance.

Core claim

The paper's core discovery is that Pareto-efficient coordination in a linearly constrained multi-agent system is exactly characterized by the feasibility of a finite set of linear inequalities: there exist positive numbers u_i^s and λ_i^t such that u_i^s − u_i^t − λ_i^t α_t^T(β_i^s − β_i^t) ≤ 0 for every pair of time steps and every agent. If these inequalities hold, the observed responses are rationalizable as weighted-sum maximizers with strictly positive weights—and conversely; moreover, the rationalizing utilities can be written explicitly as the pointwise minimum of affine functions. The paper extends this deterministic equivalence into a statistical detector: for noisy observations, th

What carries the argument

The load-bearing object is the system of per-agent Afriat inequalities (10): for each agent i and time pairs (s,t), u_i^s − u_i^t − λ_i^t α_t^T(β_i^s − β_i^t) ≤ 0 with positive u,λ. This finite LP does three jobs at once: feasibility tests rationalizability; its minimal uniform relaxation Φ(D) defines the proximity-to-coordination metric (12); and feasible (u,λ) plug directly into the piecewise-affine utility formula (11). The second piece of machinery is the Wasserstein ambiguity set around the empirical noisy dataset: the DRO program (23) minimizes the worst-case expected violation h over distributions within 1-Wasserstein radius ϵ, and Theorem 3 rewrites this minimax problem as the semi-i

Load-bearing premise

The argument stands or falls on the claim that a single per-agent system of linear inequalities captures all the coupling in the shared constraint; the proof of that equivalence is only cited, not shown, so if the shared constraint links agents more tightly than the inequalities express, the detector and the robust estimator are both built on a mis-specified foundation.

What would settle it

Generate a dataset D by solving (8) with strictly positive weights and strictly concave utilities for many random probes α_t, then test feasibility of (10): any infeasible instance falsifies the 'necessary' direction of Theorem 1. Symmetrically, take a random dataset that satisfies (10), reconstruct utilities by (11), and check that the observed responses are optimal for (8) with those utilities; a mismatch falsifies sufficiency. Both checks are finite LPs plus one convex solve and can be run on synthetic data independent of the paper's simulations.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Coordination testing becomes a finite LP feasibility check on observed data alone, with no need to estimate utilities first.
  • When the LP is feasible, explicit concave utilities can be reconstructed and used to predict responses to unseen probes, including identifying which agents' objectives are weighted most heavily.
  • The detector (17) comes with a guaranteed Type-I error bound γ, and the chosen statistic Φ* is provably the best Type-I-error minimizer among statistics in [Φ*,Ψ].
  • The Wasserstein DRO estimator reduces worst-case reconstruction error (measured by Hausdorff distance between Pareto frontiers) while keeping average error comparable to naive reconstruction.
  • Because the radar–UAV derivation maps a covertness constraint to the same linear inequality, the same test works for a physical-layer coordination-detection problem, and by the paper's argument extends to any shared-linear-resource multi-agent system.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial: the rationalizability test inherits the usual finite-data limitation of Afriat-style theorems—if utilities exist at all, continuous concave monotone ones exist—so evidence against coordination can be read as evidence against the Pareto model, not just against a particular utility class; conversely, passing the test does not certify uniqueness of the recovered preferences.
  • Editorial: because the reconstructed utilities are ordinal, the scientifically meaningful output is the induced Pareto frontier, not the utility values; the paper's Hausdorff-distance evaluation could be turned into an active-learning objective for choosing future probes α_t that most sharply separate coordination from non-coordination.
  • Editorial: the detector's Type-I guarantee assumes the noise distribution Λ_i^t is known well enough to compute F_Ψ; a natural extension would quantify how threshold misspecification or heavy-tailed noise degrades the bound, since in covert settings noise statistics themselves may be contested.
  • Editorial: the paper treats a single shared linear constraint; multi-constraint or nonlinear resource bounds would call for a generalized revealed-preference system, and it is an open question whether the same LP equivalence survives coupling across constraints.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper develops a two-stage framework for multi-agent inverse reinforcement learning under a shared linear constraint. Stage 1 tests whether observed probe-response pairs (α_t, {β_t^i}) are consistent with Pareto-efficient coordination by checking per-agent Afriat inequalities (10). It also proposes a statistical detector with a claimed Type-I-error guarantee for noisy observations (Theorem 2). Stage 2 introduces a Wasserstein distributionally robust utility reconstruction procedure (23), reformulated as a semi-infinite program (24) and solved by an exchange method. The framework is instantiated in a radar-UAV covert coordination example, with reproducible numerical experiments.

Significance. The paper addresses a timely problem and has attractive components: a nonparametric revealed-preference characterization, a statistical detection guarantee, distributionally robust reconstruction, and open-source reproducible code. If the central equivalence were sound, the paper would be a valuable multi-agent extension of Afriat's theorem and could support practical coordination detection in adversarial settings. However, the main theorem is not proved in the manuscript and its statement appears to omit cross-agent coupling in the shared constraint; moreover, the DRO objective does not match the headline accuracy metric. These issues are load-bearing for both the detector and the reconstruction algorithm, so the contribution cannot be accepted as it stands.

major comments (3)
  1. [Section III.A, Theorem 1 / Eqs. (9)-(10)] The equivalence is the foundation of the paper, but the proof is only a citation to the authors' own conference paper [5], and the statement as given is suspect. Model (9) has one shared constraint α_t^T Σ_i β_i ≤ 1, with a single Lagrange multiplier λ_t. The KKT conditions imply μ_i λ_i^t = λ_t, so the per-agent slopes in (10) must be proportional across i with constants independent of t. Condition (10) imposes no such cross-agent restriction. The construction (11) therefore rationalizes each agent under an individual budget α_t^T β_i ≤ α_t^T β_i^t, which is weaker than joint weighted-sum optimality under the shared budget. Collective revealed-preference theory (e.g., [32]) requires additional cross-agent conditions. Since Corollary 2, the null hypothesis (15)-(17), and the DRO slack (20)-(23) all rely on (10), the paper must either supply a complete proof or correct the theorem.
  2. [Section IV.C / VI.D, Eq. (23) and Table I] The stated contribution is minimizing worst-case reconstruction error, but the objective h(ψ,Φ) in (23) is the LP feasibility slack, i.e., the proximity measure of Corollary 2. Section VI evaluates reconstruction error by Hausdorff distance between true and reconstructed Pareto frontiers, and Table I reports improvements in that metric. No theorem or proposition connects the worst-case expected slack to the Hausdorff distance. Thus the numerical improvement is an empirical observation, not a consequence of (23). The authors should either prove a bound between the feasibility slack and the Hausdorff distance, or restate the contribution as minimizing worst-case expected feasibility slack.
  3. [Section III.B, Theorem 2 / Abstract] The phrase 'optimal statistical detector' and 'minimizes Type-I statistical detection error' is stronger than what is proved. Part 3 of Theorem 2 only compares statistics ar Φ with Φ* in [Φ*,Ψ]; the Type-I bound is essentially a p-value calibration for a fixed threshold γ. No optimality in a Neyman-Pearson class is established. The authors should define the class of detectors and the optimality criterion precisely; otherwise the claim should be softened to 'a detector with controlled Type-I error.'
minor comments (5)
  1. [Section III.A, Theorem 1] The theorem statement omits the monotonicity assumption that appears in Definition 3, and the proof later assumes β_i^t > 0. These assumptions should be stated explicitly in the theorem.
  2. [Appendix B, Lemma 1 proof] The proof is garbled in places, e.g., the set-builder notation `fβi_tg_{i≠j}[f \barβ^k_t g` is not well-formed, and the line beginning 'Now take' appears to be incorrect. Please rewrite the proof cleanly.
  3. [Section VI.C, Figure 5] The plotted statistic is constant at 1 over the displayed noise range. The text discusses Type-II error, but the figure would be more informative with a larger noise range or a panel showing the threshold behavior.
  4. [Appendix C, proof of Theorem 2] In the first paragraph, 'substituting β_i^t = β_i^t − ε_i^t' should read 'substituting \tilde β_i^t = β_i^t + ε_i^t'; as written it is a typo.
  5. [Section IV.D, Eq. (24)] The symbol V is used both for the scalar bound and for the set of auxiliary variables; this notation clash should be fixed.

Circularity Check

3 steps flagged

Theorem 1's per-agent Afriat inequalities are the load-bearing foundation for both the detector and the DRO, yet their proof is deferred entirely to the authors' own prior paper [5]; the claimed equivalence with the shared-constraint coordination problem is thus imported via self-citation rather than derived.

specific steps
  1. self citation load bearing [Section III.A.1, Theorem 1]
    "THEOREM 1 Let D be a set of observations. The following are equivalent: 1) there exist ... (9) 2) there exist numbers u_i^j > 0, λ_i^j > 0 such that ... (10). Proof: See Theorem 1 of [5]"

    The paper's central premise—that Pareto-efficient coordination (9) is equivalent to the per-agent linear program (10)—is not proved here. The proof is deferred to [5], the authors' own ICASSP paper. This equivalence is then used to define the proximity metric (12), the null hypothesis of the statistical detector (15)-(17), and the DRO objective (23). Since (10) contains no reference to the shared constraint α_t^T Σ_i β_i ≤ 1, the claimed equivalence with the joint optimization is not self-evident and is not independently verified in this paper. Any error in [5] propagates to every downstream guarantee.

  2. self citation load bearing [Section III.A.2, Corollary 1]
    "COROLLARY 1 Given constants u_i^t, λ_i^t, t ∈ [T], i ∈ [M] which make (10) feasible, explicit monotone and continuous objective functions that 'rationalize' the dataset ... are given by (11) ... i.e., (9) is satisfied. Proof: See Lemma 1 of [5]."

    The utility reconstruction formula (11) is asserted to satisfy the joint weighted-sum optimality condition (9), and this assertion is again justified solely by a lemma from the authors' prior work [5]. This construction is used in Algorithm 1 to predict future system responses and serves as the baseline for the robust reconstruction in Section IV. Thus the reconstruction step inherits the unverified equivalence of Theorem 1 and adds a second self-citation in place of an independent proof.

  3. other [Section IV.C, DRO formulation, after equation (23)]
    "The objective (23) finds the set of parameters ψ which minimizes the worst-case expected proximity to feasibility ... Thus, when compared to the naive estimation procedure (22), (23) will better approximate the true utility functions in the worst case, making (23) a robust estimation procedure."

    The DRO objective h(ψ, Φ) is exactly the same feasibility slack that defines proximity to Pareto-efficiency in (20). The paper then interprets minimizing this slack as 'better approximating the true utility functions in the worst case.' But the external reconstruction metric used in the numerical section is the Hausdorff distance between Pareto surfaces, which is not the same as the slack objective. No theorem in the paper connects worst-case slack minimization to worst-case Hausdorff error; the claimed robustness is therefore asserted through the surrogate objective, with the actual metric only checked in simulation. This is a surrogate-objective gap rather than a fitted prediction, but it is a missing support for the paper's main DRO claim.

full rationale

The statistical detector (Theorem 2) and the SIP reformulation (Theorem 3) are internally self-consistent: the Type-I error bound is a standard super-uniform p-value argument given the null hypothesis defined by (10), and the Wasserstein reformulation is an application of known duality. These are not circular. The circularity concern is concentrated at the base of the derivation chain: the equivalence (10)⇔(9) is the single load-bearing premise for the detector, the proximity measure, and the DRO objective, and its proof is deferred to the authors' own [5]. On inspection, (10) is the standard per-agent Afriat inequality and does not contain the shared budget α_t^⊤Σ_i β_i ≤ 1, so the claimed necessity/sufficiency for the joint problem is not self-evident and is not independently established here. Corollary 1's reconstruction is likewise imported from [5] and inherits this gap. The DRO section also equates minimizing worst-case feasibility slack with minimizing worst-case reconstruction error, while the actual Hausdorff validation is empirical. The paper does contain substantial independent content—the detector calibration, the SIP reduction, the exchange-method algorithm, and the UAV case study—so the score is 4 rather than 6 or 8. The core issue is a load-bearing self-citation chain, not a fitted parameter renamed as a prediction.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

The central claim pulls several results from prior publications without derivation: Theorem 1 and Corollaries from the authors' own [5]; the spectral-constraint equivalence from [21]; the SIP reformulation from [38]; and exchange-method convergence from [35]. The paper contributes an application of these pieces plus an assumed-noise statistical calibration. No new physical entities are postulated; the free parameters are user-chosen algorithm settings and the assumed noise model.

free parameters (4)
  • Wasserstein radius epsilon = 0.2 in experiments
    User-chosen robustness level; no sensitivity analysis reported; directly changes the objective and all DRO results.
  • Exchange method tolerance delta = 0.1
    Stopping tolerance in Algorithm 3; chosen by hand, not derived.
  • Detection threshold gamma = not specified in experiments
    Free parameter in the detector (17); Type-I error control depends on it but experiments do not sweep it.
  • Noise covariance sigma^2 = 0.01 to 0.1 for detector; 0.5 to 3 for DRO
    The noise model Lambda_t is assumed known (Gaussian); the Type-I guarantee and all simulations depend on this assumption.
axioms (7)
  • domain assumption Agents' utility functions are concave, continuous, and monotone increasing (Definition 3)
    Assumed throughout; needed for weighted-sum scalarization to cover the Pareto frontier via the known inclusion (6).
  • domain assumption Theorem 1 equivalence: per-agent Afriat inequalities (10) characterize joint Pareto efficient coordination (8)
    Proof deferred to self-cited [5]; the entire detector and DRO build on this equivalence without re-derivation.
  • domain assumption Noise distribution Lambda_t is known exactly for the Type-I guarantee
    Theorem 2 uses the cdf of Psi under the true noise distribution; the paper acknowledges approximation in practice, which breaks exact Type-I control.
  • ad hoc to paper Assumptions 1-3: finite support noise, lower bounded probes, bounded LP parameter set
    Imposed to make h(psi,Phi) bounded and Lipschitz continuous, needed to apply Corollary 3.8 of [38].
  • standard math Corollary 3.8 of [38] reformulates Wasserstein DRO as a semi-infinite program
    External result, not proved in the paper.
  • standard math Exchange method convergence from [35]
    External result used to assert Algorithm 3 terminates with a delta-optimal solution.
  • ad hoc to paper Lemma 3 of [21] maps the spectral covariance bound to the linear inequality alpha^T beta <= 1
    Self-cited prior result used for Proposition 1 and Theorem 4 in the UAV application.

pith-pipeline@v1.3.0-alltime-deepseek · 27964 in / 21075 out tokens · 225303 ms · 2026-08-04T19:56:06.267145+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Multi-Agent Inverse Reinforcement Learning for Identifying Pareto-Efficient Coordination -- A Distributionally Robust Approach." pith.science (2026). https://pith.science/paper/KWJLYYN2

@misc{pith2026250908956,
  author       = {Pith},
  title        = {Pith review of: Multi-Agent Inverse Reinforcement Learning for Identifying Pareto-Efficient Coordination -- A Distributionally Robust Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KWJLYYN2}},
  note         = {Machine review of arXiv:2509.08956}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-agent inverse reinforcement learning (IRL) aims to identify Pareto-efficient behavior in a multi-agent system, and reconstruct utility functions of the individual agents. Motivated by the problem of detecting UAV coordination, how can we construct a statistical detector for Pareto-efficient behavior given noisy measurements of the decisions of a multi-agent system? This paper approaches this IRL problem by deriving necessary and sufficient conditions for a dataset of multi-agent system dynamics to be consistent with Pareto-efficient coordination, and providing algorithms for recovering utility functions which are consistent with the system dynamics. We derive an optimal statistical detector for determining Pareto-efficient coordination from noisy system measurements, which minimizes Type-I statistical detection error. Then, we provide a utility estimation algorithm which minimizes the worst-case estimation error over a statistical ambiguity set centered at empirical observations; this min-max solution achieves distributionally robust IRL, which is crucial in adversarial strategic interactions. We illustrate these results in a detailed example for detecting Pareto-efficient coordination among multiple UAVs given noisy measurement recorded at a radar. We then reconstruct the utility functions of the UAVs in a distributionally robust sense.

Figures

Figures reproduced from arXiv: 2509.08956 by Luke Snow, Vikram Krishnamurthy.

Figure 1
Figure 1. Figure 1: Radar – Multi-UAV Interaction. We represent the high-level radar tracking waveform (parameters) by αt, and the target network maneuvers by {β i t }i∈[M] . The analyst observes the dataset D = {αt, {β i t }M i=1, t ∈ [0, T]} and has two goals. First, to determine whether the dataset is consistent with Pareto-efficient coordination. Second, to produce distributionally robust estimates of the utility function… view at source ↗
Figure 2
Figure 2. Figure 2: Pareto-efficiency in objective space for a 2-objective optimization. The interior (blue) is the feasible image f(Xc t ). Points on the upper boundary (red) are Pareto-efficient: improving one objective necessarily worsens the other. Black points on the frontier illustrate efficient solutions, while any points in the interior are dominated. The aim of this paper is to detect Pareto-efficiency given a datase… view at source ↗
Figure 3
Figure 3. Figure 3: Proberesponse revealed preference framework. The environment issues a probe αt, the system responds optimally with {β i t }, and the analyst observes (αt, {β i t }). Testing coordination reduces to checking LP feasibility (10). computing the constrained multi-objective optimal solu￾tions (8). 8 3. Predicting Future Responses Having reconstructed rationalizing utilities via Corol￾lary 1, we can predict how … view at source ↗
Figure 4
Figure 4. Figure 4: f i (β) is the true objective function of the i’th target, inducing the responses {βˆi t } 10 t=1. Ui (β) is the reconstructed objective function for radar i, computed using the deterministic dataset β and (11). 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 coordinating network non-coordinating network [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Statistic 1 − FˆΨ(Φ∗(β¯)) as a function of variance of the noise distribution Λt. Error bars represent one standard deviation within the dataset produced by 300 Monte-Carlo simulations. Higher 1 − FˆΨ(Φ∗(β¯)) corresponds to higher likelihood of radar network coordination in the statistical detector (17). Let us interpret the simulation results displayed in [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Convergence of CV in Algorithm 3, with respect to varying Wasserstein-radii ϵ. Algorithm 3 terminates when CV < δ, and produces a δ-optimal solution of (23). It can be seen that Algorithm 3 produces a δ = 0.1-optimal solution within 10 iterations for each Wasserstein proximity ϵ. estimation, without sacrificing average performance. The distributional robustness is apparent from the reduced worst-case error… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 3 linked inside Pith

  1. [1]

    UA V swarm intelligence: Recent advances and future trends,

    Y . Zhou, B. Rao, and W. Wang, “UA V swarm intelligence: Recent advances and future trends,” IEEE Access , vol. 8, pp. 183 856– 183 878, 2020

    2020

  2. [2]

    UA V communication networks issues: A review,

    H. Nawaz, H. M. Ali, and A. A. Laghari, “UA V communication networks issues: A review,” Archives of Computational Methods in Engineering , vol. 28, no. 3, pp. 1349–1369, 2021

    2021

  3. [3]

    P . E. Pace, Detecting and classifying low probability of intercept radar. Artech house, 2009

    2009

  4. [4]

    Electronic warfare systems,

    A. E. Spezio, “Electronic warfare systems,” IEEE Transactions on Microwave Theory and Techniques , vol. 50, no. 3, pp. 633–644, 2002

    2002

  5. [5]

    Identifying coor- dination in a cognitive radar network–a multi-objective inverse reinforcement learning approach,

    L. Snow, V . Krishnamurthy, and B. M. Sadler, “Identifying coor- dination in a cognitive radar network–a multi-objective inverse reinforcement learning approach,” International Conference on Acoustics, Speech, and Signal Processing , 2022

    2022

  6. [6]

    Statistical detection of coordination in a cognitive radar network through inverse multi-objective optimization,

    ——, “Statistical detection of coordination in a cognitive radar network through inverse multi-objective optimization,” in 2023 26th International Conference on Information Fusion (FU- SION). IEEE, 2023, pp. 1–8

    2023

  7. [7]

    Revealed preference,

    H. R. V arian et al. , “Revealed preference,” Samuelsonian eco- nomics and the twenty-first century , pp. 99–115, 2006

    2006

  8. [8]

    C. P . Chambers and F. Echenique, Revealed preference theory . Cambridge University Press, 2016, no. 56

    2016

  9. [9]

    The construction of utility functions from expendi- ture data,

    S. N. Afriat, “The construction of utility functions from expendi- ture data,” International Economic Review , vol. 8, no. 1, pp. 67–77, 1967

    1967

  10. [10]

    Distributionally robust optimiza- tion: A review,

    H. Rahimian and S. Mehrotra, “Distributionally robust optimiza- tion: A review,” arXiv preprint arXiv:1908.05659 , 2019

    Pith/arXiv arXiv 1908

  11. [11]

    Frameworks and results in distributionally robust optimiza- tion,

    ——, “Frameworks and results in distributionally robust optimiza- tion,” Open Journal of Mathematical Optimization , vol. 3, pp. 1–85, 2022

    2022

  12. [12]

    Wasserstein distributionally robust optimization: The- ory and applications in machine learning,

    D. Kuhn, P . M. Esfahani, V . A. Nguyen, and S. Shafieezadeh- Abadeh, “Wasserstein distributionally robust optimization: The- ory and applications in machine learning,” in Operations re- search & management science in the age of analytics . Informs, 2019, pp. 130–166

    2019

  13. [13]

    Thirty years of machine learning: The road to Pareto-optimal wireless networks,

    J. Wang, C. Jiang, H. Zhang, Y . Ren, K.-C. Chen, and L. Hanzo, “Thirty years of machine learning: The road to Pareto-optimal wireless networks,” IEEE Communications Surveys & Tutorials , vol. 22, no. 3, pp. 1472–1514, 2020

    2020

  14. [14]

    Pruning and ranking the Pareto optimal set, application for the dynamic multi-objective network design problem,

    L. J. Wismans, T. Brands, E. C. V an Berkum, and M. C. Bliemer, “Pruning and ranking the Pareto optimal set, application for the dynamic multi-objective network design problem,” Journal of advanced transportation , vol. 48, no. 6, pp. 588–607, 2014

    2014

  15. [15]

    Energy and age Pareto optimal trajectories in UA V-assisted wireless data collection,

    Y . Liao and V . Friderikos, “Energy and age Pareto optimal trajectories in UA V-assisted wireless data collection,” IEEE Transactions on V ehicular Technology, vol. 71, no. 8, pp. 9101– 9106, 2022

    2022

  16. [16]

    Topology control of unmanned aerial vehicle (UA V) mesh networks: A multi-objective evolutionary algorithm approach,

    S. Sabino and A. Grilo, “Topology control of unmanned aerial vehicle (UA V) mesh networks: A multi-objective evolutionary algorithm approach,” in Proceedings of the 4th ACM Workshop on Micro Aerial V ehicle Networks, Systems, and Applications , 2018, pp. 45–50

    2018

  17. [17]

    Multi-UA V task allo- cation based on improved algorithm of multi-objective parti- cle swarm optimization,

    Y . Gao, Y . Zhang, S. Zhu, and Y . Sun, “Multi-UA V task allo- cation based on improved algorithm of multi-objective parti- cle swarm optimization,” in 2018 International conference on cyber-enabled distributed computing and knowledge discovery (CyberC). IEEE, 2018, pp. 443–447

    2018

  18. [18]

    Pareto optimal pid tuning for px4-based unmanned aerial vehicles by using a multi-objective particle swarm optimization algorithm,

    V . Gomez, N. Gomez, J. Rodas, E. Paiva, M. Saad, and R. Gregor, “Pareto optimal pid tuning for px4-based unmanned aerial vehicles by using a multi-objective particle swarm optimization algorithm,” Aerospace, vol. 7, no. 6, p. 71, 2020

    2020

  19. [19]

    Achieving Pareto opti- mality through distributed learning,

    J. R. Marden, H. P . Y oung, and L. Y . Pao, “Achieving Pareto opti- mality through distributed learning,” SIAM Journal on Control and Optimization , vol. 52, no. 5, pp. 2753–2770, 2014

    2014

  20. [20]

    Multi- objective multi-agent decision making: a utility-based analysis and survey,

    R. R ˘adulescu, P . Mannion, D. M. Roijers, and A. Nowé, “Multi- objective multi-agent decision making: a utility-based analysis and survey,” Autonomous Agents and Multi-Agent Systems , vol. 34, no. 1, p. 10, 2020

    2020

  21. [21]

    Iden- tifying cognitive radars-inverse reinforcement learning using revealed preferences,

    V . Krishnamurthy, D. Angley, R. Evans, and B. Moran, “Iden- tifying cognitive radars-inverse reinforcement learning using revealed preferences,” IEEE Transactions on Signal Processing , vol. 68, pp. 4529–4542, 2020

    2020

  22. [22]

    Meta-cognition. an inverse-inverse reinforcement learning approach for cogni- tive radars,

    K. Pattanayak, V . Krishnamurthy, and C. Berry, “Meta-cognition. an inverse-inverse reinforcement learning approach for cogni- tive radars,” in 2022 25th International Conference on Informa- tion Fusion (FUSION) . IEEE, 2022, pp. 01–08

    2022

  23. [23]

    How can a cognitive radar mask its cognition?

    ——, “How can a cognitive radar mask its cognition?” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 5897–5901

    2022

  24. [24]

    Maximum en- tropy deep inverse reinforcement learning,

    M. Wulfmeier, P . Ondruska, and I. Posner, “Maximum en- tropy deep inverse reinforcement learning,” arXiv preprint arXiv:1507.04888, 2015

    Pith/arXiv arXiv 2015

  25. [25]

    Large-scale cost function learning for path planning using deep inverse reinforcement learning,

    M. Wulfmeier, D. Rao, D. Z. Wang, P . Ondruska, and I. Posner, “Large-scale cost function learning for path planning using deep inverse reinforcement learning,” The International Journal of Robotics Research, vol. 36, no. 10, pp. 1073–1087, 2017

    2017

  26. [26]

    Inferring group intent as a cooperative game. an nlp-based framework for trajectory analysis,

    Y . Zhang, V . Krishnamurthy, and S. Jain, “Inferring group intent as a cooperative game. an nlp-based framework for trajectory analysis,” 2025

    2025

  27. [27]

    The computational algorithm for the para- metric objective function,

    S. Gass and T. Saaty, “The computational algorithm for the para- metric objective function,” Naval Research Logistics Quarterly , vol. 2, no. 1-2, pp. 39–45, 1955

    1955

  28. [28]

    Miettinen, Nonlinear multiobjective optimization

    K. Miettinen, Nonlinear multiobjective optimization . Springer Science & Business Media, 1999, vol. 12

    1999

  29. [29]

    A CA V platoon control method for isolated intersections: Guaranteed feasible multi-objective approach with priority,

    C. Wang, Y . Dai, and J. Xia, “A CA V platoon control method for isolated intersections: Guaranteed feasible multi-objective approach with priority,” Energies, vol. 13, no. 3, p. 625, 2020

    2020

  30. [30]

    Multi-objective optimization framework for enhancing efficiency and sustainability in smart grids,

    R. Revathi, N. Senthilnathan, V . K. Chinnaiyan, and J. S. Ra- jesh, “Multi-objective optimization framework for enhancing efficiency and sustainability in smart grids,” Energy Conversion and Management , vol. 341, p. 120079, 2025

    2025

  31. [31]

    Revealed stochastic preference: a synthesis,

    D. L. McFadden, “Revealed stochastic preference: a synthesis,” in Rationality and Equilibrium: A Symposium in Honor of Marcel K. Richter . Springer, 2006, pp. 1–20

    2006

  32. [32]

    The revealed pref- erence approach to collective consumption behaviour: Testing and sharing rule recovery,

    L. Cherchye, B. De Rock, and F. V ermeulen, “The revealed pref- erence approach to collective consumption behaviour: Testing and sharing rule recovery,” The Review of Economic Studies , vol. 78, no. 1, pp. 176–198, 2011

    2011

  33. [33]

    Recent advances in robust optimization: An overview,

    V . Gabrel, C. Murat, and A. Thiele, “Recent advances in robust optimization: An overview,” European journal of operational research, vol. 235, no. 3, pp. 471–483, 2014

    2014

  34. [34]

    Distributionally robust optimization: A review on theory and applications,

    F. Lin, X. Fang, and Z. Gao, “Distributionally robust optimization: A review on theory and applications,” Numerical Algebra, Control and Optimization , vol. 12, no. 1, pp. 159–212, 2022. : 15

    2022

  35. [35]

    Wasserstein distributionally robust inverse multiobjective optimization,

    C. Dong and B. Zeng, “Wasserstein distributionally robust inverse multiobjective optimization,” in Proceedings of the AAAI Con- ference on Artificial Intelligence , vol. 35, no. 7, 2021, pp. 5914– 5921

    2021

  36. [36]

    Adaptive distributionally robust optimization,

    D. Bertsimas, M. Sim, and M. Zhang, “Adaptive distributionally robust optimization,” Management Science , vol. 65, no. 2, pp. 604–618, 2019

    2019

  37. [37]

    Wasserstein distributionally robust look-ahead economic dispatch,

    B. K. Poolla, A. R. Hota, S. Bolognani, D. S. Callaway, and A. Cherukuri, “Wasserstein distributionally robust look-ahead economic dispatch,” IEEE Transactions on Power Systems , vol. 36, no. 3, pp. 2010–2022, 2020

    2010

  38. [38]

    Decomposition algorithm for distribu- tionally robust optimization using wasserstein metric,

    F. Luo and S. Mehrotra, “Decomposition algorithm for distribu- tionally robust optimization using wasserstein metric,” arXiv preprint arXiv:1704.03920 , 2017

    Pith/arXiv arXiv 2017

  39. [39]

    Semi-infinite programming: the- ory, methods, and applications,

    R. Hettich and K. O. Kortanek, “Semi-infinite programming: the- ory, methods, and applications,” SIAM review , vol. 35, no. 3, pp. 380–429, 1993

    1993

  40. [40]

    Cutting-plane training of structural svms,

    T. Joachims, T. Finley, and C.-N. J. Y u, “Cutting-plane training of structural svms,” Machine learning , vol. 77, pp. 27–59, 2009

    2009

  41. [41]

    Power allocation for target detection in radar networks based on low probability of intercept: A cooperative game theoretical strategy,

    C. Shi, S. Salous, F. Wang, and J. Zhou, “Power allocation for target detection in radar networks based on low probability of intercept: A cooperative game theoretical strategy,” Radio Science, vol. 52, no. 8, pp. 1030–1045, 2017

    2017

  42. [42]

    Hiding unmanned aerial vehicles for wireless transmissions by covert communications,

    S. Y an, S. V . Hanly, I. B. Collings, and D. L. Goeckel, “Hiding unmanned aerial vehicles for wireless transmissions by covert communications,” in ICC 2019-2019 IEEE International Con- ference on Communications (ICC) . IEEE, 2019, pp. 1–6

    2019

  43. [43]

    UA V-enabled covert wireless data collection,

    X. Zhou, S. Y an, F. Shu, R. Chen, and J. Li, “UA V-enabled covert wireless data collection,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 11, pp. 3348–3362, 2021

    2021

  44. [44]

    UA V-enabled covert federated learning,

    X. Hou, J. Wang, C. Jiang, X. Zhang, Y . Ren, and M. Debbah, “UA V-enabled covert federated learning,”IEEE Transactions on Wireless Communications, vol. 22, no. 10, pp. 6793–6809, 2023

    2023

  45. [45]

    Survey of maneuvering target tracking. part i. dynamic models,

    X. R. Li and V . P . Jilkov, “Survey of maneuvering target tracking. part i. dynamic models,” IEEE Transactions on aerospace and electronic systems , vol. 39, no. 4, pp. 1333–1364, 2003

    2003

  46. [46]

    Bar-Shalom, X

    Y . Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with applications to tracking and navigation: theory algorithms and software. John Wiley & Sons, 2004

    2004

  47. [47]

    Multi-target tracking using joint probabilistic data association,

    T. E. Fortmann, Y . Bar-Shalom, and M. Scheffe, “Multi-target tracking using joint probabilistic data association,” in 1980 19th IEEE Conference on Decision and Control including the Symposium on Adaptive Processes . IEEE, 1980, pp. 807–812

    1980

  48. [48]

    Joint probabilistic data association revisited,

    S. H. Rezatofighi, A. Milan, Z. Zhang, Q. Shi, A. Dick, and I. Reid, “Joint probabilistic data association revisited,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 3047–3055

    2015

  49. [49]

    H. L. V an Trees,Detection, estimation, and modulation theory, part I: detection, estimation, and linear modulation theory . John Wiley & Sons, 2004

    2004

  50. [50]

    Optimal waveform selection for tracking systems,

    D. J. Kershaw and R. J. Evans, “Optimal waveform selection for tracking systems,” IEEE Transactions on Information Theory , vol. 40, no. 5, pp. 1536–1550, 1994

    1994

  51. [51]

    Bar-Shalom and X.-R

    Y . Bar-Shalom and X.-R. Li, Multitarget-multisensor tracking: principles and techniques . YBs Storrs, CT, 1995, vol. 19. VIII. Appendix A. Kalman Filter Dynamics Consider the linear Gaussian dynamics ( 31), ( 32). Based on observations yi 1,...,y i k associated to target i, the tracking functionality in the radar computes the target i state posterior πi ...

    1995

  52. [52]

    collective rationality

    Extracting a Covertness Bound By Lemma 3 of [ 21], we can represent a limit ¯Σ−1 on the radar’s precision of target i measurement, Σ∗−1 t (αt,βi t) as the simple linear inequality α⊤ t βi t 1, i.e., α⊤ t βi t 1() Σ∗−1 t (αt,βi t) ¯Σ−1 where the constant 1 bound is taken without loss of gener- ality. The key idea behind this equivalence is to show the asym...

    2025

  53. [53]

    Triangular Pulse - Continuous Wave s(t) = {√ 3 2θ ( 1 |t| θ ) θ<t<θ 0 otherwise R(α) = [ c2θ2 12η 0 0 5c2 2ω2cθ2η ]

  54. [54]

    Gaussian Pulse - Continuous Wave s(t) = ( 1 πθ2 )1/4 exp (t2 2θ2 ) R(α) = [ c2θ2 sη 0 0 c2 2ω2cθ2η ]

  55. [55]

    Gaussian Pulse - Linear Frequency Modulation Chirp s(t) = ( 1 πθ2 1 )1/4 exp ( ( 1 2θ2 1 jθ2 ) t2 ) R(α) =   c2θ2 1 2η −c2θ2θ2 1 ωcη −c2θ2θ2 1 ωcη c2 ω2cη ( 1 2θ2 1 + 2θ2 2θ2 1 )   E. Multi-T arget Filtering: Joint Probabilistic Data Association Filter The joint probabilistic data association filter (JPDAF) operates under the regime where n measurement...

  56. [56]

    none of the targets

    Filter Dynamics Define the empirical validation matrix Ω = [ ωjt,j 2 [n],t 2 f0,...,m g, with ωjt = 1 if measurement j is in the validation gate of target t, and 0 otherwise. It is common to let the t = 0 index correspond to "none of the targets". Now we construct an object θ known as the "joint association event", as θ = m∩ j=1 θjtj where - θjt represents...

  57. [57]

    a measurement is associated to only one source, m∑ t=0 ˆωjt(θ) = 1, 8j2 [n] (42)

  58. [58]

    target detection indicator

    at most one measurement originates from each target, δt(θ) := n∑ j=1 ˆωjt(θ) 1, 8t2 [m] (43) Denote by Θ the set of all feasible events. The binary variable δt(θ) is known as the "target detection indicator" since it indicates whether, in event θ, a measurement j has been associated to target t. We may also define a "measurement association indicator" τj(θ...

    2025

  59. [59]

    Extracting a Revealed Preference Bound The crucial observation is that, as in the Kalman filter algebraic Riccati equation ( 33), the covariance ( 46) is monotone decreasing in βi t for all i, since this corre- sponds to increasing ˆRk for fixed k. Thus, the asymptotic measurement precision (inverse of asymptotic predicted covariance) is monotone non-decrea...

    2025