Pith. sign in

REVIEW 4 major objections 6 minor 38 references

Enhancing Swarms Durability to Threats via Graph Signal Processing and GNN-based Generative Modeling

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that swarm formations trade detection risk against post-detection survival, and that a GNN generative model can find shapes—most notably a kite—that improve both sides.

desk verdict A genuinely new trade-off and a useful generative framework, but the theory linking diffusion to extinction has a real exponent error that needs fixing before the paper's central proxy claim is taken at face value. read the letter →

arxiv 2507.03039 v1 pith:YHHYTPYH submitted 2025-07-03 q-bio.QM cs.LGphysics.bio-ph

Jonathan Karin , Zoe Piran , Mor Nitzan This is my paper · ORCID
classification q-bio.QMcs.LGphysics.bio-ph
keywords swarmbehaviorgraphsignalprocessingdetectability-durabilitytrade-offdomainofdangerdiffusionneuralnetworksgenerativedesignpredationmodeling
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

Treating a swarm as a graph, this paper claims that every spatial formation is a compromise between two threats: staying compact enough to avoid detection and staying spread out enough to slow a predator once detected. It defines detectability by the domain of danger, the union of detection disks around agents, and durability by how easily a heat signal diffuses through the swarm graph, with faster diffusion meaning faster depletion. Those two quantities both depend on pairwise agent distances, so compactness improves one and worsens the other; the paper calls this the detectability-durability trade-off. The authors then build SWAGEN, a graph-neural-network generator whose loss function makes both quantities differentiable and optimizes them together, and report that the optima cluster into a 'kite' formation—a dense block with a curved tail—that improves both detection avoidance and post-detection survival in simulations.

What carries the argument

Two linked objects carry the argument. First, the domain of danger (DOD): the union of radius-$\rho$ disks around agents, whose relative area sets detection probability. Second, the diffusion evaluator: an impulse $\delta_i$ is propagated through the heat filter $H = e^{-\tau \tilde{L}}$ on the normalized graph Laplacian, and the coefficient of variation $\mathrm{CV} = \sigma/\mu$ of the smoothed signal measures how easily a perturbation spreads, with lower CV meaning faster spread and, per the paper, faster predation. The Gaussian heat-kernel approximation $K_{ij}(\tau)\propto \exp(-d(i,j)^2/(4\tau^2))$ is what connects diffusion time to inter-agent distance and makes the trade-off geometric. SWAGEN converts both quantities into differentiable loss terms, with the DOD approximated by a smooth sigmoid count of grid points covered by agent disks.

What would settle it

Run the authors' own predator simulation on a diverse collection of swarm graphs—random planar graphs, lattices, and elongated shapes, not only chain-like swarms—and compute the Spearman correlation between the CV-based diffusion loss and mean simulated time to extinction; if that correlation is not close to 0.985, or if the ordering of configurations changes, the diffusion proxy is not the quantity that determines survival. A second test is to replace the predator's nearest-agent targeting rule with a centroid-seeking rule and check whether the kite still outperforms the rectangle and arrow.

Watch

Extended reading notes

Core claim

The central claim is that a predator's attack can be modeled as a signal propagating through the swarm graph, and that two graph quantities—the relative area of the domain of danger, $D = \bigcup_i B(c_i,\rho)$, and the coefficient of variation of a heat-kernel impulse response—jointly determine a swarm's fate. The paper proves that detection probability is proportional to the DOD's relative area and argues, via the Gaussian approximation of the heat kernel, that the minimal time to extinction scales with graph distances, so slow diffusion means long survival. Because both quantities are driven by the same inter-agent spacings, no static formation can simultaneously minimize them; the authors present this as an inherent trade-off, confirm it in simulations of v-formations, arrows, and rectangles, and use it as the objective for SWAGEN. The generator's optimized outputs are best matched by a kite-shaped template (by Gromov-Wasserstein distance), and simulations give the kite the best mean survival rate, 71.14%, versus 64–68% for the three baseline shapes.

Load-bearing premise

The load-bearing premise is that the coefficient of variation of a heat-kernel impulse response faithfully measures how quickly a predator depletes a detected swarm, with the paper's direct support being a distance-scaling argument and a Spearman correlation of 0.985 on synthetic chain-like swarms rather than a derivation from predator dynamics.

Editorial extensions

If this is right

  • Static swarm formations are necessarily compromises: improving detection avoidance via compaction makes a detected swarm easier to deplete, so design tasks must specify which side of the trade-off matters more.
  • The two-term DOD-diffusion plane gives a cheap screening tool: candidate formations can be ranked without full predation simulations.
  • Because both loss terms are differentiable, a GNN generator can explore the formation landscape automatically; the paper reports that repeated optimization converges to the kite motif rather than to any baseline shape.
  • Under the paper's predation model, the kite raises mean survival to 71.14% versus 66.5% (v-formation), 64.42% (arrow), and 67.61% (rectangle), while also being the least detected configuration in its simulations.
  • Extreme weighting of either loss term reproduces the trade-off's endpoints: pure DOD minimization collapses agents to one point, and pure diffusion minimization pushes agents apart into disconnected components.

Reading between the lines

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

  • My inference: the same two-term decomposition should transfer to non-predator perturbations such as wind gusts, jamming, or communication loss by keeping the DOD machinery and swapping the signal model; the paper mentions such extensions but does not test them.
  • My inference: the kite's structure suggests a design principle the paper does not state explicitly—pack a dense core to shrink the detection footprint and add a curved trailing filament to stretch the predator's travel path—which would predict that other 'core-plus-tail' formations perform similarly.
  • My inference: since both objective terms are smooth functions of positions, the optimization could be swept across trade-off weights to trace a full Pareto front of formations; the paper shows only the endpoints and the recurring kite, leaving the intermediate family uncharacterized.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper models animal or robotic swarms as unit-disk graphs and uses graph signal processing to relate a swarm's spatial configuration to two quantities: the domain of danger (DOD), which controls the probability of being detected, and the graph diffusion of a heat-kernel impulse response, which is proposed as a measure of how quickly a predator can deplete the swarm after detection. It argues that these two quantities are in tension, yielding a 'detectability-durability trade-off', and provides supporting evidence from predator-prey simulations for three canonical shapes (v-formation, arrow, rectangle). It then introduces SwaGen, a GNN-based variational generative model whose loss combines a DOD term and a diffusion term, and reports that the model discovers a new 'kite' motif that simultaneously improves both objectives, with improved survival in independent simulations. The paper includes proofs in Appendix B, an ablation-style study of loss weighting, stochasticity robustness checks, and publicly released code.

Significance. If the central claims hold, the detectability-durability trade-off is a conceptually valuable organizing principle for swarm design, and SwaGen provides a flexible generative framework for task-specific swarm configurations. The paper's strengths include a clear empirical demonstration of the trade-off in Figure 3b with robustness across stochastic variants (Figure A13), a publicly available implementation, and a novel GNN-based generative approach with a differentiable surrogate for the trade-off. However, the theoretical foundation of the durability proxy contains a mathematical error (the heat-kernel Gaussian approximation), and the proxy is validated only on a narrow class of chain-like swarms. Because the SwaGen diffusion loss and the kite's claimed advantage rest on this proxy, the current manuscript cannot be accepted without substantial revision and additional validation.

major comments (4)
  1. [§2.3.1 and Appendix B.3] The Gaussian heat-kernel approximation is written as K_{i,j}(τ)=(4πτ)^{-n/2} exp(-d(i,j)^2/(4τ^2)), with τ^2 in the denominator and n described as 'the number of nodes in the graph'. The standard heat kernel on R^n is (4πτ)^{-n/2} exp(-d(i,j)^2/(4τ)), with n the spatial dimension. This is not a typographical detail: the authors' version makes the exponent dimensionless at τ∼d, which is exactly the linear scaling used in B.3 to conclude τ_min∼D_min. With the correct kernel, the scaling is τ∼d^2, so the claimed linear equivalence between diffusion time and predator travel time is not established. This derivation is load-bearing because the diffusion loss in §4.1 is justified by connecting the CV of the heat-kernel impulse response to the duration of the predation phase. The authors should correct the kernel, re-derive the scaling, and either prove the CV-to-extinction link or present a different theoretical argument.
  2. [Appendix B.3] The only quantitative support for the durability proxy is a Spearman correlation of 0.985 computed on random chain-like swarms, where the metric compared is the mean minimal extinction time Tmin (a shortest-path quantity) rather than the actual survival outcome from the predator simulation. Section 3 asserts that the diffusion term measures the predator's effectiveness in hunting the swarm, but this link is not validated on the non-chain geometries used elsewhere in the paper (v-formation, arrow, rectangle, kite) nor against actual extinction rates. If the proxy is misaligned, the SwaGen diffusion loss is mis-specified and the kite's advantage may be an artifact. Please add an experiment that correlates the diffusion loss value with the empirically observed survival rate (or extinction probability) across many random configurations and the canonical shapes, using the full predator simulation rather than the minimal travel time.
  3. [Table 3 and Figure 7] The reported improvement of the kite over the rectangle (71.14% vs. 67.61% mean survival) is presented without error bars, confidence intervals, or significance testing. The violin plots in Figure 3a show very large inter-simulation spread (e.g., v-formation 66.5±34.8%, rectangle 67.6±46.8%), so the difference could easily be within stochastic noise. The claim that the kite 'improves the overall survival rate' is central to the paper's applied conclusion and needs statistical support, for example bootstrap confidence intervals or a permutation test over the 250 simulations per configuration. The same issue applies to the comparison of detection and extinction percentages in Figure 7.
  4. [§5.1 and §4.1] The evaluation in Figure 6 shows that SwaGen improves the normalized DOD and diffusion terms relative to random initializations. Since these are precisely the terms being optimized by the loss in §4.1, this outcome is expected by construction and does not by itself demonstrate that the optimized configurations are more durable. The independent predator simulations in Figure 7 provide external anchoring, but the kite's measured DOD and diffusion loss values are not reported in the main text, so the reader cannot see whether the kite's survival benefit actually coincides with better values of the two surrogates. Please report the normalized DOD and diffusion terms for the kite and the three canonical shapes (beyond the qualitative statement that the kite sits above the diagonal in Figure A12), which would also help disentangle a genuine effect from a proxy optimization artifact.
minor comments (6)
  1. [§2.3.1] The symbol n is used both for the number of nodes (denoted N elsewhere) and as the dimension of the Euclidean space in the Gaussian kernel; please rename to avoid confusion.
  2. [§4] The text reads 'acting as regularizes' and should be 'acting as regularizers'.
  3. [Table 3] The table should include standard deviations, standard errors, or confidence intervals for each survival percentage, not just the point estimate.
  4. [§5.2, Table 2] The Gromov-Wasserstein distances (kite 0.0259, arrow 0.0612, rectangle 0.0911) are all quite small; please state what distance would be considered a good match, or provide a baseline such as the GW distance between two independently sampled kite configurations, to make the validation interpretable.
  5. [Figures 2 and A12] The captions for Figure 2 and Figure A12 appear nearly identical; consider merging the figures or clarifying the difference (e.g., one uses N=1000, d=5 and the other presumably also uses the same parameters but with the kite included).
  6. [Throughout] The model name is written inconsistently as both SWAGEN and SwaGen; please choose one spelling and use it consistently.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial circularity: Figure 6 validates SwaGen by the same DOD/diffusion terms it optimizes; the central trade-off is nevertheless anchored by independent predator simulations, and the Gaussian-kernel exponent issue in §2.3.1/B.3 is a correctness risk rather than a circularity.

  1. fitted input called prediction [Section 5.1 (Figure 6, Table 1) vs. the SwaGen loss in Section 4.1]
    "We first consider the DOD and diffusion terms, of random (model input) and optimized (model output) configurations (considering 100 random swarm initializations, see 4.1, A.2.3). Terms are normalized to the range of [0, 1] such that 1 is the optimal value of each–minimal DOD and minimal diffusion. Visualizing these results over a 2D plane (DOD-diffusion), we observe that optimized points (representing spatial configurations) improve the random initializations, according to both trade-off terms, closer to the optimal top right corner (Figure 6, Figure A12, Table 1)."

    The two quantities used for this improvement are exactly the terms βLDOD and γLdiff that SwaGen minimizes in LSwaGen = δLKLD + ϕLrec + βLDOD + γLdiff (Section 4.1). Showing that optimized configurations score better on the optimized objectives mainly confirms that gradient descent minimized the loss; it is not an independent test of detectability or durability. The paper does provide an external anchor in Figure 7 and Table 3, where the kite is evaluated by predator simulations and survival percentages, so the central trade-off does not reduce entirely to the loss. The self-referential Figure 6 evaluation is therefore a partial, not total, circularity.

full rationale

The derivation chain is mostly not circular. Theorem B.1 derives detection probability from the uniform-position assumption and the predator speed bound, and the DOD-diffusion plane in Figure 2 is a computed trade-off between two geometric quantities; neither reduces to a fitted parameter. The durability-to-diffusion link in B.3 is the weakest step: it defines Tmin = Dmin/(vp−vprey), uses a Gaussian-kernel approximation written with a non-standard τ^2 denominator to obtain τ∼Dmin, and then validates the CV proxy with a Spearman 0.985 correlation on random chain-like swarms. That is a mathematical/proxy-validity concern, not an input-output tautology, so it is not scored as circularity here. The one true self-referential step is the Section 5.1 evaluation: SwaGen's output is judged by the DOD and diffusion terms it was trained to minimize, making the Figure 6/Table 1 improvement a check of optimization rather than a prediction. Because Figure 7 and Table 3 use independent predator simulations and the kite configuration also improves mean survival (71.14% vs. 64–68% for baselines), the paper's central claim is externally anchored. Overall score 3 reflects this partial self-evaluation while acknowledging the independent evidence.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central derivations rest on a handful of modeling choices (uniform predator distribution, unit-disk graphs, Gaussian heat kernel approximation) and several hand-set hyperparameters (tau, k, loss weights). The kite motif is a generated configuration, not an externally validated entity.

free parameters (4)
  • Heat diffusion time tau = 50, 100, 150 in DOD-diffusion planes; value in SwaGen loss not reported
    Controls how far a heat signal spreads; the CV diffusion measure and the claimed geometric ordering depend on tau.
  • DOD sigmoid slope k = not reported
    Determines sharpness of the soft union-of-disks approximation in LDOD; affects optimized configurations.
  • SwaGen loss weights beta (DOD) and gamma (diffusion) = baseline unstated; variants use 2x weights in Figure A11
    Balance the two contradictory terms; no principled selection or ablation beyond a 2x sweep.
  • Kite model parameters (block fraction, spacing alpha, curvature f) = block fraction=3/4, spacing alpha, curvature f chosen in Appendix A.2.6
    The 'theoretical' kite is constructed post hoc to match SwaGen outputs; these parameters are hand-picked, not inferred from data.
assumptions (6)
  • domain assumption Predators are uniformly distributed over the grid
    Used in Theorem B.1 to make detection probability proportional to DOD area; real predators may not be uniformly distributed.
  • domain assumption Marginal predation: predator always attacks the closest prey within detection range and kills it
    Inherited from [12,20] and imposed in Section 2.2; the trade-off result depends on this attack rule.
  • domain assumption Swarm graph is a unit disk graph with edge radius equal to interaction range rho
    Section 3 and Appendix A.1.1; all diffusion and DOD calculations use this graph construction.
  • domain assumption Heat kernel on a locally Euclidean graph is approximated by a Gaussian, leading to tau ~ distance scaling
    Section 2.3.1 and Appendix B.3; the stated formula uses exponent -n/2 (number of nodes rather than dimension) and tau squared, so it is an approximation rather than a theorem.
  • ad hoc to paper The coefficient of variation of heat-kernel impulse responses is a valid proxy for predation-phase duration
    This is the load-bearing link used in the SwaGen diffusion loss and in the durability claim; it is supported only by a synthetic correlation in Appendix B.3, not by derivation.
  • domain assumption Static initial configuration determines detectability and durability; dynamic reconfiguration during an attack is ignored in the analytical model
    The analysis and SwaGen loss use initial positions; the paper acknowledges this limitation in the Discussion.
invented entities (1)
  • kite motif
    purpose: a spatial configuration claimed to improve the detectability-durability trade-off relative to V, arrow, and rectangle shapes
    The kite is extracted from SwaGen outputs and validated only within the same simulation/loss framework; no natural swarms or physical experiments are offered as external support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Swarms Durability to Threats via Graph Signal Processing and GNN-based Generative Modeling." pith.science (2026). https://pith.science/paper/YHHYTPYH

@misc{pith2026250703039,
  author       = {Pith},
  title        = {Pith review of: Enhancing Swarms Durability to Threats via Graph Signal Processing and GNN-based Generative Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YHHYTPYH}},
  note         = {Machine review of arXiv:2507.03039}
}
read the original abstract

Swarms, such as schools of fish or drone formations, are prevalent in both natural and engineered systems. While previous works have focused on the social interactions within swarms, the role of external perturbations--such as environmental changes, predators, or communication breakdowns--in affecting swarm stability is not fully understood. Our study addresses this gap by modeling swarms as graphs and applying graph signal processing techniques to analyze perturbations as signals on these graphs. By examining predation, we uncover a "detectability-durability trade-off", demonstrating a tension between a swarm's ability to evade detection and its resilience to predation, once detected. We provide theoretical and empirical evidence for this trade-off, explicitly tying it to properties of the swarm's spatial configuration. Toward task-specific optimized swarms, we introduce SwaGen, a graph neural network-based generative model. We apply SwaGen to resilient swarm generation by defining a task-specific loss function, optimizing the contradicting trade-off terms simultaneously.With this, SwaGen reveals novel spatial configurations, optimizing the trade-off at both ends. Applying the model can guide the design of robust artificial swarms and deepen our understanding of natural swarm dynamics.

Figures

Figures reproduced from arXiv: 2507.03039 by the authors.

Figure 1
Figure 1. An abstract visualization of the GSP approach used to model swarm behavior. a, Initial swarming configuration used in our framework: v-formation (top), arrow (middle), rectangle (bottom). b, The swarming model (see 2.1), along with the introduced graph constructions. Agents are in blue triangles, and orange lines represent the induced swarm graph (edges added in accordance with the interaction range). Characteristic… view at source ↗
Figure 2
Figure 2. The DOD-diffusion plane. The tension between the domain of danger (DOD) and diffusion ability in common spatial configurations of swarms. Evaluation of the DOD (x-axis) and diffusion term (y-axis) normalized to the range of [0, 1]. Evaluation for each spatial configuration is performed using N = 1000 agents and d = 5 (the distance between neighboring agents, see A.1.3). To evaluate the diffusion we computed the mean… view at source ↗
Figure 3
Figure 3. Swarm’s susceptibility to predation as a function of spatial configurations. Results based on 250 simulations for each configuration (v-formation, arrow, rectangle). a, Violin plots of the number of living agents at the end of the simulation for different initial shape structures: v-formation, an arrow, and a rectangle. b, The detectability-durability plane, the percentage of simulations in which the swarm avoided e… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the SWAGEN model architecture. The SWAGEN model input is a swarm represented as a graph with N neighbors and the following per-node features: position, distance and angle from the swarm’s center of mass. The graph is processed through two GraphSAGE [11] lay…
Figure 5
Figure 5. Figure 5: The models’ gradients depict the DOD-diffusion trade-off Scatter plot of the swarm with the direction of the diffu￾sion term gradient (blue) and the DOD term (red) for each agent. The gradients are negatively correlated without completely cancel￾ing, with Pearson corre…
Figure 6
Figure 6. Figure 6: Visualization of the trade-off considering random and optimized swarm spatial configurations. The normalized dif￾fusion (y-axis) compared to the normalized DOD (x-axis). For comparison, values are normalized to the range [0, 1] such that the top-right corner (1, 1) is …
Figure 7
Figure 7. Figure 7: The kite improves the swarms’ durability trade-off. Evaluation of the percentage of simulations in which the swarm avoided detection (none of the agents were predated; x-axis) com￾pared to those that avoided extinction (at least one agent survived; y-axis) for the diff…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

  1. [1]

    Heat kernels, mani- folds and graph embedding

    Xiao Bai and Edwin R Hancock. Heat kernels, mani- folds and graph embedding. In Structural, Syntactic, and Statistical Pattern Recognition: Joint IAPR In- ternational Workshops, SSPR 2004 and SPR 2004, Lisbon, Portugal, August 18-20, 2004. Proceedings , pages 198–206. Springer, 2004

    work page 2004

  2. [2]

    Col- lective response to perturbations in a data-driven fish school model

    Daniel S Calovi, Ugo Lopez, Paul Schuhmacher, Hugues Chat´e, Cl´ement Sire, and Guy Theraulaz. Col- lective response to perturbations in a data-driven fish school model. Journal of The Royal Society Interface, 12(104):20141362, 2015

    work page 2015

  3. [3]

    Modeling bird flight formations using diffusion adaptation

    Federico S Cattivelli and Ali H Sayed. Modeling bird flight formations using diffusion adaptation. IEEE transactions on signal processing, 59(5):2038–2051, 2011

    work page 2011

  4. [4]

    Spatial temporal graph neural net- 9 Enhancing Swarms’ Durability to Threats via GNN-based Generative Modeling works for decentralized control of robot swarms

    Siji Chen, Yanshen Sun, Peihan Li, Lifeng Zhou, and Chang-Tien Lu. Spatial temporal graph neural net- 9 Enhancing Swarms’ Durability to Threats via GNN-based Generative Modeling works for decentralized control of robot swarms. In Proceedings of the 31st ACM International Confer- ence on Advances in Geographic Information Systems, pages 1–4, 2023

    work page 2023

  5. [5]

    A minimal model of predator–swarm interactions

    Yuxin Chen and Theodore Kolokolnikov. A minimal model of predator–swarm interactions. Journal of The Royal Society Interface, 11(94):20131208, 2014

    work page 2014

  6. [6]

    Self-organization and collective behavior in vertebrates

    Iain D Couzin, Jens Krause, et al. Self-organization and collective behavior in vertebrates. Advances in the Study of Behavior, 32(1):10–1016, 2003

    2003

  7. [7]

    Effective leadership and decision- making in animal groups on the move

    Iain D Couzin, Jens Krause, Nigel R Franks, and Simon A Levin. Effective leadership and decision- making in animal groups on the move. Nature, 433(7025):513–516, 2005

    work page 2005

  8. [8]

    Pygsp: Graph signal processing in python (v0.5.0), 2017

    Micha¨el Defferrard, Lionel Martin, Rodrigo Pena, and Nathana¨el Perraudin. Pygsp: Graph signal processing in python (v0.5.0), 2017

    work page 2017

Show all 38 references
  1. [9]

    Graph signal process- ing for machine learning: A review and new perspec- tives

    Xiaowen Dong, Dorina Thanou, Laura Toni, Michael Bronstein, and Pascal Frossard. Graph signal process- ing for machine learning: A review and new perspec- tives. IEEE Signal processing magazine, 37(6):117– 127, 2020

    2020

  2. [10]

    Fast graph rep- resentation learning with pytorch geometric

    Matthias Fey and Jan Eric Lenssen. Fast graph rep- resentation learning with pytorch geometric. arXiv preprint arXiv:1903.02428, 2019

    1903 arXiv

  3. [11]

    Induc- tive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Induc- tive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

  4. [12]

    Geometry for the selfish herd

    William D Hamilton. Geometry for the selfish herd. Journal of theoretical Biology, 31(2):295–311, 1971

    1971

  5. [13]

    Graph diffusion distance: A difference measure for weighted graphs based on the graph laplacian exponen- tial kernel

    David K Hammond, Yaniv Gur, and Chris R Johnson. Graph diffusion distance: A difference measure for weighted graphs based on the graph laplacian exponen- tial kernel. In 2013 IEEE global conference on signal and information processing , pages 419–422. IEEE, 2013

    2013

  6. [14]

    Angle-encoded swarm optimization for uav formation path planning

    VT Hoang, Manh Duong Phung, Tran Hiep Dinh, and Quang Phuc Ha. Angle-encoded swarm optimization for uav formation path planning. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5239–5244. IEEE, 2018

    2018

  7. [15]

    Geometry for mutualistic and selfish herds: the limited domain of danger

    R James, PG Bennett, and J Krause. Geometry for mutualistic and selfish herds: the limited domain of danger. Journal of Theoretical Biology, 228(1):107– 113, 2004

    2004

  8. [16]

    Op- timal network topology for responsive collective be- havior

    David Mateo, Nikolaj Horsevad, Vahid Hassani, Mo- hammadreza Chamanbaz, and Roland Bouffanais. Op- timal network topology for responsive collective be- havior. Science advances, 5(4):eaau0999, 2019

    2019

  9. [17]

    Effect of correlations in swarms on collective response

    David Mateo, Yoke Kong Kuan, and Roland Bouf- fanais. Effect of correlations in swarms on collective response. Scientific reports, 7(1):10388, 2017

    2017

  10. [18]

    Gromov–wasserstein distances and the metric approach to object matching

    Facundo M´emoli. Gromov–wasserstein distances and the metric approach to object matching. Foundations of computational mathematics, 11:417–487, 2011

    2011

  11. [19]

    A graph-based approach for shep- herding swarms with limited sensing range

    Reem E Mohamed, Saber Elsayed, Robert Hunjet, and Hussein Abbass. A graph-based approach for shep- herding swarms with limited sensing range. In 2021 IEEE congress on evolutionary computation (CEC) , pages 2315–2322. IEEE, 2021

    2021

  12. [20]

    Spatial positioning in the selfish herd

    Lesley J Morrell, Graeme D Ruxton, and Richard James. Spatial positioning in the selfish herd. Be- havioral Ecology, 22(1):16–22, 2011

    2011

  13. [21]

    Evolution of swarming behavior is shaped by how predators attack

    Randal S Olson, David B Knoester, and Christoph Adami. Evolution of swarming behavior is shaped by how predators attack. Artificial life, 22(3):299–318, 2016

    2016

  14. [22]

    Graph signal processing: Overview, challenges, and appli- cations

    Antonio Ortega, Pascal Frossard, Jelena Kova ˇcevi´c, Jos´e MF Moura, and Pierre Vandergheynst. Graph signal processing: Overview, challenges, and appli- cations. Proceedings of the IEEE , 106(5):808–828, 2018

    2018

  15. [23]

    Variational autoencoder

    Lucas Pinheiro Cinelli, Matheus Ara ´ujo Marins, Ed- uardo Ant´unio Barros da Silva, and S´ergio Lima Netto. Variational autoencoder. In Variational Methods for Machine Learning with Applications to Deep Net- works, pages 111–149. Springer, 2021

    2021

  16. [24]

    Bird flocks

    Steven J Portugal. Bird flocks. Current Biology, 30(5):R206–R210, 2020

    2020

  17. [25]

    Swarm attack: A self-organized model to recover from malicious communication ma- nipulation in a swarm of simple simulated agents

    Giuseppe Primiero, Elio Tuci, Jacopo Tagliabue, and Eliseo Ferrante. Swarm attack: A self-organized model to recover from malicious communication ma- nipulation in a swarm of simple simulated agents. InIn- ternational Conference on Swarm Intelligence, pages 213–224. Springer, 2018

    2018

  18. [26]

    De- tecting anomalous swarming agents with graph signal processing

    Kevin Schultz, Anshu Saksena, Elizabeth P Reilly, Rahul Hingorani, and Marisel Villafa˜ne-Delgado. De- tecting anomalous swarming agents with graph signal processing. In 2021 IEEE International Conference on Autonomous Systems (ICAS), pages 1–5. IEEE, 2021. 10 Enhancing Swarms...

    2021

  19. [27]

    Ana- lyzing collective motion using graph fourier analysis

    Kevin Schultz, Marisel Villafane-Delgado, Elizabeth P Reilly, Grace M Hwang, and Anshu Saksena. Ana- lyzing collective motion using graph fourier analysis. arXiv preprint arXiv:2103.08583, 2021

    2021 arXiv

  20. [28]

    Vertex-frequency analysis on graphs

    David I Shuman, Benjamin Ricaud, and Pierre Van- dergheynst. Vertex-frequency analysis on graphs. Applied and Computational Harmonic Analysis , 40(2):260–291, 2016

    2016

  21. [29]

    Implicit neu- ral representations with periodic activation functions

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neu- ral representations with periodic activation functions. Advances in neural information processing systems , 33:7462–7473, 2020

    2020

  22. [30]

    The geometry of decision-making in in- dividuals and collectives

    Vivek H Sridhar, Liang Li, Dan Gorbonos, M´at´e Nagy, Bianca R Schell, Timothy Sorochkin, Nir S Gov, and Iain D Couzin. The geometry of decision-making in in- dividuals and collectives. Proceedings of the National Academy of Sciences, 118(50):e2102157118, 2021

    2021

  23. [31]

    Shared decision- making drives collective movement in wild baboons

    Ariana Strandburg-Peshkin, Damien R Farine, Iain D Couzin, and Margaret C Crofoot. Shared decision- making drives collective movement in wild baboons. Science, 348(6241):1358–1361, 2015

    2015

  24. [32]

    Learning decentralized controllers for robot swarms with graph neural networks

    Ekaterina Tolstaya, Fernando Gama, James Paulos, George Pappas, Vijay Kumar, and Alejandro Ribeiro. Learning decentralized controllers for robot swarms with graph neural networks. In Conference on robot learning, pages 671–682. PMLR, 2020

    2020

  25. [33]

    Environ- mental perturbations induce correlations in midge swarms

    Kasper van der Vaart, Michael Sinhuber, Andrew M Reynolds, and Nicholas T Ouellette. Environ- mental perturbations induce correlations in midge swarms. Journal of The Royal Society Interface , 17(164):20200018, 2020

    2020

  26. [34]

    Collective motion

    Tam´as Vicsek and Anna Zafeiris. Collective motion. Physics reports, 517(3-4):71–140, 2012

    2012

  27. [35]

    A com- prehensive survey on graph neural networks

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A com- prehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4–24, 2020

    2020

  28. [36]

    Graph neural networks: A review of methods and applications

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. Graph neural networks: A review of methods and applications. AI open, 1:57–81, 2020. 11 Enhancing Swarms’ Durability to Threats via GNN-based Generative Mod...

    2020

  29. [37]

    The mean position values µx and µy are sampled from normal distributions centered at µ with standard deviation √ σ2. 2. The range values σx and σy are sampled from normal distributions centered at µr with standard deviation p σ2r. 3. The positions of each agent are sampled fro...

  30. [38]

    The output is a 2D array with the generated positions of all agents

    The positions are clipped to ensure they stay within the bounds of the 2D space (0 ≤ xi, yi ≤ 1000). The output is a 2D array with the generated positions of all agents. A.2.4. T RAINING DETAILS Input normalization: In the first layer, the input x is normalized and scaled as f...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.