Pith. sign in

REVIEW 4 major objections 8 minor 1 cited by

Advancing network resilience theories with symbolized reinforcement learning

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A machine-learning pipeline turns an AI network-attack strategy into the resilience formula d·s, the first to couple topology and dynamics.

desk verdict A serious empirical paper that finds a simple, apparently general d·s node-importance formula via RL and symbolic regression, but overstates its status as a derived theory. read the letter →

arxiv 2507.08827 v1 pith:ZL3OOA7N submitted 2025-07-04 physics.soc-ph cs.AI

Yu Zheng , Jingtao Ding , Depeng Jin , Jianxi Gao , Yong Li This is my paper
classification physics.soc-phcs.AI PACS 89.75.Fb89.75.Hc
keywords networkresiliencekeystonenodesdismantlingsymbolicregressionreinforcementlearningdegree-statecorrelationearlywarningsignalscomplexnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to establish that an AI system can do more than predict network collapse: it can express the cause of collapse as a human-readable law. The authors train a reinforcement-learning agent to dismantle networks by removing keystone nodes, then symbolize the agent's strategy into formulas. The central discovery is $d\cdot s$—degree times steady state—claimed to be the first resilience theory that couples network structure with the nodes' dynamical states, finding keystone nodes with up to 74.1% fewer removals than existing theories on real networks. The same approach refines classical topological and homogeneous-dynamics formulas, and yields a network-level indicator, the degree–state correlation, and an early-warning signal. If the claim holds, node importance in many real complex systems is governed by a simple product of local connectivity and local activity, which would change how resilience is measured and how collapses are anticipated.

What carries the argument

The load-bearing mechanism is a three-stage self-inductive loop: a deep reinforcement-learning agent, built on a graph neural network, solves the network dismantling problem by removing one node per step until the system loses resilience; a differentiable feature-selection mask then distills the agent by finding the sparse input features that carry most of its predictive information; and an evolutionary symbolic-regression search combines those features with only addition and multiplication to reproduce the agent's logits, after which human experts strip constants. The theory that carries the new result is $d\cdot s$, the product of a node's weighted degree $d$ and its steady state $s$: it encodes the node's topological reach and dynamical influence in a single ranking. For pure-topology and homogeneous-dynamics settings, the same machinery yields the refinements $d^2/\bar d$ and $2\bar d+d(\bar d-2\beta)$.

What would settle it

Train the same pipeline on networks with strong negative degree–state correlation, modular community structure, or dynamics lacking a unique steady state, and check by exhaustive search whether $d\cdot s$ still achieves near-zero removal error; a single network class where it overestimates keystone nodes substantially would falsify the universality claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that an AI system can generate, rather than merely approximate, network resilience theory. By training a graph-neural-network reinforcement-learning agent to dismantle networks with the fewest node removals and then symbolically regressing its decision logits onto a sparse set of distilled input features, the authors obtain the formula $d\cdot s$, where $d$ is a node's weighted degree and $s$ is its steady state. This is claimed to be the first resilience theory that accounts for both topology and dynamics. The paper further claims that the correlation between the degree and state distributions, $\mathrm{corr}(d,s)$, is a network-level resilience indicator, that $\alpha = \sum_{i\in V_c} d_i\,s_i \,/\, \sum_{i\in V} d_i\,s_i$ works as an early-warning signal, and that the same pipeline refines existing formulas to $d^2/\bar d$ for structural resilience and $2\bar d + d(\bar d-2\beta)$ for homogeneous dynamics.

Load-bearing premise

The load-bearing premise is that a single reinforcement-learning agent trained on 30 synthetic 80-node scale-free networks with Michaelis–Menten dynamics generalizes to every tested topology, dynamics, and size, so that the formula extracted from its decisions is a general law of network resilience rather than a quirk of that training set.

Editorial extensions

If this is right

  • Keystone-node identification in heterogeneous dynamical networks reduces to ranking nodes by $d\cdot s$; on real gene-regulatory and neuronal networks this removes up to 74.1% fewer nodes than existing topological and learned baselines.
  • Network-level resilience is predicted by $\mathrm{corr}(d,s)$; decorrelating degree from state raises the number of nodes that must be removed by over 183% in synthetic scale-free networks and over 300% in random networks, and the gap between ecology and supply-chain networks helps explain why natural systems look more resilient.
  • An early-warning signal built from $d\cdot s$ reaches 50% roughly 38% before functional collapse, giving operators time to intervene.
  • The same pipeline refines established theories: $d^2/\bar d$ beats collective influence by up to 3.53% for structural dismantling, and $2\bar d+d(\bar d-2\beta)$ beats resilience centrality by up to 37.5% under homogeneous dynamics.
  • The symbolized $d\cdot s$ formula transfers to networks outside its training distribution and infers keystone nodes faster than the neural agent it was extracted from.

Reading between the lines

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

  • An unstated corollary is that resilience can be engineered: deliberately decorrelating degree from state—for example by shifting high activity off hub nodes—should harden a network, a prediction that can be tested by rewiring supply-chain or ecological networks and measuring removal cost.
  • The $d\cdot s$ rule assumes a well-defined steady state; for oscillatory, chaotic, or time-varying dynamics, the formula would likely need a time-averaged or distributional definition of $s$, a regime the paper does not test.
  • The same self-inductive recipe—solve a hard search problem with reinforcement learning, distill it, symbolize it—could be applied to percolation, network growth, or influence maximization to yield closed-form rules instead of black-box policies.
  • Because the discovered law came from an agent trained on small scale-free graphs, its universality is best probed on strongly modular, spatially embedded, or negatively correlated networks; a failure there would show the law is a training-distribution artifact rather than a general principle.
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

4 major / 8 minor

Summary. The paper proposes 'symbolized reinforcement learning,' a self-inductive pipeline that (i) trains a DRL agent (PPO with a GNN policy) to dismantle networks by removing nodes until the system loses resilience, (ii) uses GNNExplainer-based feature selection to identify the most important node features, and (iii) applies symbolic regression (PySR) to distill the agent's node-selection logits into closed-form formulas. Applied to three settings, the pipeline yields d·s (degree times steady state) for heterogeneous dynamics (Michaelis-Menten and Wilson-Cowan), d²/d̄ for pure topological dismantling, and 2d̄+d(d̄−2β) refining resilience centrality. The paper further introduces corr(d,s) as a network-level resilience indicator and an early warning signal α. The authors validate the formulas on real-world cellular, neuronal, social, ecology, and supply-chain networks, reporting large improvements over existing topological and machine-learning baselines.

Significance. If the central claims are established, this is a valuable AI-for-science contribution: it demonstrates an automated way to convert a DRL policy into interpretable formulas with strong empirical generalization across topologies, dynamics, and network sizes. The paper ships code on GitHub, and the extracted formulas make concrete, falsifiable predictions (node rankings via d·s; network-level resilience ordering via corr(d,s)). The empirical results in Tables 1-3 and Figure 2 are encouraging. However, the 'theory discovery' claim is currently overstated: the formulas are symbolic-regression fits to one trained policy, not derivations from the resilience problem, and the universality claim rests on unexamined transfer assumptions. The significance will be greatly enhanced if the authors address the robustness and methodological gaps detailed below.

major comments (4)
  1. [M.3–M.5] The discovered formula d·s is derived entirely from a single PPO agent trained on 30 scale-free 80-node networks with Michaelis-Menten dynamics (M.3). Feature selection (M.4) and symbolic regression (M.5) are then performed on the same 30 networks, with the agent's logits (Eq. 27) as the SR target. Thus d·s is an in-sample distillation of one policy, and the paper provides no random-seed replication for the agent or the symbolic-regression pipeline. The generalization results in Tables 1-2 and Figure 2 do not control for agent instability: if the agent overfits this training distribution, the formula inherits that bias. The authors should train the agent with multiple seeds and on multiple training distributions (e.g., ER/BA/RP/SW, different sizes, both dynamics) and report whether d·s is robustly recovered.
  2. [Figure 2e-f] The 'ground-truth values obtained by exhaustive search' are not described in the main text or Methods. It is unclear whether the ground truth refers to the global optimum of the original optimization problem (5-7) or the recursive version (8-12), and what algorithm was used for networks up to N=200. Without this description, the claim that d·s achieves 'near-zero errors' and is optimal is unverifiable. In addition, Figure 2a-f shows curves averaged over 10 networks per size without error bars or confidence intervals; the authors should report variance.
  3. [Figure 3a-b] The network-level indicator corr(d,s) is tested by 'setting dynamical parameters to different values to achieve different degree-state correlation' (Figure 3 caption). This manipulation changes the marginal distribution of s as well as the correlation, so the observed resilience differences could be driven by changes in the state distribution rather than by the correlation itself. To support the claim that corr(d,s) is the causal network-level indicator, the authors should vary corr(d,s) while controlling the marginals (e.g., by shuffling states among nodes) and show that the resilience effect persists.
  4. [Abstract and Discussion] The claims of 'the first resilience theory that accounts for both topology and dynamics' and of discovering theories 'without introducing any assumptions' are not supported by the methodology. The formula is obtained by symbolic regression on a learned policy (M.5), with human expert selection (M.1, M.6) and a manually restricted operator set (+, ×). The authors should either temper these claims to describe an interpretable heuristic distilled from a trained policy, or provide a rigorous derivation and evidence that d·s is not an artifact of the training distribution. A comparison with existing dynamical importance measures (e.g., eigenvector-based or spectral influence measures) would also be needed to substantiate the 'first' claim.
minor comments (8)
  1. [Abstract] The phrase 'financial crise' should be 'financial crises.'
  2. [Figure 4] The axis labels read 'Funtionality'; they should be 'Functionality.'
  3. [Results, paragraph 1] The network model name 'erd˝os-rényi' should be rendered as 'Erdős–Rényi.'
  4. [M.3] The 11-dimensional feature set is listed with blank placeholders and a typo ('neightbor state'); the features should be enumerated explicitly (with equations or names) for reproducibility.
  5. [Results, d·s section] The 'state' s in d·s is not formally defined in the main text; the authors should state that s is the steady-state activity x_T from Eq. (43) for cellular dynamics and the corresponding steady state for neuronal dynamics.
  6. [Abstract and Table 3] The abstract states 'over 37.5% improvement in accuracy,' but Table 3 reports a maximum improvement of 37.5%; the wording should be 'up to 37.5%.'
  7. [Eq. (1) and Figure 4] The early warning signal α equals 1 by construction when all keystone nodes Vc are removed; the claim of being '38% ahead of the actual collapse' should be reported with error bars across multiple networks, not just a single exemplar.
  8. [Methods] Several core methodological details (DRL hyperparameters λ1, λ2, λ3, γ, ε, the exhaustive-search algorithm for Figure 2e-f, and full results) are relegated to Supplementary Notes. The main text should at least summarize the exhaustive-search procedure and the key hyperparameters so that the central claims are self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: d·s is a distilled formula validated out-of-sample, not a self-defined prediction.

full rationale

The paper's derivation chain is a legitimate distillation pipeline, not a circular reduction. The DRL agent is trained on a network-dismantling objective (M.3), then XAI selects important features (M.4), and symbolic regression fits formulas to the agent's logits (M.5). This is an honest 'self-inductive' procedure: d·s is not defined as the resilience objective, nor is the resilience objective defined in terms of d·s. The key question is whether the distilled formula is merely a restatement of its training signal. It is not, because the paper validates d·s out-of-sample on unseen topologies (ER, BA, RP, SW), larger network sizes up to 3125, real-world networks, and a different dynamical model (Wilson-Cowan) that the agent never saw during training (Table 1, Figure 2). It also compares against exhaustive-search ground truth for small networks and against external baselines (DC, RC, CI, GND, EI, CoreHD, GDM9, FINDER), showing near-zero error. The paper further reports that the symbolized d·s formula outperforms its DRL agent precursor on networks not included in training data (Supplementary Note 6), which is direct evidence that the formula captures generalizable structure rather than memorizing the agent's fit. The only same-author citations (Refs. 2 and 8 for β and RC) are established external baselines used for comparison, not load-bearing uniqueness theorems or ansatz-smuggling devices. The corr(d,s) network-level analysis is a new simulation study, not the training set. Consequently, the central claim has independent, externally falsifiable content and does not reduce to its inputs by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 3 invented entities

The central claim rests on modeling choices inherited from prior resilience definitions, on the transferability of one DRL policy, and on the fidelity of the distillation/symbolic regression chain. The formulas themselves contain no fitted constants, but the pipeline has many hand-chosen settings and the synthetic corr(d,s) experiments fit dynamical parameters to target correlations.

free parameters (5)
  • heterogeneity parameter a = not reported
    Controls the power-law distribution of decay rates b_i in Eqs. (40)-(41) and (45); the value sets the dynamical heterogeneity and thus the state distribution used in d·s and corr(d,s).
  • dynamical parameters for corr(d,s) experiments = not reported
    In Figure 3a-b, parameters are set to different values to achieve target degree-state correlations; this fits the input to produce the claimed resilience relationship.
  • DRL training hyperparameters (λ1, λ2, λ3, γ, ε, entropy weight) = not reported in main text
    The discovered formula is distilled from the trained policy, so these hyperparameters indirectly determine the formula; exact values are not given.
  • number of important features P and final operator set = P unspecified; operators {+,×}
    The topP selection in Eq. (34) and the final operator set are human choices that constrain symbolic search and therefore affect which formula is found.
  • integration horizon T = 400
    Steady states in Eqs. (43) and (47) are defined by integrating to T=400; the state values feeding d·s depend on this choice.
assumptions (6)
  • domain assumption Resilience is correctly defined by the aggregate steady-state criteria in Eqs. (44) and (48).
    Adopted from refs. 2 and 5; all ground-truth keystone-node evaluations inherit this modeling choice.
  • domain assumption The recursive greedy dismantling formulation (Eqs. 8-12) is an adequate proxy for the NP-hard minimum keystone set (Eqs. 5-7).
    The paper transforms the intractable set-selection problem into a greedy recursion; optimality claims are relative to this objective.
  • ad hoc to paper A DRL agent trained on 30 synthetic 80-node scale-free networks with Michaelis-Menten dynamics generalizes to all tested topologies, dynamics, and sizes.
    The universality of d·s depends on this transfer; no theoretical guarantee is provided (M.3).
  • ad hoc to paper The distilled GNN logits (Eq. 27) faithfully represent the DRL agent's node-selection strategy.
    Symbolic regression fits formulas to these logits (M.5); if the binary-classifier distillation loses information, the formula inherits the distortion.
  • standard math PySR's evolutionary search with hall-of-fame selection finds the simplest accurate formula within the allowed operator set.
    Symbolic regression is NP-hard (ref. 44), so the search is heuristic; the final formula is not guaranteed to be globally optimal.
  • domain assumption The node state s in d·s remains a valid importance indicator after previous node removals during the recursive dismantling.
    The formula is applied at each step of the recursion; if states are not recomputed after removals, the sequential use is unjustified.
invented entities (3)
  • d·s node importance metric independent evidence
    purpose: Identify keystone nodes whose removal most quickly collapses network resilience with heterogeneous dynamics.
    Evaluated on held-out real-world cellular and neuronal networks (Table 1), giving external predictive evidence.
  • corr(d,s) network-level resilience indicator independent evidence
    purpose: Quantify whole-network resilience from the degree-state correlation.
    Applied to real ecology and supply-chain networks (Fig. 3e-f), though synthetic experiments fit parameters to produce the correlation.
  • early warning signal α independent evidence
    purpose: Provide a 0-to-1 signal of proximity to collapse (Eq. 1).
    Demonstrated on four real networks (Fig. 4), but the denominator uses the d·s-defined keystone set, so it is partly self-referential.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing network resilience theories with symbolized reinforcement learning." pith.science (2026). https://pith.science/paper/ZL3OOA7N

@misc{pith2026250708827,
  author       = {Pith},
  title        = {Pith review of: Advancing network resilience theories with symbolized reinforcement learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZL3OOA7N}},
  note         = {Machine review of arXiv:2507.08827}
}
read the original abstract

Many complex networks display remarkable resilience under external perturbations, internal failures and environmental changes, yet they can swiftly deteriorate into dysfunction upon the removal of a few keystone nodes. Discovering theories that measure network resilience offers the potential to prevent catastrophic collapses--from species extinctions to financial crise--with profound implications for real-world systems. Current resilience theories address the problem from a single perspective of topology, neglecting the crucial role of system dynamics, due to the intrinsic complexity of the coupling between topology and dynamics which exceeds the capabilities of human analytical methods. Here, we report an automatic method for resilience theory discovery, which learns from how AI solves a complicated network dismantling problem and symbolizes its network attack strategies into theoretical formulas. This proposed self-inductive approach discovers the first resilience theory that accounts for both topology and dynamics, highlighting how the correlation between node degree and state shapes overall network resilience, and offering insights for designing early warning signals of systematic collapses. Additionally, our approach discovers formulas that refine existing well-established resilience theories with over 37.5% improvement in accuracy, significantly advancing human understanding of complex networks with AI.

Figures

Figures reproduced from arXiv: 2507.08827 by the authors.

Figure 1
Figure 1. a. The proposed self-inductive symbolized reinforcement learning framework. It first dismantles networks with a DRL agent searching for keystone nodes, then symbolizes the DRL agent with mathematical formulas unraveling its node selection strategy, leading to the discovery of network resilience theories. b. The GNN-based DRL agent. It encodes rich features of both network topology and complex dynamics into node repr… view at source ↗
Figure 2
Figure 2. a-d. The node removal ratio of different methods (lower means more accurate) for a. ER networks b. BA networks c. RP networks and d. SW networks of growing network sizes, with each size containing 10 synthetic networks driven by gene regulatory dynamics. Results demonstrate the robustness of our discovered theory, which constantly provides precise identification of keystone nodes, with the number of removed nodes re… view at source ↗
Figure 3
Figure 3. a-b. Resilience measured by the number of removed nodes under different correlations between node degree distribution and node state distribution for a. neuronal BA networks and b. neuronal ER networks. The dynamical parameters are set to different values to achieve different degree-state correlation. The results demonstrate negative relation between degree-state correlation and network resilience, where networks ar… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Network functionality (blue) and early warning signal α (red) under DC attack strategy (solid) that removes hub nodes first and d ·s attack strategy (dashed) that prioritizes nodes with larger d ·s. The first response time (orange circle and exclamation mark) indicates…

Discussion (0). Sign in 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. AutoSND: From Execution Evidence to Structural Policies for Automated Network Dismantling Heuristic Discovery

    cs.AI 2026-08 conditional novelty 7.0 of 10

    A three-stage LLM search framework converts execution evidence into structural policies and discovers network dismantling programs with strong quality and low runtime.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    V ., Parshani, R., Paul, G., Stanley, H

    Buldyrev, S. V ., Parshani, R., Paul, G., Stanley, H. E. & Havlin, S. Catastrophic cascade of failures in interdependent networks. Nature 464, 1025–1028 (2010)

    work page 2010

  2. [2]

    & Barabási, A.-L

    Gao, J., Barzel, B. & Barabási, A.-L. Universal resilience patterns in complex networks. Nature 530, 307–312 (2016)

    work page 2016

  3. [3]

    & Barabási, A.-L

    Albert, R., Jeong, H. & Barabási, A.-L. Error and attack tolerance of complex networks. Nature 406, 378–382 (2000)

    work page 2000

  4. [4]

    & Liu, Y .-Y

    Fan, C., Zeng, L., Sun, Y . & Liu, Y .-Y . Finding key players in complex networks through deep reinforcement learning.Nat. machine intelligence 2, 317–324 (2020)

    work page 2020

  5. [5]

    Sanhedrai, H. et al. Reviving a failed network through microscopic interventions. Nat. Phys. 18, 338–349 (2022)

    work page 2022

  6. [6]

    Artime, O. et al. Robustness and resilience of complex networks. Nat. Rev. Phys. 1–18 (2024)

    work page 2024

  7. [7]

    & Makse, H

    Morone, F. & Makse, H. A. Influence maximization in complex networks through optimal percolation. Nature 524, 65–68 (2015)

    work page 2015

  8. [8]

    & Gao, J

    Zhang, Y ., Shao, C., He, S. & Gao, J. Resilience centrality in complex networks.Phys. Rev. E 101, 022304 (2020)

    work page 2020

Show all 45 references
  1. [9]

    & Mangioni, G

    Grassia, M., De Domenico, M. & Mangioni, G. Machine learning dismantling and early-warning signals of disintegration in complex systems. Nat. Commun. 12, 5190 (2021)

    2021

  2. [10]

    & Brown, M

    Ulgiati, S. & Brown, M. T. Monitoring patterns of sustainability in natural and man-made ecosystems. Ecol. modelling 108, 23–36 (1998)

    1998

  3. [11]

    Lalou, M., Tahraoui, M. A. & Kheddouci, H. The critical node detection problem in networks: A survey. Comput. Sci. Rev. 28, 92–117 (2018)

    2018

  4. [12]

    Davies, A. et al. Advancing mathematics by guiding human intuition with ai. Nature 600, 70–74 (2021)

    2021

  5. [13]

    Clusella, P., Grassberger, P., Pérez-Reche, F. J. & Politi, A. Immunization and targeted destruction of networks using explosive percolation. Phys. review letters 117, 208301 (2016). 14/22

    2016

  6. [14]

    & Shamir, R

    Karlebach, G. & Shamir, R. Modelling and analysis of gene regulatory networks. Nat. reviews Mol. cell biology 9, 770–780 (2008)

    2008

  7. [15]

    & Barabási, A.-L

    Barzel, B. & Barabási, A.-L. Universality in network dynamics. Nat. physics 9, 673–681 (2013)

    2013

  8. [16]

    & Barabási, A.-L

    Barzel, B., Liu, Y .-Y . & Barabási, A.-L. Constructing minimal models for complex system dynamics.Nat. communications 6, 7186 (2015)

    2015

  9. [17]

    & Hwang, D.-U

    Boccaletti, S., Latora, V ., Moreno, Y ., Chavez, M. & Hwang, D.-U. Complex networks: Structure and dynamics.Phys. reports 424, 175–308 (2006)

    2006

  10. [18]

    Laurence, E., Doyon, N., Dubé, L. J. & Desrosiers, P. Spectral dimension reduction of complex dynamical networks. Phys. Rev. X 9, 011042 (2019)

    2019

  11. [19]

    & Shi, G.-Y

    Wu, R.-J., Kong, Y .-X., Di, Z., Bascompte, J. & Shi, G.-Y . Rigorous criteria for the collapse of nonlinear cooperative networks. Phys. Rev. Lett. 130, 097401 (2023)

    2023

  12. [20]

    Wilson, H. R. & Cowan, J. D. Excitatory and inhibitory interactions in localized populations of model neurons. Biophys. journal 12, 1–24 (1972)

    1972

  13. [21]

    Wilson, H. R. & Cowan, J. D. A mathematical theory of the functional dynamics of cortical and thalamic nervous tissue. Kybernetik 13, 55–80 (1973)

    1973

  14. [22]

    & Antulov-Fantulin, N

    Ren, X.-L., Gleinig, N., Helbing, D. & Antulov-Fantulin, N. Generalized network dismantling. Proc. national academy sciences 116, 6554–6559 (2019)

    2019

  15. [23]

    & Zhou, H.-J

    Zdeborová, L., Zhang, P. & Zhou, H.-J. Fast and simple decycling and dismantling of networks. Sci. reports 6, 37954 (2016)

    2016

  16. [24]

    R., Stevenson, M., Busby, J

    Tukamuhabwa, B. R., Stevenson, M., Busby, J. & Zorzini, M. Supply chain resilience: definition, review and theoretical foundations for further study. Int. journal production research 53, 5592–5623 (2015)

    2015

  17. [25]

    & Bascompte, J

    Dakos, V . & Bascompte, J. Critical slowing down as early warning for the onset of collapse in mutualistic communities. Proc. Natl. Acad. Sci. 111, 17546–17551 (2014)

    2014

  18. [26]

    & Havlin, S

    Cohen, R., Erez, K., Ben-Avraham, D. & Havlin, S. Breakdown of the internet under intentional attack. Phys. review letters 86, 3682 (2001)

    2001

  19. [27]

    M., Moreira, A

    Schneider, C. M., Moreira, A. A., Andrade Jr, J. S., Havlin, S. & Herrmann, H. J. Mitigation of malicious attacks on networks. Proc. Natl. Acad. Sci. 108, 3838–3841 (2011)

    2011

  20. [28]

    & Kasmi, C

    Engsig, M., Tejedor, A., Moreno, Y ., Foufoula-Georgiou, E. & Kasmi, C. Domirank centrality reveals structural fragility of complex networks via node dominance. Nat. Commun. 15, 56 (2024)

    2024

  21. [29]

    & Albert, R

    Barabási, A.-L. & Albert, R. Emergence of scaling in random networks. Science 286, 509–512 (1999)

    1999

  22. [30]

    Fawzi, A. et al. Discovering faster matrix multiplication algorithms with reinforcement learning. Nature 610, 47–53 (2022)

    2022

  23. [31]

    & Barabási, A.-L

    Albert, R. & Barabási, A.-L. Statistical mechanics of complex networks. Rev. modern physics 74, 47 (2002)

    2002

  24. [32]

    Watts, D. J. & Strogatz, S. H. Collective dynamics of ‘small-world’networks. Nature 393, 440–442 (1998)

    1998

  25. [33]

    S., Newman, M

    Callaway, D. S., Newman, M. E., Strogatz, S. H. & Watts, D. J. Network robustness and fragility: Percolation on random graphs. Phys. review letters 85, 5468 (2000)

    2000

  26. [34]

    Romera-Paredes, B. et al. Mathematical discoveries from program search with large language models. Nature 1–3 (2023)

    2023

  27. [35]

    & Luong, T

    Trinh, T., Wu, Y ., Le, Q., He, H. & Luong, T. Solving olympiad geometry without human demonstrations.Nature DOI: 10.1038/s41586-023-06747-5 (2024)

    2024 doi

  28. [36]

    Senior, A. W. et al. Improved protein structure prediction using potentials from deep learning. Nature 577, 706–710 (2020)

    2020

  29. [37]

    Jumper, J. et al. Highly accurate protein structure prediction with alphafold. Nature 596, 583–589 (2021)

    2021

  30. [38]

    Merchant, A. et al. Scaling deep learning for materials discovery. Nature 1–6 (2023)

    2023

  31. [39]

    Meena, C. et al. Emergent stability in complex network dynamics. Nat. Phys. 1–10 (2023)

    2023

  32. [40]

    Kipf, T. N. & Welling, M. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR) (2017)

    2017

  33. [41]

    Raffin, A. et al. Stable-baselines3: Reliable reinforcement learning implementations. J. Mach. Learn. Res. 22, 1–8 (2021). 15/22

    2021

  34. [42]

    & Klimov, O

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A. & Klimov, O. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  35. [43]

    & Leskovec, J

    Ying, Z., Bourgeois, D., You, J., Zitnik, M. & Leskovec, J. Gnnexplainer: Generating explanations for graph neural networks. Adv. neural information processing systems 32 (2019)

    2019

  36. [44]

    & Pissis, S

    Virgolin, M. & Pissis, S. P. Symbolic regression is np-hard. arXiv preprint arXiv:2207.01018 (2022)

    2022 arXiv

  37. [45]

    𝜔#⋯topological & dynamicalimportant features inferencekeystones+× 𝜔!×𝜔

    Cranmer, M. Interpretable machine learning for science with pysr and symbolicregression. jl. arXiv preprint arXiv:2305.01582 (2023). Author contributions Y .Z., J.D., and Y .L. conceived of the project and designed the research methods. Y .Z. developed the self-inductive symbo...

    2023 arXiv

Pith tools

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