Pith. sign in

REVIEW 3 major objections 6 minor 60 references

Topology-Assisted Spatio-Temporal Pattern Disentangling for Scalable MARL in Large-scale Autonomous Traffic Control

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that routing traffic-signal agents by local topological signatures, not spatial adjacency, makes MARL scale to large heterogeneous road networks.

desk verdict A genuinely new topology-routed MoE mechanism for MARL in traffic control, but the central expressiveness proof assumes the key premise and the 'superior performance' claim is not supported on two of three maps. read the letter →

arxiv 2506.12453 v1 pith:NTC2SOAX submitted 2025-06-14 cs.AI

Rongpeng Li , Jianhang Zhu , Jiahao Huang , Zhifeng Zhao , Honggang Zhang This is my paper
classification cs.AI
keywords multi-agentreinforcementlearningtrafficsignalcontroldynamicgraphneuralnetworkstopologicaldataanalysispersistencediagramsmixtureofexpertsMAPPOrepresentation
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

Large-scale traffic signal control is a multi-agent problem in which intersections differ in shape, lane count, and connectivity, so a single shared representation must serve structurally heterogeneous agents. This paper proposes TGN-TMoE, a MARL architecture that computes a persistence-diagram topological signature for every lane vertex, then routes vertices to specialized experts by local topological similarity rather than spatial adjacency. The authors argue this topology-assisted routing (TSD) lets a Temporal Graph Network backbone disentangle heterogeneous observations, and they back it with a theorem claiming TGN-TMoE is at least as expressive as the TOGL graph model and strictly more expressive when local topology differs while global topology matches. In simulations of three real city maps, TGN-TMoE-empowered MAPPO achieves the best or second-best fuel, CO2, travel time, and delay figures among baselines including GLOSA, PressLight, and CoTV. The authors position the approach as a communication-free, topology-aware route to scalable and deployable MARL.

What carries the argument

The carrying mechanism is the vertex-level topological signature $T_v(j) \in \mathbb{R}^{d_1}$, computed as $T_v, T_e = \mathrm{MLP}\circ\Psi\circ\Upsilon\circ\Phi(G)$, where $\Phi$ maps vertices and edges through $U$ filtration functions and $\Psi$ embeds the resulting persistence diagram via rational-hat, triangle, Gaussian, and line transformations. Each routing score $Q_{jp} = \exp(W_p \cdot T_v(j)) / \sum_j \exp(W_p \cdot T_v(j))$ measures the distance between a lane vertex and an expert's learned anchor topology $W_p$, and these scores build the expert's input slot from the multi-head GAT outputs. This replaces spatial adjacency with topological similarity as the grouping principle, so topologically equivalent but spatially distant lanes are processed by the same specialist, which is what the paper claims lifts the expressiveness bottleneck.

What would settle it

Compute, for a fixed intersection, the vertex-level topological signatures $T_v(j)$ and the routing weights $Q_{jp}$ they produce, then compare against ground-truth structural roles such as controlled lanes, peripheral lanes, and lanes in distinct cycles; if vertices with clearly different local roles receive near-identical signatures or are not routed to distinguishable experts, the mechanism loses its justification. A simpler ablation: replace $T_v(j)$ with random per-vertex vectors or with degree and cycle counts and re-run the Shenzhen experiment; if fuel, delay, and expert-diversity metrics do not worsen, the persistence-diagram signature is not what carries the reported gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that the bottleneck for MARL in large-scale traffic control is representational: fixed or shallow per-intersection encodings obscure structural differences between junctions, and vertex-wise MoE routing based on learned similarity of raw features is biased by GNN oversmoothing. TGN-TMoE addresses this by feeding each lane's vertex through a temporal graph network enriched with topological signatures derived from the augmented subgraph's persistence diagrams, and by using those signatures in a SoftMoE-style routing module so that lanes with similar local cycle and connectivity roles are grouped into the same expert slots. The authors prove that this construction preserves injectivity of the routing and fusion maps, and conclude from Theorem 1 that the model matches TOGL's expressive power and exceeds it when two graphs share global topology but differ in local structures. Empirically they report that MAPPO with TGN-TMoE obtains the best or second-best results on nearly every metric across three real-world city maps, and that the TSD router yields lower inter-expert similarity and more balanced utilization than a topology-blind MoE.

Load-bearing premise

The load-bearing premise is that the per-vertex topological signature $T_v(j)$, extracted from the augmented subgraph's persistence diagram, faithfully and discriminatively encodes each lane's local structural role; the paper asserts this in Section IV-A and Appendix B but never proves or directly measures that global persistence diagrams yield per-vertex local discriminability, and if that fails the routing is just a learned softmax over graph summaries.

Editorial extensions

If this is right

  • TGN-TMoE-empowered MAPPO reports the best or second-best fuel consumption, CO2 emissions, travel time, and delay in Shenzhen, Shanghai, and Hangzhou, with all four core metrics best in Shenzhen.
  • Because routing is driven by local topological signatures, the model needs no vehicle-to-infrastructure communication; a virtual mean-field vertex supplies global context, which the paper argues makes deployment more practical than communication-dependent baselines like CoTV.
  • The TSD router produces more specialized experts than a topology-blind MoE, measured by lower inter-expert output similarity and more even weight distribution, supporting the paper's claim of structure-aware disentangling.
  • Theorem 1 entails that TGN-TMoE is at least as expressive as TOGL and strictly more expressive in distinguishing graphs with equal global persistence diagrams but differing local cycles or connectivity, provided the graphs are 1-WL distinguishable.
  • Ablations show hidden-state MSE falls as TDA and TMoE are added, which the paper reads as each module improving graph representation learning.

Reading between the lines

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

  • An extension the authors leave implicit: the same topology-assisted routing could apply to any MARL setting with heterogeneous local graph structure, such as warehouse fleets, power grids, or wireless networks, wherever a shared policy must serve differently shaped local neighborhoods.
  • A testable consequence of the paper's reasoning is that replacing the persistence-diagram signature with cheaper per-vertex descriptors, such as degree and cycle counts or WL colors, should degrade routing quality; measuring that gap would isolate how much TDA itself contributes.
  • If Theorem 1 is to transfer, the per-vertex topological signatures must be locally discriminative; a direct check would compare routing-weight similarity against ground-truth structural roles at a junction, since the proof assumes rather than demonstrates that discriminability.
  • The comparison against CoTV suggests a boundary condition: when dense sensing is cheap, communication-based control may match or beat topology-aware learning, so the method's advantage is largest in networks too large for communication.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper proposes TGN-TMoE, a MARL architecture for large-scale traffic signal control. It models each intersection as a dynamic lane-level graph, augments it with a mean-field virtual vertex, and injects topological signatures computed from persistence diagrams into a TGN backbone. A topology-assisted spatial disentangling (TSD) routing mechanism, inspired by SoftMoE, routes vertex representations to specialized experts based on per-vertex topological signatures, and the same module is inserted into MAPPO policy and value networks. The paper claims Theorem 1: that TGN-TMoE subsumes TOGL's expressiveness and surpasses it when local topological differences matter, and it presents experiments on Shenzhen, Shanghai, and Hangzhou reporting improvements over GLOSA, PressLight, and CoTV. The central theoretical and performance claims are not fully supported by the presented evidence.

Significance. If the main claims held, the paper would make a useful contribution: a communication-free, topology-aware routing mechanism for large-scale MARL, with an architecture that integrates TDA and MoE in a novel way. The paper has concrete strengths: experiments on three real-world maps in SUMO, comparison with nontrivial baselines, ablations isolating the effect of TDA and MoE, and qualitative analyses of expert specialization and representation geometry. However, the theoretical expressiveness result rests on an unproven premise about per-vertex local discriminability of the topological signatures, and the headline 'superior performance' claim is contradicted by Table IV in two of the three cities. These are load-bearing issues for the paper's stated contributions, so the contribution is conditional rather than established.

major comments (3)
  1. [Section IV-A, Eq. (6); Appendix B] Theorem 1's distinguishing power depends on the premise that the vertex-level signature T_v(j) computed by Eq. (6) from the global persistence diagram of the augmented subgraph faithfully and discriminatively encodes vertex j's local structural role. The proof of Theorem 1 in Appendix B simply asserts that 'the generated vertex-level topological representations contain rich local topological information' and then states Eq. (34) (T^{G1}_v_j != T^{G2}_v_j implies rV^{G1}_j != rV^{G2}_j), which is exactly the claim to be established. Eq. (6) is only MLP after Psi after Upsilon after Phi(G); the mechanism by which points of the global persistence diagram are attributed to individual vertices is never specified, and no experiment measures local discriminability. If T_v(j) degenerates to a function of the global diagram and vertex filtration values, the TSD routing is a learned softmax over graph summaries and the claimed expressiveness gain over TOGL collapses. The theorem should either be proved under an explicit and verified local-attribution assumption or substantially weakened.
  2. [Theorem 1 and Appendix B] The comparison with TOGL is not established because TOGL, as characterized in Eq. (30), also uses vertex features and a graph representation function f(V, T_v, G), such as a GAT. The theorem's conditions include graphs that are distinguishable by the 1-WL test, and condition 1 only forces identical global persistence diagrams. A TOGL instantiation with a 1-WL-powerful GNN component would already distinguish such graphs, so the conclusion that TOGL 'cannot distinguish' them does not follow from the stated premises. The theorem needs a clearer specification of the TOGL variant being compared and of what 'global topological features' exclude.
  3. [Section V-B1, Table IV] The paper's abstract and Section I claim 'superior performance' and 'superiority' of TGN-TMoE, and Section V-B1 says it achieves best or second-best on nearly all metrics. Table IV shows the opposite in two of three scenarios: in Shanghai, CoTV is better than TGN-TMoE on all four metrics (Fuel 14.770 vs 14.844, CO2 347.312 vs 349.044, Travel Time 231.485 vs 231.556, Delay 106.027 vs 106.875), and in Hangzhou, CoTV is better on Fuel and CO2. The text's qualification that CoTV 'slightly outperforms' is accurate, but the blanket claims should be revised, and the comparison should report variance or statistical significance before claiming practical equivalence.
minor comments (6)
  1. [Appendix B, Lemma 2] The statement of Lemma 2 reads f3(T_v^1) != f1(T_v^2); it should be f3(T_v^1) != f3(T_v^2).
  2. [Eqs. (15) and (16)] Q_jp is defined as a matrix in R^{d x H} through broadcasting, but the notation in Eq. (16) treats Q_p as though it were a per-vertex scalar; the indexing should be cleaned up.
  3. [Appendix B, Eq. (31)] Eq. (31) in the proof of Lemma 1 uses inconsistent indices W1j and W1k, and the passage from the first display to the log expression is hard to follow; a subscript rewrite would improve readability.
  4. [Table VI and Section V-A2] Table VI omits several hyperparameters that appear in the method, notably alpha and beta in the reward function and the teleport threshold T_wp; these should be listed for reproducibility.
  5. [Table IV and Section V-A1] No error bars or standard deviations are reported for Table IV, and the statement 'averaged across these parallel environments' does not establish variability across independent training runs; at least seed-level statistics should be provided.
  6. [Section V-B3, Fig. 9] Fig. 9 is presented as evidence for vertex-level topological similarity, but it shows a single intersection and no quantitative agreement between the T_v similarity matrix and expert routing; a quantitative evaluation across many intersections would be needed to support the claim.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1's proof assumes the vertex-level local discriminability of T_v that it is supposed to establish; the claimed expressiveness gain over TOGL reduces to that assumption.

  1. self definitional [Section IV-C / Appendix B, Theorem 1 proof, Eq. (34)]
    "Due to the embedding functions of TDA, the generated vertex-level topological representations contain rich local topological information, such as their contributions to cycles or connectivity patterns. ... If the local topological signatures of two graphs are different, the model will generate different embeddings for vertices within the two graphs, that is, rVG1 j ‰ rVG2 j , if TG1 j ‰ TG2 j . (34)"

    Theorem 1's conclusion is that TGN-TMoE, unlike TOGL, distinguishes graphs with identical global persistence diagrams but different local topology. The proof's only bridge from graph structure to model outputs is the bare assertion that T_v 'contain rich local topological information.' Eq. (34) then states the theorem's desired distinguishing property as a conditional on T_v. Since Eqs. (15) and (29) define routing and the whole composite F as functions of T_v, any discrimination the model performs is, by construction, exactly the discrimination of the T_v signatures. The proof never shows that T_v from Eq. (6), which is a global-PD embedding, assigns different signatures to the locally different vertices; it assumes precisely what Theorem 1 needs.

full rationale

The one significant circular step is in the proof of Theorem 1: the model's claimed ability to resolve local topological differences that TOGL misses is made to follow from the assertion that the vertex-level topological signatures T_v already encode those local differences, without any derivation or attribution mechanism. Because the routing (Eq. 15) and the composite representation (Eq. 29) are defined as functions of T_v, Eq. (34) is not a consequence of the lemmas alone; it is the theorem's conclusion restated as an assumption. This makes the theoretical expressiveness claim partly definitional: if T_v is locally discriminative, the architecture inherits that power by construction, and if T_v is not, the theorem is unsupported. I did not find other circularity patterns: the self-citation [54] is used only to justify a design detail (T=0 suffices for short-term dynamics) and is not load-bearing; the TOGL expressiveness comparison cites external work; and the traffic experiments are external and independent of the theorem. Still, the paper's central advertised theoretical contribution, subsuming TOGL's expressiveness, rests on this circular step, so the score is 6 rather than 0-2.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The framework rests on the assumption that persistence-diagram-derived vertex signatures capture local topology, on rank and injectivity guarantees imported from a citation, and on a characterization of TOGL that is asserted rather than demonstrated. The main fitted quantities are the number of experts and the reward coefficients; no code release means these cannot be audited.

free parameters (3)
  • No. of TMoE experts P = 16
    Chosen on the Shenzhen map by best performance (Table III), then reused for all maps; this is model selection on one of the three evaluation scenarios.
  • Reward coefficients alpha, beta = not reported
    Reward R = -alpha R_sta - beta R_wait in Section V-A2; values are tunable but omitted from Table VI.
  • Waiting teleport threshold T_wp = not reported
    Introduced in Eq. (21) to normalize R_wait; the threshold value is not given, which affects reward scale and cross-method comparability.
assumptions (3)
  • ad hoc to paper Vertex-level topological signatures T_v(j) from the persistence diagram of G_i^+ faithfully represent each vertex's local topological role.
    This is the key unproved premise for TSD routing (Section IV-A, Eq. 6, and Appendix B, Eq. 34); without it, routing is not topology-guided and Theorem 1 fails.
  • ad hoc to paper Random initialization guarantees rank(W1)=d1 and rank[W1,1]=d1+1, hence injectivity of f1 and f3.
    Lemma 1 and Lemma 2 invoke this to establish injectivity; the paper cites reference [60] rather than proving the rank conditions for the specific softmax routing used here.
  • domain assumption TOGL cannot distinguish graphs with identical global persistence diagrams but different local topology.
    Theorem 1 assumes TOGL relies on global topological descriptors; the paper does not show that TOGL's node-level topological features fail on such pairs, despite citing [22].
invented entities (1)
  • Virtual mean-field vertex v_MF
    purpose: Aggregates global context into each intersection subgraph via Eq. (1) and Eq. (14); borrowed from GAT-MF [49].
    A modeling construct, not an observable; its benefit is only demonstrated through ablation and performance, and no external falsifiable handle is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Topology-Assisted Spatio-Temporal Pattern Disentangling for Scalable MARL in Large-scale Autonomous Traffic Control." pith.science (2026). https://pith.science/paper/NTC2SOAX

@misc{pith2026250612453,
  author       = {Pith},
  title        = {Pith review of: Topology-Assisted Spatio-Temporal Pattern Disentangling for Scalable MARL in Large-scale Autonomous Traffic Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NTC2SOAX}},
  note         = {Machine review of arXiv:2506.12453}
}
read the original abstract

Intelligent Transportation Systems (ITSs) have emerged as a promising solution towards ameliorating urban traffic congestion, with Traffic Signal Control (TSC) identified as a critical component. Although Multi-Agent Reinforcement Learning (MARL) algorithms have shown potential in optimizing TSC through real-time decision-making, their scalability and effectiveness often suffer from large-scale and complex environments. Typically, these limitations primarily stem from a fundamental mismatch between the exponential growth of the state space driven by the environmental heterogeneities and the limited modeling capacity of current solutions. To address these issues, this paper introduces a novel MARL framework that integrates Dynamic Graph Neural Networks (DGNNs) and Topological Data Analysis (TDA), aiming to enhance the expressiveness of environmental representations and improve agent coordination. Furthermore, inspired by the Mixture of Experts (MoE) architecture in Large Language Models (LLMs), a topology-assisted spatial pattern disentangling (TSD)-enhanced MoE is proposed, which leverages topological signatures to decouple graph features for specialized processing, thus improving the model's ability to characterize dynamic and heterogeneous local observations. The TSD module is also integrated into the policy and value networks of the Multi-agent Proximal Policy Optimization (MAPPO) algorithm, further improving decision-making efficiency and robustness. Extensive experiments conducted on real-world traffic scenarios, together with comprehensive theoretical analysis, validate the superior performance of the proposed framework, highlighting the model's scalability and effectiveness in addressing the complexities of large-scale TSC tasks.

Figures

Figures reproduced from arXiv: 2506.12453 by the authors.

Figure 1
Figure 1. The illustration of intersection heterogeneities in the real-word map [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of our topology-enhanced spatial disentangling model for state processing and decision-making. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. The illustration of the TSD module. B. Implementations of TGN-TMoE Recalling the drawbacks of regarding vertex as the input of SoftMoE, we adopt the TMoE architecture by treating the outputs of MH-GAT in Eq. (8a) as the “token” while operating “slot” calculation on the specific vertices. In this design, the outputs of Routep¨q in Eq. (12), denoted as Xp mh p , comprise both the topological embeddings Tv and the orig… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The map of Shenzhen, Shanghai, and Hangzhou. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Episode reward evolution in Shenzhen, Shanghai, and Hangzhou. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: The involved vehicle number in Shenzhen, Shanghai, and Hangzhou. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: MSE loss during training. the TDA-based structural features, TGN-MoE that integrates a standard mixture-of-experts module. Besides the numerical results in Table IV, [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The comparisons of expert output similarities between TGN-MoE and TGN-TMoE architectures in Shenzhen. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Illustration of the local topological structure and routing behavior of the model at a randomly selected intersection in Shenzhen. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: t-SNE visualizations of agent representation distributions over time in Shenzhen using TGN, TGN-MoE, and TGN-TMoE architectures. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 59 canonical work pages

  1. [1]

    A mobility-aware vehicular caching scheme in content centric networks: Model and optimization,

    Y . Zhang, C. Li, T. H. Luan , et al., “A mobility-aware vehicular caching scheme in content centric networks: Model and optimization,” IEEE Trans. Veh. Technol., vol. 68, no. 4, pp. 3100–3112, 2019

    work page 2019

  2. [2]

    Intelligent transportation sys- tems,

    G. Dimitrakopoulos and P. Demestichas, “Intelligent transportation sys- tems,” IEEE Veh. Technol. Mag., vol. 5, no. 1, pp. 77–84, 2010

    work page 2010

  3. [3]

    Irregular traffic time series forecasting based on asynchronous spatio-temporal graph convolutional networks,

    W. Zhang, L. Zhang, J. Han , et al. , “Irregular traffic time series forecasting based on asynchronous spatio-temporal graph convolutional networks,” in Proc. SIGKDD, Barcelona, Spain, Aug. 2024

    work page 2024

  4. [4]

    A deep reinforcement learning network for traffic light cycle control,

    X. Liang, X. Du, G. Wang , et al. , “A deep reinforcement learning network for traffic light cycle control,” IEEE Trans. Veh. Technol. , vol. 68, no. 2, pp. 1243–1253, 2019

    work page 2019

  5. [5]

    Settings for fixed-cycle traffic signals,

    A. J. Miller, “Settings for fixed-cycle traffic signals,” J. Oper. Res. Soc., vol. 14, no. 4, pp. 373–386, 1963

    work page 1963

  6. [6]

    Arterial-based control of traffic flow in urban grid networks,

    N. H. Gartner and C. Stamatiadis, “Arterial-based control of traffic flow in urban grid networks,” Math. Comput. Model. , vol. 35, no. 5-6, pp. 657–671, 2002

    work page 2002

  7. [7]

    Deep reinforcement learning for intelligent transportation systems: A survey,

    A. Haydari and Y . Yılmaz, “Deep reinforcement learning for intelligent transportation systems: A survey,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 1, pp. 11–32, 2022

    work page 2022

  8. [8]

    Intellilight: A reinforcement learning approach for intelligent traffic light control,

    H. Wei, G. Zheng, H. Yao , et al., “Intellilight: A reinforcement learning approach for intelligent traffic light control,” in Proc. SIGKDD, London, UK, Aug. 2018

    work page 2018

Show all 60 references
  1. [9]

    Max pressure control of a network of signalized intersec- tions,

    P. Varaiya, “Max pressure control of a network of signalized intersec- tions,” Transp. Res. Part C Emerg. Technol., vol. 36, pp. 177–195, 2013

    2013

  2. [10]

    Presslight: Learning max pres- sure control to coordinate traffic signals in arterial network,

    H. Wei, C. Chen, G. Zheng , et al. , “Presslight: Learning max pres- sure control to coordinate traffic signals in arterial network,” in Proc. SIGKDD, Anchorage, AK, USA, Aug. 2019

    2019

  3. [11]

    Multi-agent reinforcement learning: A selective overview of theories and algorithms,

    K. Zhang, Z. Yang, and T. Bas ¸ar, “Multi-agent reinforcement learning: A selective overview of theories and algorithms,” Handb. Reinforcement Learn. Control, pp. 321–384, 2021

    2021

  4. [12]

    Recent advances in reinforcement learning for traffic signal control: A survey of models and evaluation,

    H. Wei, G. Zheng, V . Gayah , et al. , “Recent advances in reinforcement learning for traffic signal control: A survey of models and evaluation,” SIGKDD Explor. Newsl., vol. 22, no. 2, p. 12–18, Jan. 2021

    2021

  5. [13]

    A survey on reinforcement learning- based control for signalized intersections with connected automated vehicles,

    K. Zhang, Z. Cui, and W. Ma, “A survey on reinforcement learning- based control for signalized intersections with connected automated vehicles,” Transp. Rev., vol. 44, no. 6, pp. 1187–1208, 2024

    2024

  6. [14]

    CoTV: Cooperative control for traffic light signals and connected autonomous vehicles using deep reinforcement learning,

    J. Guo, L. Cheng, and S. Wang, “CoTV: Cooperative control for traffic light signals and connected autonomous vehicles using deep reinforcement learning,” IEEE Trans. Intell. Transp. Syst. , vol. 24, no. 10, pp. 10 501–10 512, 2023

    2023

  7. [15]

    IG-RL: Inductive graph reinforcement learning for massive-scale traffic signal control,

    F.-X. Devailly, D. Larocque, and L. Charlin, “IG-RL: Inductive graph reinforcement learning for massive-scale traffic signal control,” IEEE Trans. Intell. Transp. Syst. , vol. 23, no. 7, pp. 7496–7507, 2022

    2022

  8. [16]

    Mixtures of experts unlock parameter scaling for deep RL,

    J. S. Obando Ceron, G. Sokar, T. Willi , et al. , “Mixtures of experts unlock parameter scaling for deep RL,” in Proc. ICML, Vienna, Austria, Jul. 2024

    2024

  9. [17]

    Temporal graph networks for deep learning on dynamic graphs,

    E. Rossi, B. Chamberlain, F. Frasca , et al. , “Temporal graph networks for deep learning on dynamic graphs,” in Proc. ICML 2020 Workshop on Graph Representation Learning , Virtual Edition, Jul. 2020

    2020

  10. [18]

    Green, K

    D. Green, K. Lewis, J. Ward , et al. , Guide to traffic management part 6: intersections, interchanges and crossings management , 2020, no. AGTM06-20

    2020

  11. [19]

    Topological graph neural networks,

    M. Horn, E. De Brouwer, M. Moor , et al. , “Topological graph neural networks,” in Proc. ICLR, Virtual Edition, Apr. 2021

    2021

  12. [20]

    A user’s guide to topological data analysis,

    E. Munch, “A user’s guide to topological data analysis,” J. Learn. Anal., vol. 4, no. 2, pp. 47–61, 2017

    2017

  13. [21]

    Hatcher, Algebraic Topology

    A. Hatcher, Algebraic Topology. Cambridge: Cambridge University Press, 2002

    2002

  14. [22]

    The expressive power of graph neural networks: A survey,

    B. Zhang, C. Fan, S. Liu , et al. , “The expressive power of graph neural networks: A survey,” IEEE Trans. Knowl. Data Eng. , vol. 37, no. 3, pp. 1455–1474, 2025

    2025

  15. [23]

    A survey on mixture of experts in large language models,

    W. Cai, J. Jiang, F. Wang , et al. , “A survey on mixture of experts in large language models,” IEEE Trans. Knowl. Data Eng. , vol. 37, no. 7, pp. 3896–3915, 2025

    2025

  16. [24]

    From sparse to soft mixtures of experts,

    J. Puigcerver, C. R. Ruiz, B. Mustafa , et al. , “From sparse to soft mixtures of experts,” in Proc. ICLR, Vienna, Austria, May 2024

    2024

  17. [25]

    Gradient gating for deep multi-rate learning on graphs,

    T. K. Rusch, B. P. Chamberlain, M. W. Mahoney, et al., “Gradient gating for deep multi-rate learning on graphs,” in Proc. ICLR, Kigali, Rwanda, May 2023

    2023

  18. [26]

    A new approach to green light optimal speed advisory (GLOSA) systems for high-density traffic flowe,

    H. Suzuki and Y . Marumo, “A new approach to green light optimal speed advisory (GLOSA) systems for high-density traffic flowe,” in Proc. ITSC. Maui, HI, USA: IEEE, Nov. 2018

    2018

  19. [27]

    The sydney coordinated adaptive traffic (SCAT) system philosophy and benefits,

    A. G. Sims and K. W. Dobinson, “The sydney coordinated adaptive traffic (SCAT) system philosophy and benefits,” IEEE Trans. Veh. Technol., vol. 29, no. 2, pp. 130–137, 1980

    1980

  20. [28]

    Learning an interpretable traffic signal control policy,

    J. Ault, J. P. Hanna, and G. Sharon, “Learning an interpretable traffic signal control policy,” in Proc. AAMAS, Virtual Edition, May 2020

    2020

  21. [29]

    Playing Atari with deep reinforcement learning,

    V . Mnih, “Playing Atari with deep reinforcement learning,” in Proc. NeurIPS, Lake Tahoe, NV , USA, Dec. 2013

    2013

  22. [30]

    Maximum pressure controller for stabilizing queues in signalized arterial networks,

    A. Kouvelas, J. Lioris, S. A. Fayazi , et al. , “Maximum pressure controller for stabilizing queues in signalized arterial networks,” Transp. Res. Rec., vol. 2421, no. 1, pp. 133–141, 2014

    2014

  23. [31]

    Toward a thousand lights: Decentralized deep reinforcement learning for large-scale traffic signal control,

    C. Chen, H. Wei, N. Xu , et al., “Toward a thousand lights: Decentralized deep reinforcement learning for large-scale traffic signal control,” in Proc. AAAI, New York, NY , USA, Feb. 2020

    2020

  24. [32]

    Hierarchical reinforcement learning for dynamic autonomous vehicle navigation at intelligent intersections,

    Q. Sun, L. Zhang, H. Yu , et al. , “Hierarchical reinforcement learning for dynamic autonomous vehicle navigation at intelligent intersections,” in Proc. SIGKDD, Long Beach, CA, USA, Aug. 2023

    2023

  25. [33]

    Meta-learning based spatial-temporal graph attention network for traffic signal control,

    M. Wang, L. Wu, M. Li , et al. , “Meta-learning based spatial-temporal graph attention network for traffic signal control,” Knowl-based Syst. , vol. 250, p. 109166, 2022

    2022

  26. [34]

    Colight: Learning network-level cooperation for traffic signal control,

    H. Wei, N. Xu, H. Zhang , et al. , “Colight: Learning network-level cooperation for traffic signal control,” in Proc. CIKM, Beijing, China, Nov. 2019

    2019

  27. [35]

    Graph attention net- works,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova , et al. , “Graph attention net- works,” in Proc. ICLR, Vancouver, BC, Canada, Apr. 2018

    2018

  28. [36]

    A large-scale traffic signal control algorithm based on multi-layer graph deep reinforcement learning,

    T. Wang, Z. Zhu, J. Zhang , et al. , “A large-scale traffic signal control algorithm based on multi-layer graph deep reinforcement learning,” Transp. Res. Part C Emerg. Technol. , vol. 162, p. 104582, 2024. 18

    2024

  29. [37]

    Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey,

    J. Skarding, B. Gabrys, and K. Musial, “Foundations and modeling of dynamic networks using dynamic graph neural networks: A survey,” IEEE Access, vol. 9, pp. 79 143–79 168, May 2021

    2021

  30. [38]

    Scaling multi- agent reinforcement learning with selective parameter sharing,

    F. Christianos, G. Papoudakis, M. A. Rahman , et al. , “Scaling multi- agent reinforcement learning with selective parameter sharing,” in Proc. ICML, Virtual Edition, Jul. 2021

    2021

  31. [39]

    Adaptive mixtures of local experts,

    R. A. Jacobs, M. I. Jordan, S. J. Nowlan , et al. , “Adaptive mixtures of local experts,” Neural Comput., vol. 3, no. 1, pp. 79–87, 1991

    1991

  32. [40]

    Hash layers for large sparse models,

    S. Roller, S. Sukhbaatar, J. Weston , et al. , “Hash layers for large sparse models,” Dec. 2021

    2021

  33. [41]

    Base layers: Simplifying training of large, sparse models,

    M. Lewis, S. Bhosale, T. Dettmers , et al. , “Base layers: Simplifying training of large, sparse models,” in Proc. ICML, Virtual Edition, Jul. 2021

    2021

  34. [42]

    Conditional computation in neural networks for faster models,

    E. Bengio, P.-L. Bacon, J. Pineau , et al. , “Conditional computation in neural networks for faster models,” in Proc. ICLR , San Juan, Puerto Rico, May 2015

    2015

  35. [43]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz , et al., “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” in Proc. ICLR, Toulon, France, Apr. 2017

    2017

  36. [44]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

    W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,” J. Mach. Learn. Res., vol. 23, no. 120, pp. 1–39, 2022

    2022

  37. [45]

    Mixture-of-experts with expert choice routing,

    Y . Zhou, T. Lei, H. Liu , et al. , “Mixture-of-experts with expert choice routing,” in Proc. NeurIPS, New Orleans, LA, USA, Nov. 2022

    2022

  38. [46]

    LLaVa-MoLE: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms,

    S. Chen, Z. Jie, and L. Ma, “LLaVa-MoLE: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms,” arXiv preprint arXiv:2401.16160 , 2024

    2024 arXiv

  39. [47]

    Federated mixture of experts,

    M. Reisser, C. Louizos, E. Gavves , et al., “Federated mixture of experts,” in Proc. ICLR, Vienna, Austria, May 2021

    2021

  40. [48]

    Mixture of experts in a mixture of RL settings,

    T. Willi, J. S. O. Ceron, J. N. Foerster , et al. , “Mixture of experts in a mixture of RL settings,” in Proc. RLC, Amherst, MA, USA, Aug. 2024

    2024

  41. [49]

    GAT-MF: Graph attention mean field for very large scale multi-agent reinforcement learning,

    Q. Hao, W. Huang, T. Feng , et al. , “GAT-MF: Graph attention mean field for very large scale multi-agent reinforcement learning,” in Proc. SIGKDD, Long Beach, CA, USA, Aug. 2023

    2023

  42. [50]

    High-dimensional continuous control using generalized advantage estimation,

    J. Schulman, P. Moritz, S. Levine , et al., “High-dimensional continuous control using generalized advantage estimation,” in Proc. ICLR , San Juan, Puerto Rico, May 2016

    2016

  43. [51]

    Topological persis- tence and simplification,

    H. Edelsbrunner, D. Letscher, and A. Zomorodian, “Topological persis- tence and simplification,” Discrete Comput. Geom., vol. 28, pp. 511–533, 2002

    2002

  44. [52]

    Perslay: A neural network layer for persistence diagrams and new graph topological signatures,

    M. Carri `ere, F. Chazal, Y . Ike, et al. , “Perslay: A neural network layer for persistence diagrams and new graph topological signatures,” in Proc. AISTATS, Virtual Edition, Aug. 2020

    2020

  45. [53]

    Learning representations of persistence barcodes,

    C. D. Hofer, R. Kwitt, and M. Niethammer, “Learning representations of persistence barcodes,” J. Mach. Learn. Res. , vol. 20, no. 126, pp. 1–45, 2019

    2019

  46. [54]

    Semantics-enhanced temporal graph networks for content popularity prediction,

    J. Zhu, R. Li, X. Chen , et al. , “Semantics-enhanced temporal graph networks for content popularity prediction,” IEEE Trans. Mob. Comput., vol. 23, no. 8, pp. 8478–8492, 2024

    2024

  47. [55]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho , et al. , “Empirical evaluation of gated recurrent neural networks on sequence modeling,” in Proc. NeurIPS , Montreal, QC, Canada, Dec. 2014

    2014

  48. [56]

    How attentive are graph attention networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” in Proc. ICLR, Virtual Edition, Apr. 2022

    2022

  49. [57]

    The reduction of a graph to canonical form and the algebra which appears therein,

    B. Weisfeiler and A. Leman, “The reduction of a graph to canonical form and the algebra which appears therein,” Nauchno-Technicheskaya Informatsia, Ser. 2, vol. 2, no. 9, pp. 12–16, 1968

    1968

  50. [58]

    Microscopic traffic simulation using SUMO,

    P. A. Lopez, M. Behrisch, L. Bieker-Walz , et al. , “Microscopic traffic simulation using SUMO,” in Proc. ITSC, Maui, HI, USA, Nov. 2018

    2018

  51. [59]

    Flow: A modular learning framework for mixed autonomy traffic,

    C. Wu, A. R. Kreidieh, K. Parvate , et al. , “Flow: A modular learning framework for mixed autonomy traffic,” IEEE Trans. Robot. , vol. 38, no. 2, pp. 1270–1286, 2021

    2021

  52. [60]

    Bridging the divide: Reconsidering softmax and linear attention,

    D. Han, Y . Pu, Z. Xia , et al. , “Bridging the divide: Reconsidering softmax and linear attention,” in Proc. NeurIPS, Vancouver, CA, Canada, Dec. 2024

    2024

Pith tools

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