TrajAware: Graph Cross-Attention and Trajectory-Aware for Generalisable VANETs under Partial Observations

Eiman Kanjo; Xiaolu Fu; Ziyuan Bao

REVIEW 4 major objections 7 minor 28 references

TrajAware: Graph Cross-Attention and Trajectory-Aware for Generalisable VANETs under Partial Observations

T0 review · 4 major / 7 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read TrajAware claims a single trained RL router can route near-shortest paths in unseen city networks under partial observations, without retraining.

desk verdict A genuinely new RL routing system with a serious simulation setup, but the comparative claims outrun the evidence because baseline training protocols are undisclosed. read the letter →

arxiv 2509.06665 v1 pith:3YV7OHEC submitted 2025-09-08 cs.LG cs.AI

Xiaolu Fu , Ziyuan Bao , Eiman Kanjo This is my paper

classification cs.LGcs.AI

keywords VANETroutingreinforcementlearninggraphneuralnetworkscross-attentionaction-spacepruningtrajectorypredictionpartialobservabilityzero-shotgeneralisation

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

TrajAware is an attempt to make reinforcement-learning routing for vehicular networks work in the real world, where a vehicle sees only part of the network and city layouts change. The paper argues that three mechanisms together are enough: pruning the set of next-hop neighbours down to those that still reach all two-hop neighbours, replacing fixed-size MLP layers with graph cross-attention that handles variable graph sizes, and predicting vehicle positions from historical routes when observations are outdated. Evaluated on six real-world city maps with a leave-one-city-out split, TrajAware reports average path lengths close to the shortest possible and packet delivery ratios above 0.93 under partial observations, beating four baselines without retraining. If the results hold, a single trained model could be deployed to a new city on constrained edge hardware with no per-city tuning.

What carries the argument

The load-bearing object is the pruned, attention-routed graph representation. Action-space pruning greedily keeps a subset of direct neighbours that together can still reach every two-hop neighbour, capping node degree at eight and cutting the action space from 21 to 8; the paper assumes two-hop coverage is enough to preserve routing quality. Graph cross-attention then maps each retained neighbour's features against the full node set, so the policy sees permutation-invariant, variable-size inputs and can be applied inductively to new graphs. Trajectory prediction supplies real-time positions from a 6-dimensional GRU input (position plus two upcoming segment nodes) with road projection, bridg

What would settle it

Build a small graph where the current node has two neighbours, A and B; both reach a common two-hop node, so greedy pruning may discard A, but only A connects to the destination three hops away while B's route to that destination is much longer. Run TrajAware's pruning on that graph and measure the chosen next hop and the resulting shortest-path ratio; if pruning removes A and the delivered path is no longer near-shortest, the two-hop-coverage assumption fails.

Watch

Extended reading notes

Core claim

In the paper's own terms, the central discovery is that a GraphSAGE-based DQN router generalises to completely unseen road networks when its action space is pruned and its final layer is a cross-attention module instead of an MLP. The full TrajAware pipeline adds a GRU trajectory predictor that uses planned routes and segment-node coordinates to estimate real-time positions, so the routing policy can be trained under complete observation and used unchanged when observations are partial. The paper reports average shortest-path ratios of 1.0031 without congestion and 1.1271 with congestion in unseen cities, and a reached ratio of 0.9307 under partial observations, compared with at most 0.2175

Load-bearing premise

The load-bearing premise is that a direct neighbour can be safely discarded whenever the remaining neighbours still reach all two-hop neighbours; the paper asserts this preserves reachability with the same hop count but gives no proof or error bound for destinations that are farther than two hops away.

Editorial extensions

If this is right

A single trained model can be deployed to a new city without retraining, since the policy trained on five cities transfers to a held-out sixth city.
Partial observability can be handled by proactive message passing plus trajectory prediction, leaving the routing policy itself unchanged.
Action-space pruning cuts the action space from 21 to 8 neighbours per node, reducing the compute and memory footprint enough for edge devices.
The ablation attributes the largest accuracy gain to cross-attention, while pruning mainly improves robustness and sample efficiency; both are needed for delivery ratios near 1.
Routing decisions stay near-shortest even under congestion (SPR 1.127) while delivery ratios remain above 0.96 in complete-observation settings.

Reading between the lines

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

The pruning guarantee is only about two-hop reachability, so on networks with long corridors or sparse bridges a pruned neighbour could be the sole gateway to a distant destination; testing on such topologies would reveal how often the near-shortest-path result degrades.
Because cross-attention removes the need for fixed-size MLP outputs, the same architecture should transfer to other graph-based RL problems with variable node degrees, such as drone swarms or mesh networks, without architectural changes.
Trajectory error grows with hop distance, so delivery ratios should be sensitive to broadcast frequency and destination distance; varying the broadcast frequency in the same simulator would quantify that trade-off.

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.

Desk Editor's Note

read the letter

Short version: this is a real attempt at a hard problem, and the architecture is more than a bag of tricks. But the paper's headline comparative claims are not backed by the evidence as written, mainly because the baselines in the partial-observation experiment are not described well enough to know whether they were even trained on the same data.

The genuinely new thing is the combination: action-space pruning based on 2-hop coverage, a cross-attention layer between neighbor embeddings and global context, and a GRU trajectory predictor that lets a proactive router estimate positions under stale observations. The pruning heuristic is plausible and the ablation suggests cross-attention is doing the heavy lifting. The leave-one-city-out protocol on six real cities in SUMO is a step ahead of the usual 20-node regular graphs. The authors also state the limitation of proactive routing clearly.

The reader's worry about pruning is mostly answerable. If the retained neighbors cover every 2-hop neighbor, a shortest path to any farther destination can be rerouted through a retained neighbor with the same first hop length. The unaddressed edge case is a leaf neighbor at distance 1, and the random discarding of excess neighbors in rare high-degree cases, but these are minor relative to the main claim. The heavier problem is in Table III. Four baseline methods are compared, but the paper doesn't say whether they were retrained on the same five training cities, with the same observation window, message-passing schedule, or action space. The baselines are also modified into hybrid methods and the replay buffer is shrunk. If those models were run out-of-the-box from a 20-node, degree-3 setting, their very low delivery ratios reflect distribution shift, not inferiority. The full-observation table compares only ablations, so the abstract's 'outperforming state-of-the-art baselines in both full and partial observation' is unsupported. No error bars, no code.

That said, the weaknesses are fixable. A revision that discloses baseline training protocols, adds more held-out cities, reports variance, and releases artifacts would make the evidence match the claims. The core system is coherent and worth engaging with. I'd send it to review rather than desk reject.

Referee Report

4 major / 7 minor

Summary. The paper proposes TrajAware, an RL-based VANET routing framework combining three components: action-space pruning that removes redundant neighbour options while preserving two-hop reachability, graph cross-attention to handle variable-size neighbour sets, and GRU-based trajectory prediction for partial observations. It is evaluated in the SUMO simulator on six real-world city maps with a leave-one-city-out protocol, using full-observation ablations and a partial-observation comparison against four methods from Weil et al. The central claims are near-shortest routing paths, high packet delivery ratios, zero-shot generalisation to unseen cities, and superiority over state-of-the-art baselines under both full and partial observation.

Significance. If the main claims hold, this is a useful step forward for graph-based RL routing: it relaxes the fixed-size/fixed-degree assumptions of prior GRL routing work, introduces an attention mechanism that scales to variable neighbourhoods, and evaluates on realistic city-scale networks with a held-out city. The leave-one-city-out protocol is a genuine generalisation test, and the ablation design cleanly separates the contributions of pruning and attention. However, the current evidence does not fully support the headline comparative claims: the full-observation evaluation is internal only, and the partial-observation baseline comparison is under-specified. The method's internal consistency is good, but external validity needs strengthening before the paper can support its abstract and conclusion.

major comments (4)

[Section IV-A, Table I and Abstract] The abstract claims 'outperforming state-of-the-art baselines in both full and partial observation scenarios.' Table I, however, reports only ablations of TrajAware's own components (pruning and attention) under full observation; no external routing method is compared in this setting. Therefore the full-observation SOTA claim is unsupported by the presented evidence. Please add comparisons to prior GRL routing methods (e.g., [3], [4], [6]) under the same environment and metrics, or revise the claim to describe an ablation-based improvement.
[Section IV-C, Table III] The training protocol for the four baselines (RMP, GConv-LSTM, ADGN, GraphSAGE-DQN) is not stated. It is unclear whether these methods were (a) retrained on the same five-city training data with the same observation protocol, hyperparameters, TTL, and reward function, or (b) used as pretrained checkpoints from Weil et al.'s 20-node, degree-3 setting. The paper explicitly says the replay buffer size was reduced relative to [6], so the setup already differs. If (b) is the case, the baselines' low RR values (0.0887–0.2175) may reflect distribution shift rather than algorithmic inferiority. This is load-bearing for the partial-observation SOTA claim. Please specify the baseline training procedure and, ideally, retrain all baselines under an identical protocol.
[Section III-A, Figure 3] The pruning rule is asserted to preserve reachability and path lengths, but no proof is given. For destinations at distance d>=2, a rigorous argument is available: a path s-v1-v2-...-vd can be replaced by s-w-v2-...-vd when w covers all of v1's 2-hop neighbours, preserving length. The paper should state this invariant explicitly. The d=1 case is not addressed: a leaf neighbour with no 2-hop links can be pruned even though it may be the destination, unless direct relaying is handled separately before pruning. Additionally, the fallback of randomly selecting eight neighbours when the pruned degree exceeds eight can break the coverage guarantee; the text says the impact is negligible but does not quantify the frequency of this fallback. Please provide a formal statement, handle the direct-destination case, and report the incidence of the random fallback.
[Sections IV-B and IV-C] The contribution of trajectory prediction to downstream routing is not isolated. Table II reports trajectory prediction error, but there is no routing ablation comparing TrajAware with the prediction module against, for example, using stale broadcast positions or an oracle with ground-truth positions. Given the paper's claim that trajectory-aware prediction is a key component enabling partial-observation performance, please add an ablation that reports routing metrics with and without the prediction module in the partial-observation setting.

minor comments (7)

[Throughout] The spelling 'V ANET' is used inconsistently; the title and abstract use 'VANET' while the body frequently uses 'V ANET'.
[Figure 1] The label 'Trajectory predection' contains a typo ('prediction').
[Section IV-A] The text contains 'sTable I' — likely intended as 'Table I'.
[Section III-A] 'If a vehicle has more than eight neighbours, we remain eight neighbours' should read 'retain eight neighbours'.
[Section III-B, Eq. (10)] The permutation map f in the equivariance/invariance equations is not formally defined. Please specify that f is a permutation applied elementwise to the input list.
[Section IV-C] The sentence 'Consequently, reactive methods generally achieve higher RR and SPR' appears to contradict the preceding discussion and the results in Table III. Please clarify what is meant.
[Related Work and References] Reference [6] shares an author with the current manuscript, so this connection should be explicitly acknowledged when [6] is used to define the baseline and environment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the central derivation and evaluation are self-contained empirical claims.

full rationale

The paper's claimed derivation chain is empirical rather than formal: TrajAware's components (action-space pruning, graph cross-attention, trajectory-aware prediction) are ML modules whose outputs are measured against held-out city networks under a leave-one-city-out protocol. No evaluation metric (SPR, PSPR, RR) is used as a training label in a way that would force the reported results, and the test city is excluded from training for both the routing policy and the trajectory predictor. The action-space pruning step is a structural heuristic with an explicit coverage condition ('the agent only needs to consider a set of direct neighbours that can reach all 2-hop neighbours'); even if the preservation-of-shortest-paths argument is not fully proved, it is an independent claim about graph structure, not a definition of the performance metric. The cross-attention and trajectory modules are adapted from external prior work (Transformer, Kim et al.) and from the authors' own prior environment code, but the environment borrowing is for simulation and baseline comparison, not for deriving the central generalisation result. No equation is shown to reduce to another by construction, and no fitted parameter is renamed as a prediction. The only noticeable self-citation is to Weil et al. [6], co-authored by Z. Bao, which provides the open-source baseline environment and the four comparison methods; however, [6] is used as a reproducible external baseline and is not invoked to prove TrajAware's effectiveness, so it does not constitute load-bearing circularity. The absence of a stated training protocol for the baselines in Table III is a reporting/correctness concern, not a circularity concern, and does not affect the circularity score.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

Free parameters: the maximum pruned neighbour count (8), broadcast frequency (4 Hz), observation horizon (4 hops), TTL (20), connection range (800 m), density factor per city, full-observation network update period (5 steps), and reduced replay buffer size all shape the reported metrics. The two most load-bearing are the max-neighbour cap and the broadcast frequency, because they determine the action space dimensionality and the trajectory prediction error. Axioms: the paper assumes planned paths are broadcast by vehicles, that pruning to a two-hop covering set is safe, that SUMO on major roads with 40 to 70 vehicles represents real VANETs, and that mobility is predictable enough for proactive message passing. It also relies on standard RL equations. Invented entity: the segment node mechanism for encoding road geometry into the trajectory predictor; it has no separate ablation, so independent evidence is absent.

free parameters (8)

max_neighbors_after_pruning = 8 neighbours
Hand-chosen cap on the pruned action space, described in Section III-A ('we reduced the action space from 21 to 8 in our experimental environment' and 'we remain eight neighbours'). Determines the output dimensionality of the cross-attention and FC layers, so it shapes both efficiency and generalisation claims.
broadcast_frequency_f = 4 Hz
In partial observation, 'communication occurs four times per second'. This feeds into the missing-steps formula (Eq. 12) and therefore controls trajectory prediction error.
observation_horizon = 4 hops
In partial observation, the proactive method obtains positions of vehicles within four hops; 'Reactive methods have access to the positions of vehicles within four hops (inclusive)'. Controls the information available to all methods.
packet_ttl = 20 hops
Packet time-to-live used to compute RR and PSPR in Eq. 11; 'We set the time-to-live (TTL) to 20 hops'.
connection_range = 800 m
Assumed wireless range: 'We assume a signal range of 800 meters, slightly below the theoretical upper limit of DSRC'. Determines the graph topology.
traffic_density_factor = per city
Adjusted per city to keep 40-70 simultaneously active vehicles; 'The density factor was adjusted per city to account for variations in road density'. This environmental parameter affects connectivity and task difficulty in all cities.
network_update_period = 5 steps
Complete observation experiments 'computed the network structure every five steps'. Not a model parameter but affects the graph fed to the model in the full-observation setting.
replay_buffer_size = reduced from Weil et al.
Adjusted down to fit 32 GB memory; 'we reduced the replay buffer size to ensure that the experiments fit within the 32 GB memory limit'. May affect baseline training dynamics.

assumptions (5)

domain assumption Vehicles broadcast their planned paths and local observations to neighbours.
Section III states 'we assume it is feasible to transmit a reasonable amount of additional information, including the vehicle's planned path and local observations' and 'We assume such information can typically be obtained from navigation systems or autonomous driving agents'. The trajectory prediction module depends on this.
ad hoc to paper A direct neighbour can be pruned when all 2-hop neighbours reachable through it are also reachable through retained neighbours, without harming route optimality.
Section III-A asserts 'the agent only needs to consider a set of direct neighbours that can reach all 2-hop neighbours'. This is a heuristic stated without proof; the near-shortest-path results rely on it.
domain assumption SUMO simulation on major roads with 40 to 70 active vehicles is representative of real VANET dynamics.
Section IV describes the simulation setup; this assumption underpins the external validity of all performance claims.
domain assumption Proactive message passing before routing is possible because vehicle mobility is predictable.
Section IV states 'Our method, assuming VANET with predictable mobility, performs all message passing before routing and relies on position prediction during routing'.
standard math Standard RL/Bellman equations describe the routing objective.
Section II-B introduces the RL framework and the value-function definitions; this is a standard formal assumption for RL-based routing.

invented entities (1)

segment node
purpose: Marks the end of straight road segments (typically at junctions) to encode direction changes for the GRU trajectory predictor; long straight roads are split with segment nodes to balance their distribution.
Section III-C introduces segment nodes as an input representation for trajectory prediction. No separate ablation isolates their contribution, so there is no falsifiable handle outside the paper's own evaluation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TrajAware: Graph Cross-Attention and Trajectory-Aware for Generalisable VANETs under Partial Observations." pith.science (2026). https://pith.science/paper/3YV7OHEC

@misc{pith2026250906665,
  author       = {Pith},
  title        = {Pith review of: TrajAware: Graph Cross-Attention and Trajectory-Aware for Generalisable VANETs under Partial Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3YV7OHEC}},
  note         = {Machine review of arXiv:2509.06665}
}

read the original abstract

Vehicular ad hoc networks (VANETs) are a crucial component of intelligent transportation systems; however, routing remains challenging due to dynamic topologies, incomplete observations, and the limited resources of edge devices. Existing reinforcement learning (RL) approaches often assume fixed graph structures and require retraining when network conditions change, making them unsuitable for deployment on constrained hardware. We present TrajAware, an RL-based framework designed for edge AI deployment in VANETs. TrajAware integrates three components: (i) action space pruning, which reduces redundant neighbour options while preserving two-hop reachability, alleviating the curse of dimensionality; (ii) graph cross-attention, which maps pruned neighbours to the global graph context, producing features that generalise across diverse network sizes; and (iii) trajectory-aware prediction, which uses historical routes and junction information to estimate real-time positions under partial observations. We evaluate TrajAware in the open-source SUMO simulator using real-world city maps with a leave-one-city-out setup. Results show that TrajAware achieves near-shortest paths and high delivery ratios while maintaining efficiency suitable for constrained edge devices, outperforming state-of-the-art baselines in both full and partial observation scenarios.

Figures

Figures reproduced from arXiv: 2509.06665 by the authors.

**Figure 2.** Our graph reinforcement-learning method, TrajAware. A graph is represented by a tensor that encodes the topology (adjacency matrix or edge index) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 4.** Node degree distributions before and after pruning. The modified [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Road network of Portland in the simulation environment. The black [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Vehicle position heatmap illustrating the spatial distribution of traffic [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages

[3]

Dealing with changes: Resilient routing via graph neural networks and multi-agent deep reinforcement learning,

S. S. Bhavanasi, L. Pappone, and F. Esposito, “Dealing with changes: Resilient routing via graph neural networks and multi-agent deep reinforcement learning,”IEEE Transactions on Network and Service Management, vol. 20, no. 3, pp. 2283–2294, 2023

work page 2023
[4]

Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case,

P. Almasan, J. Su ´arez-Varela, K. Rusek, P. Barlet-Ros, and A. Cabellos- Aparicio, “Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case,”Computer Communications, vol. 196, pp. 184–194, 2022

2022
[6]

Towards Generalizability of Multi-Agent Reinforcement Learning in Graphs with Recurrent Message Passing

J. Weil, Z. Bao, O. Abboud, and T. Meuser, “Towards generalizability of multi-agent reinforcement learning in graphs with recurrent message passing,”arXiv preprint arXiv:2402.05027, 2024

work page Pith review arXiv 2024
[1]

Interworking of dsrc and cellular network technologies for v2x communications: A survey,

K. Abboud, H. A. Omar, and W. Zhuang, “Interworking of dsrc and cellular network technologies for v2x communications: A survey,”IEEE Transactions on Vehicular Technology, vol. 65, no. 12, pp. 9457–9470, 2016

work page 2016
[2]

Qmr: Q-learning based multi-objective optimization routing protocol for flying ad hoc networks,

J. Liu, Q. Wang, C. He, K. Jaffr `es-Runser, Y . Xu, Z. Li, and Y . Xu, “Qmr: Q-learning based multi-objective optimization routing protocol for flying ad hoc networks,”Computer Communications, vol. 150, pp. 304–316, 2020

work page 2020
[5]

Gddr: Gnn-based data-driven routing,

O. Hope and E. Yoneki, “Gddr: Gnn-based data-driven routing,” in2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS). IEEE, 2021, pp. 517–527

work page 2021
[7]

Graph neural networks for wireless networks: Graph representation, architecture and evaluation,

Y . Lu, Y . Li, R. Zhang, W. Chen, B. Ai, and D. Niyato, “Graph neural networks for wireless networks: Graph representation, architecture and evaluation,”IEEE Wireless Communications, 2024

work page 2024
[8]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

2017

Show all 28 references

[9]

Gtat: empowering graph neural networks with cross attention,

J. Shen, Q. T. Ain, Y . Liu, B. Liang, X. Qiang, and Z. Kou, “Gtat: empowering graph neural networks with cross attention,”Scientific Reports, vol. 15, no. 1, p. 4760, 2025

2025
[10]

Convolutional neural networks on graphs with fast localized spectral filtering,

M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural networks on graphs with fast localized spectral filtering,”Advances in neural information processing systems, vol. 29, 2016

2016
[11]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, 2016

2016 arXiv
[12]

Representation learning on graphs with jumping knowledge networks,

K. Xu, C. Li, Y . Tian, T. Sonobe, K.-i. Kawarabayashi, and S. Jegelka, “Representation learning on graphs with jumping knowledge networks,” inInternational conference on machine learning. PMLR, 2018, pp. 5453–5462

2018
[13]

Inductive representation learning on large graphs,

W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”Advances in neural information processing systems, vol. 30, 2017

2017
[14]

Graph convolutional reinforce- ment learning,

J. Jiang, C. Dun, T. Huang, and Z. Lu, “Graph convolutional reinforce- ment learning,”arXiv preprint arXiv:1810.09202, 2018

2018 arXiv
[15]

Playing atari with deep reinforcement learn- ing,

V . Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wier- stra, and M. Riedmiller, “Playing atari with deep reinforcement learn- ing,”arXiv preprint arXiv:1312.5602, 2013

2013 arXiv
[16]

Reinforcement learning: A survey,

L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Reinforcement learning: A survey,”Journal of artificial intelligence research, vol. 4, pp. 237–285, 1996

1996
[17]

Human-level control through deep reinforcement learning,

V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

2015
[18]

Approximation by superpositions of a sigmoidal function,

G. Cybenko, “Approximation by superpositions of a sigmoidal function,” Mathematics of control, signals and systems, vol. 2, no. 4, pp. 303–314, 1989

1989
[19]

Deep reinforcement learning for autonomous driving: A survey,

B. R. Kiran, I. Sobh, V . Talpaert, P. Mannion, A. A. Al Sallab, S. Yo- gamani, and P. P ´erez, “Deep reinforcement learning for autonomous driving: A survey,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2021

2021
[20]

Intelligent routing based on reinforcement learning for software-defined networking,

D. M. Casas-Velasco, O. M. C. Rendon, and N. L. da Fonseca, “Intelligent routing based on reinforcement learning for software-defined networking,”IEEE Transactions on Network and Service Management, vol. 18, no. 1, pp. 870–881, 2020. 10

2020
[21]

House of graphs 2.0: A database of interesting graphs and more,

K. Coolsaet, S. D’hondt, and J. Goedgebeur, “House of graphs 2.0: A database of interesting graphs and more,”Discrete Applied Mathematics, vol. 325, pp. 97–107, 2023

2023
[22]

Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction,

Y . Xu, A. Bazarjani, H.-g. Chi, C. Choi, and Y . Fu, “Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9632–9643

2023
[23]

Robust long-term vehicle trajectory prediction using link projection and a situation-aware transformer,

M. Kim, B. I. Kwak, J.-U. Hou, and T. Kim, “Robust long-term vehicle trajectory prediction using link projection and a situation-aware transformer,”Sensors, vol. 24, no. 8, p. 2398, 2024

2024
[24]

Dedicated short-range communications (dsrc) standards in the united states,

J. B. Kenney, “Dedicated short-range communications (dsrc) standards in the united states,”Proceedings of the IEEE, vol. 99, no. 7, pp. 1162– 1182, 2011

2011
[25]

On the properties of neural machine translation: Encoder-decoder approaches,

K. Cho, B. Van Merri ¨enboer, D. Bahdanau, and Y . Bengio, “On the properties of neural machine translation: Encoder-decoder approaches,” arXiv preprint arXiv:1409.1259, 2014

2014 arXiv
[26]

Microscopic traffic simulation using sumo,

P. A. Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .-P. Fl ¨otter¨od, R. Hilbrich, L. L ¨ucken, J. Rummel, P. Wagner, and E. Wießner, “Microscopic traffic simulation using sumo,” inThe 21st IEEE International Conference on Intelligent Transportation Systems. IEEE,
[27]

Planet dump retrieved from https://planet.osm.org ,

OpenStreetMap contributors, “Planet dump retrieved from https://planet.osm.org ,” https://www.openstreetmap.org, 2017

2017
[2018]

Available: https://elib.dlr.de/124092/

[Online]. Available: https://elib.dlr.de/124092/

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [3]

Dealing with changes: Resilient routing via graph neural networks and multi-agent deep reinforcement learning,

S. S. Bhavanasi, L. Pappone, and F. Esposito, “Dealing with changes: Resilient routing via graph neural networks and multi-agent deep reinforcement learning,”IEEE Transactions on Network and Service Management, vol. 20, no. 3, pp. 2283–2294, 2023

work page 2023

[2] [4]

Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case,

P. Almasan, J. Su ´arez-Varela, K. Rusek, P. Barlet-Ros, and A. Cabellos- Aparicio, “Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case,”Computer Communications, vol. 196, pp. 184–194, 2022

2022

[3] [6]

Towards Generalizability of Multi-Agent Reinforcement Learning in Graphs with Recurrent Message Passing

J. Weil, Z. Bao, O. Abboud, and T. Meuser, “Towards generalizability of multi-agent reinforcement learning in graphs with recurrent message passing,”arXiv preprint arXiv:2402.05027, 2024

work page Pith review arXiv 2024

[4] [1]

Interworking of dsrc and cellular network technologies for v2x communications: A survey,

K. Abboud, H. A. Omar, and W. Zhuang, “Interworking of dsrc and cellular network technologies for v2x communications: A survey,”IEEE Transactions on Vehicular Technology, vol. 65, no. 12, pp. 9457–9470, 2016

work page 2016

[5] [2]

Qmr: Q-learning based multi-objective optimization routing protocol for flying ad hoc networks,

J. Liu, Q. Wang, C. He, K. Jaffr `es-Runser, Y . Xu, Z. Li, and Y . Xu, “Qmr: Q-learning based multi-objective optimization routing protocol for flying ad hoc networks,”Computer Communications, vol. 150, pp. 304–316, 2020

work page 2020

[6] [5]

Gddr: Gnn-based data-driven routing,

O. Hope and E. Yoneki, “Gddr: Gnn-based data-driven routing,” in2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS). IEEE, 2021, pp. 517–527

work page 2021

[7] [7]

Graph neural networks for wireless networks: Graph representation, architecture and evaluation,

Y . Lu, Y . Li, R. Zhang, W. Chen, B. Ai, and D. Niyato, “Graph neural networks for wireless networks: Graph representation, architecture and evaluation,”IEEE Wireless Communications, 2024

work page 2024

[8] [8]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

2017