Pith. sign in

REVIEW 3 major objections 5 minor 27 references

A Dynamic Programming Framework for Vehicular Task Offloading with Successive Action Improvement

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a two-time-scale approximate Markov decision process can schedule vehicular task offloading with a guaranteed average-cost ceiling and better simulated performance than deterministic and reinforcement-learning…

desk verdict A solid two-time-scale heuristic with real engineering value, but the advertised average-cost bound is not proven; the paper deserves review with major revision. read the letter →

arxiv 2509.05907 v1 pith:Q3MBGZQ6 submitted 2025-09-07 eess.SY cs.NIcs.SY

Qianren Li , Yuncong Hong , Bojie Lv , Rui Wang This is my paper · ORCID
classification eess.SYcs.NIcs.SY
keywords vehiculartaskoffloadingMarkovdecisionprocessapproximatedynamicprogrammingtwo-time-scaleschedulingsuccessiveactionimprovementedgecomputingrandomvehiclemobilityaveragecostbound
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 tries to establish that task offloading from vehicles with random velocities can be scheduled at near-optimal cost without solving an intractable dynamic program. It models cell association, uplink time, and throughput allocation as a finite-horizon Markov decision process, then approximates the optimal value function by a closed-form expression built from a deterministic reference schedule. Each time slot's actual action is chosen to improve on that reference, which yields a provable upper bound on average total cost. In simulations with a high-fidelity traffic simulator, the scheme outperforms deterministic allocation, pre-allocation only, single-super-slot, and model-free reinforcement learning baselines, and in small instances its cost stays within 1.85% of the optimal policy. If the claim is right, the paper offers a low-complexity, analysable alternative to reinforcement learning for vehicular edge offloading.

What carries the argument

The load-bearing object is the approximate value function $\tilde V_{n,t+1}$ of Lemma 1, a closed-form, asymptotically achievable average future cost that splits into a transmission-time and residual-bits term $f^1_{n,t+1}$ and an energy term $f^2_{n,t+1}$. It is constructed from the deterministic reference schedule $R_\iota$ by averaging over the Markov mobility model and using a high-SNR approximation that treats transmit power as an exponential function of throughput. This approximation gives the value function physical meaning: it represents an actual achievable cost, not a learned surrogate. Lemma 2 then uses the fact that the online action in P3 minimizes current cost plus $\tilde V$ to show the scheduler successively suppresses the remaining cost, which is what delivers the average-cost upper bound without value iteration.

What would settle it

In the paper's traffic-simulator scenario, compute for each slot the exact transmit power needed to deliver the scheduled throughput $r_{n,t}$ under the reference association and time allocation, and compare it with the approximate exponential power used in Lemma 1 and with $P_{\max}=5$ W; if the approximate power exceeds the peak constraint or diverges from the exact power in a nontrivial fraction of slots, the Lemma 2 ceiling is not the cost of an actually achievable policy.

Watch

Extended reading notes

Core claim

The central claim is that a two-time-scale approximate MDP can achieve the same or lower average total cost than its own reference schedule. At the beginning of each super slot, the scheduler computes a reference plan for the remaining slots by deterministic optimization along average vehicle trajectories. Within each slot, it minimizes the immediate cost plus a closed-form estimate of future cost, derived from the reference plan and the vehicles' Markov mobility statistics. Lemma 1 gives that estimate as $\tilde V_{n,t+1} = f^1_{n,t+1} + f^2_{n,t+1}$, where $f^1$ accounts for transmission time and residual bits and $f^2$ for energy, and Lemma 2 states the successive-improvement inequality $C_t(S_t,\tilde\Omega) \le c_t(S_t,A^\mathrm{ref}_{t,\iota}) + \sum_{S_{t+1}} P[S_{t+1}|S_t,A_t] \sum_n \tilde V_{n,t+1}(d_{n,t+1},l_{n,t+1},\Upsilon^{(\iota)}_{n,t+1})$. The implication is that every online decision is no worse than the reference action plus the approximate future cost, so the overall policy has an analysable average-cost ceiling rather than an opaque learned behavior.

Load-bearing premise

The provable cost ceiling in Lemma 2 rests on Lemma 1's closed-form future cost, which assumes a high-SNR power approximation and a peak power constraint large enough to make that power feasible; the simulations set $P_{\max}=5$ W without verifying that the operating point satisfies the approximation.

Editorial extensions

If this is right

  • The scheduler runs in polynomial time: per super slot $O(I_{\max}((NT)^2/\epsilon^2 + NT^2\log_2 T))$ and per time slot $O(I_{\max}(N^2/\epsilon^2 + NT^2\log_2 T))$, instead of exponential value iteration.
  • Because the online action always improves on the reference, shortening the super slot tightens performance; the simulations show the gain saturates when the slot length drops below 10, so frequent reference updates are unnecessary.
  • The paper's bound increases with the expected path-loss factor $\mathbb{E}[\|l_{n,t+\kappa}-l_m\|_2^\gamma \mid l_{n,t}]$, so the performance gain should shrink as vehicle mobility becomes harder to predict.
  • In the simulated 5-to-20 vehicle scenarios, the scheme completes all tested offloading tasks up to 2000 Mb, with lower residual bits and a more stable per-slot energy draw than the baselines.
  • Against a quantized optimal policy in one- and two-vehicle cases, the proposed policy stays at most 1.85% above optimal cost; a perfect-prediction lower bound suggests 15.02% to 26.59% of the cost is inherent trajectory uncertainty.

Reading between the lines

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

  • The same two-time-scale template, a deterministic surrogate schedule plus a closed-form stochastic value approximation, should transfer to other mobility-driven resource allocation problems, such as UAV relaying or robot task offloading, whenever the transition statistics are known.
  • The closed-form $\tilde V$ could serve as a model-based initialization or shaping reward for reinforcement-learning agents, potentially cutting the many training episodes the paper's model-free baseline required.
  • A direct test of the framework's core assumption would replace the high-SNR exponential power formula with the exact Shannon inverse and compare the two in the simulated SNR regime; the gap would isolate whether the provable bound, not just the heuristic, holds at $P_{\max}=5$ W.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies task offloading from N vehicles to an edge server via multiple base stations, with vehicle locations following a stationary Markov chain. The joint design of cell association, uplink time, and throughput allocation is formulated as a finite-horizon MDP. Because the exact optimal value function is intractable, the authors propose a two-time-scale scheme: at each super slot, a deterministic average-trajectory problem P2 produces a reference schedule, and at each time slot an approximate Bellman update P3 selects the actual action using a closed-form approximation \tilde V of the future cost (Lemma 1). The central theoretical claim is Lemma 2, which states that the proposed online policy has average total cost no larger than the reference schedule plus the approximate future cost, i.e., Eq. (40). The paper further reports CARLA-based simulations with 5–20 vehicles comparing the proposed method with deterministic allocation, pre-allocation only, single-super-slot, and PPO baselines, plus small-scale optimality-gap experiments.

Significance. If the claims hold, the paper would provide a rare example of an approximate MDP method with a physically meaningful value approximation and a non-trivial worst-case cost ceiling, together with a low-complexity scheduler and a realistic simulation study. The CARLA-based evaluation, the comparison with an exact optimal policy for small instances, and the closed-form structure of \tilde V are genuine strengths. However, the central guarantee in Lemma 2 is not established by the manuscript's argument, and the numerical experiments do not verify the operating conditions under which the closed-form value function is derived. The contribution is therefore significant but conditional on a fixable, yet load-bearing, proof gap.

major comments (3)
  1. [Section IV-C, Lemma 2, Eq. (40)] The proof of Lemma 2 is a one-sentence assertion that the online scheduling 'successively suppresses' the reference cost, and the stated inequality does not follow from the construction of P3. At time t, P3 minimizes c_t(A_t) + \tilde Q_t(S_t,A_t;R_ι), where \tilde Q_t is defined in Eq. (33) as a minimum over future throughput allocations {Υ_{n,t+1}}. Since a minimum is no larger than the value at the reference throughput Υ^{(ι)}_{n,t+1}, the selected action satisfies c_t(A*_t) + \tilde Q_t(A*_t) ≤ c_t(A^{ref}_{t,ι}) + E[Σ \tilde V_{n,t+1}(·, Υ^{(ι)}_{n,t+1})]. This only bounds current cost plus an optimistic continuation value. The quantity C_t(S_t,\tilde Ω) in Eq. (40) contains the actual future cost of the online policy, not \tilde V evaluated at the optimized Υ. Moreover, the minimizing Υ inside \tilde Q_t is not committed: at time t+1 the algorithm re-solves P3 from the new state and may select different actions. To obtain Eq. (40) one would need an induction proving that the actual online continuation cost is no larger than the optimistic min-over-Υ expectation (or an equivalent monotonicity property). Such an inequality is neither stated nor proved. The advertised guaranteed average-cost ceiling is therefore unsupported as the manuscript stands.
  2. [Appendix A, Eq. (41), and Section VIII] Lemma 1 is derived under a high-SNR approximation of transmit power as an exponential function of throughput and under the requirement that the peak power P_max be sufficiently large so that Eq. (41) holds. The simulations in Section VIII use P_max = 5 W, T_s = 1 s, B = 20 MHz, and path-loss exponent 4, but the paper never checks whether the operating regime satisfies Eq. (41) or whether the high-SNR approximation is accurate for the simulated channels. If \tilde V is not the true achievable average future cost in the simulated regime, then the approximation in Eq. (33) is not valid and the subsequent bound argument in Lemma 2 cannot be applied to the reported numerical results. The authors should either verify Eq. (41) and the high-SNR approximation numerically or state clearly that the simulations are outside the provable regime and separate the experimental findings from the theoretical claim.
  3. [Section VI, P3/P10 and Section VII Lemma 2] The online action A*_t is defined as the solution of P3, and the derivation of Eq. (40) treats it as the true argmin of the approximate Bellman expression. However, Section VI states that P3 is non-convex for t < T and proposes an alternating minimization between P9 and P10, without proving global optimality. P10 is said to be solved 'similarly to the problem P6', but P6 was solved by case comparisons for a deterministic allocation problem; no proof is given that the alternating procedure converges to the global minimizer of P3, or even to a point whose current-plus-approximate-future cost is no larger than that of the reference action. Since the inequality behind Lemma 2 needs the online action to be no worse than the reference action with respect to the surrogate objective, this missing optimality guarantee is load-bearing. The authors should either prove the needed property for their alternating minimization or modify the algorithm (for example, by explicitly keeping the reference action as a feasible candidate) and then restate Lemma 2 accordingly.
minor comments (5)
  1. [Notation section] The Landau notation definition is incorrect: O(n) means the absolute value is bounded above by c n for sufficiently large n, not bounded below.
  2. [Lemma 1 and Appendix A] The lemma is titled 'Asymptotically Achievable Average Cost', but the proof does not identify an asymptotic parameter; the condition is a 'sufficiently large' P_max satisfying Eq. (41). The terminology should be aligned with the actual assumptions.
  3. [Appendix A, power expression] In the derivation of P_{n,t+κ}, the exponent uses τ^{(ι)}_{n,t} in the denominator, but the throughput r_{n,t+κ} refers to time slot t+κ; this appears to be a typo for τ^{(ι)}_{n,t+κ}, and the table entries for f_2 should be checked for the same indexing issue.
  4. [Section VII, Remark 2] The claim that the upper bound increases monotonically with E[∥l_{n,t+κ}-l_m∥^γ | l_{n,t}] is not proved; if it is intended as part of the theoretical contribution, a short argument should be supplied, otherwise it should be marked as an observation.
  5. [Section VIII, Figures 4–7] The paper reports only mean values over 500 trials, with no confidence intervals or statistical significance tests for the reported gains over PPO and the other baselines; given the variance visible in the per-slot energy plots, this weakens the strength of the comparative claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the value-function approximation is derived from stated channel/mobility assumptions, and the simulated gains are measured against independent baselines.

full rationale

The paper's derivation chain is self-contained rather than circular. The MDP model and Bellman equations in Section III are standard; Lemma 1 in Appendix A derives a closed-form approximate value function from stated high-SNR, ergodic-capacity, and sufficiently-large-peak-power assumptions, not from the target result. The online scheduler in P3 optimizes the current cost plus this approximate continuation value, and Lemma 2 claims a performance bound relative to the reference schedule. The closest concern is that Lemma 2's proof is a one-sentence assertion and the min-over-future-throughput in (33) makes the optimization surrogate optimistic relative to the reference continuation value; however, this is a proof gap or soundness issue, not a reduction of the predicted bound to the paper's own inputs. The reference schedule is obtained by an independent deterministic optimization (P2), and the baselines DA, PAO, and PPO are external to the proposed derivation. The use of PAS3, 'similar to our prior work [1]', is a comparison against the authors' own earlier algorithm, but it is a simulation baseline rather than a load-bearing premise for the analytic claims. Self-citations such as [1] and [20] are contextual or used for trajectory statistics and do not carry the derivation. Therefore no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework is an engineering construction: it assumes known trajectory statistics, ergodic interference-averaged channels, a high-SNR power model, and globally solved subproblems. Those assumptions, not new physics, carry the result. No new physical entities are invented.

free parameters (4)
  • Cost weights ω1, ω2 = ω1=2, ω2=5
    Chosen for the main simulation sweep; they set the Pareto tradeoff among transmission time, energy and residual bits, and Fig 9 shows the proposed method's advantage over baselines nearly disappears at ω1=0.1.
  • Super slot length Γ = 10
    Set to 10 time slots in all main experiments; Fig 8 shows average cost improves as Γ shrinks, so the reported gain is partly a function of this hand-picked value.
  • Quantization levels for optimality-gap MDP = 21
    Used in the 1-2 vehicle optimality comparison; the '1.85% worst-case gap' is relative to this quantized MDP, not the continuous-state problem.
  • Peak power P_max = 5 W
    Simulation parameter; the theoretical value function assumes P_max is sufficiently large for the high-SNR power expression, but the paper does not check this in the simulated regime.
assumptions (5)
  • domain assumption Vehicle trajectories are stationary Markov chains with known transition matrices P_n.
    Stated in Section II-A, Eq (3); the scheduler is assumed to know P_n, estimated from CARLA traces in the authors' own prior work [20], so the method's tractability depends on this statistics being accurate and stationary.
  • domain assumption Uplink slots are long enough for ergodic capacity, and co-channel interference at each BS is stationary.
    Section II-B before Eq (8); this justifies replacing instantaneous rate by the expectation over fast fading h and interference I_m. It is not validated in the simulation.
  • ad hoc to paper High-SNR approximation and a sufficiently large peak power make transmit power an exact exponential function of allocated throughput.
    Appendix A, Eq (32) and surrounding text; this is the key step that makes the value function \tilde V closed-form. The paper does not verify the regime in the CARLA simulations with P_max=5W.
  • ad hoc to paper The alternating minimization for P3 and P10 finds the global optimum, or at least a solution satisfying the inequality used in Lemma 2.
    Section VI labels the result A*_t but the solver alternates between convex relaxations and does not state a global optimality or even a descent guarantee; Lemma 2's induction requires the online step to be no worse than the reference action at every state.
  • domain assumption Relax-then-round BS association, selecting the BS with largest relaxed e*, is a valid approximation.
    Used in Sections V and VI, Eq (37)-(39); no bound on the rounding loss is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Dynamic Programming Framework for Vehicular Task Offloading with Successive Action Improvement." pith.science (2026). https://pith.science/paper/Q3MBGZQ6

@misc{pith2026250905907,
  author       = {Pith},
  title        = {Pith review of: A Dynamic Programming Framework for Vehicular Task Offloading with Successive Action Improvement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3MBGZQ6}},
  note         = {Machine review of arXiv:2509.05907}
}
read the original abstract

In this paper, task offloading from vehicles with random velocities is optimized via a novel dynamic programming framework. Particularly, in a vehicular network with multiple vehicles and base stations (BSs), computing tasks of vehicles are offloaded via BSs to an edge server. Due to the random velocities, the exact locations of vehicles versus time, namely trajectories, cannot be determined in advance. Hence, instead of deterministic optimization, the cell association, uplink time, and throughput allocation of multiple vehicles during a period of task offloading are formulated as a finite-horizon Markov decision process. In order to derive a low-complexity solution algorithm, a two-time-scale framework is proposed. The scheduling period is divided into super slots, each super slot is further divided into a number of time slots. At the beginning of each super slot, we first obtain a reference scheduling scheme of cell association, uplink time and throughput allocation via deterministic optimization, yielding an approximation of the optimal value function. Within the super slot, the actual scheduling action of each time slot is determined by making improvement to the approximate value function according to the system state. Due to the successive improvement framework, a non-trivial average cost upper bound could be derived. In the simulation, the random trajectories of vehicles are generated from a high-fidelity traffic simulator. It is shown that the performance gain of the proposed scheduling framework over the baselines is significant.

Figures

Figures reproduced from arXiv: 2509.05907 by the authors.

Figure 1
Figure 1. Conceptual illustration of vehicular task offloading. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed dynamic improvement [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of simulation scenario, where two samples [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The comparison of (a) average total cost, (b) average transmission time, (c) average energy cost and (d) average residual [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The comparison of (a) average total cost, (b) average transmission time, (c) average energy cost and (d) average residual [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Illustration of per-slot energy consumption. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: The average total cost versus vehicle number. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: The average total cost versus super slot length. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: The comparison of (a) average total cost, (b) average transmission time, (c) average energy cost and (d) average residual [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 11
Figure 11. Figure 11: Illustration of optimality gap, where the average total [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    A Dynamic Improvement Framework for Vehicular Task Offloading,

    Q. Li, Y . Hong, B. Lv, and R. Wang, “A Dynamic Improvement Framework for Vehicular Task Offloading,” in2025 IEEE Wireless Communications and Networking Conference (WCNC), Mar. 2025, pp. 1–7

    work page 2025

  2. [2]

    Architectural Design Alternatives Based on Cloud/Edge/Fog Computing for Connected Vehicles,

    H. Wang, T. Liu, B. Kim, C.-W. Lin, S. Shiraishi, J. Xie, and Z. Han, “Architectural Design Alternatives Based on Cloud/Edge/Fog Computing for Connected Vehicles,”IEEE Communications Surveys & Tutorials, vol. 22, no. 4, pp. 2349–2377, Fourthquarter 2020

    work page 2020

  3. [3]

    [Online]

    (2024) Uber’s US Safety Report. [Online]. Available: https://www.uber.com/us/en/about/reports/us-safety-report/

    work page 2024

  4. [4]

    Learning-Based Queuing Delay-Aware Task Offloading in Collaborative Vehicular Networks,

    Z. Jia, Z. Zhou, X. Wang, and S. Mumtaz, “Learning-Based Queuing Delay-Aware Task Offloading in Collaborative Vehicular Networks,” in ICC 2021 - IEEE International Conference on Communications, Jun. 2021, pp. 1–6

    work page 2021

  5. [5]

    Joint Offloading De- cision and Resource Allocation in MEC-enabled Vehicular Networks,

    L. Zhang, Y . Sun, Y . Tang, H. Zeng, and Y . Ruan, “Joint Offloading De- cision and Resource Allocation in MEC-enabled Vehicular Networks,” in2021 IEEE 93rd Vehicular Technology Conference (VTC2021-Spring), Apr. 2021, pp. 1–5

    work page 2021

  6. [6]

    Energy-Efficient Cooperative Offloading for Edge Computing-Enabled Vehicular Networks,

    H. Cho, Y . Cui, and J. Lee, “Energy-Efficient Cooperative Offloading for Edge Computing-Enabled Vehicular Networks,”IEEE Transactions on Wireless Communications, vol. 21, no. 12, pp. 10 709–10 723, Dec. 2022

    work page 2022

  7. [7]

    Com- putation Offloading in Heterogeneous Vehicular Edge Networks: On- Line and Off-Policy Bandit Solutions,

    A. Bozorgchenani, S. Maghsudi, D. Tarchi, and E. Hossain, “Com- putation Offloading in Heterogeneous Vehicular Edge Networks: On- Line and Off-Policy Bandit Solutions,”IEEE Transactions on Mobile Computing, vol. 21, no. 12, pp. 4233–4248, Dec. 2022

    work page 2022

  8. [8]

    Maximizing spatial– temporal coverage in mobile crowd-sensing based on public transports with predictable trajectory,

    C. Wang, C. Li, C. Qin, W. Wang, and X. Li, “Maximizing spatial– temporal coverage in mobile crowd-sensing based on public transports with predictable trajectory,”International Journal of Distributed Sensor Networks, vol. 14, no. 8, p. 155014771879535, Aug. 2018

    work page 2018

Show all 27 references
  1. [9]

    Priority-Aware Task Offloading in Vehicular Fog Computing Based on Deep Reinforcement Learning,

    J. Shi, J. Du, J. Wang, J. Wang, and J. Yuan, “Priority-Aware Task Offloading in Vehicular Fog Computing Based on Deep Reinforcement Learning,”IEEE Transactions on Vehicular Technology, vol. 69, no. 12, pp. 16 067–16 081, Dec. 2020

    2020

  2. [10]

    Fundamentals of Transportation Engineering,

    J. D. Fricker and R. K. Whitford, “Fundamentals of Transportation Engineering,”A Multimodal Systems Approach. Inc. Upper Saddle River, New Jersey, USA, 2004

    2004

  3. [11]

    Dynamic Trajectory and Offloading Control of UA V-enabled MEC under User Mobility,

    Z. Yang, S. Bi, and Y .-J. A. Zhang, “Dynamic Trajectory and Offloading Control of UA V-enabled MEC under User Mobility,” in2021 IEEE Inter- national Conference on Communications Workshops (ICC Workshops), 2021, pp. 1–6

    2021

  4. [12]

    Popularity-Aware Online Task Offloading for Heterogeneous Vehicular Edge Computing Using Contextual Clustering of Bandits,

    Y . Lin, Y . Zhang, J. Li, F. Shu, and C. Li, “Popularity-Aware Online Task Offloading for Heterogeneous Vehicular Edge Computing Using Contextual Clustering of Bandits,”IEEE Internet of Things Journal, vol. 9, no. 7, pp. 5422–5433, Apr. 2022

    2022

  5. [13]

    MDP- Based Task Offloading for Vehicular Edge Computing Under Certain and Uncertain Transition Probabilities,

    X. Zhang, J. Zhang, Z. Liu, Q. Cui, X. Tao, and S. Wang, “MDP- Based Task Offloading for Vehicular Edge Computing Under Certain and Uncertain Transition Probabilities,”IEEE Transactions on Vehicular Technology, vol. 69, no. 3, pp. 3296–3309, Mar. 2020

    2020

  6. [14]

    Delay-Tolerant Data Traffic to Software- Defined Vehicular Networks With Mobile Edge Computing in Smart City,

    M. Li, P. Si, and Y . Zhang, “Delay-Tolerant Data Traffic to Software- Defined Vehicular Networks With Mobile Edge Computing in Smart City,”IEEE Transactions on Vehicular Technology, vol. 67, no. 10, pp. 9073–9086, Oct. 2018

    2018

  7. [15]

    Handover-Enabled Dynamic Computation Offloading for Vehicular Edge Computing Networks,

    H. Maleki, M. Bas ¸aran, and L. Durak-Ata, “Handover-Enabled Dynamic Computation Offloading for Vehicular Edge Computing Networks,” IEEE Transactions on Vehicular Technology, vol. 72, no. 7, pp. 9394– 9405, Jul. 2023

    2023

  8. [16]

    Deep Reinforcement Learning for Offloading and Resource Allocation in Vehicle Edge Computing and Networks,

    Y . Liu, H. Yu, S. Xie, and Y . Zhang, “Deep Reinforcement Learning for Offloading and Resource Allocation in Vehicle Edge Computing and Networks,”IEEE Transactions on Vehicular Technology, vol. 68, no. 11, pp. 11 158–11 168, Nov. 2019

    2019

  9. [17]

    Mobility Models for Vehicular Ad Hoc Networks: A Survey and Taxonomy,

    J. Harri, F. Filali, and C. Bonnet, “Mobility Models for Vehicular Ad Hoc Networks: A Survey and Taxonomy,”IEEE Communications Surveys & Tutorials, vol. 11, no. 4, pp. 19–41, Fourth 2009

    2009

  10. [18]

    Bertsekas,Dynamic Programming and Optimal Control, 4th ed

    D. Bertsekas,Dynamic Programming and Optimal Control, 4th ed. Belmont, MA, USA: Athena Scientific, 2012, vol. 2

    2012

  11. [19]

    Distributed Dynamic Map Fusion via Federated Learning for Intelligent Networked Vehicles,

    Z. Zhang, S. Wang, Y . Hong, L. Zhou, and Q. Hao, “Distributed Dynamic Map Fusion via Federated Learning for Intelligent Networked Vehicles,” in2021 IEEE International Conference on Robotics and Automation (ICRA), May 2021, pp. 953–959

    2021

  12. [20]

    FedCars: An Efficient Scheduling Framework for In-Vehicle Federated Learning In- tegrated With CARLA-Cosimulation,

    Y . Hong, B. Lv, S. Wang, R. Wang, H. Tan, and F. C. Lau, “FedCars: An Efficient Scheduling Framework for In-Vehicle Federated Learning In- tegrated With CARLA-Cosimulation,”IEEE Transactions on Vehicular Technology, pp. 1–14, 2024

    2024

  13. [21]

    W. B. Powell,Approximate Dynamic Programming: Solving the Curses of Dimensionality. John Wiley & Sons, Oct. 2007

    2007

  14. [22]

    Monotonic Value Function Factorisation for Deep Multi- agent Reinforcement Learning,

    T. Rashid, M. Samvelyan, C. S. De Witt, G. Farquhar, J. Foerster, and S. Whiteson, “Monotonic Value Function Factorisation for Deep Multi- agent Reinforcement Learning,”Journal of Machine Learning Research, vol. 21, no. 178, pp. 1–51, Jan. 2020

    2020

  15. [23]

    Value-decomposition Networks for Cooperative Multi-agent Learning,

    P. Sunehag, G. Lever, A. Gruslys, W. M. Czarnecki, V . Zambaldi, M. Jaderberg, M. Lanctot, N. Sonnerat, J. Z. Leibo, K. Tuylset al., “Value-decomposition Networks for Cooperative Multi-agent Learning,” arXiv preprint arXiv:1706.05296, 2017

    2017 arXiv

  16. [24]

    Beck,First-Order Methods in Optimization

    A. Beck,First-Order Methods in Optimization. Philadelphia, PA: Society for Industrial and Applied Mathematics, 2017. [Online]. Available: https://epubs.siam.org/doi/abs/10.1137/1.9781611974997

    2017 doi

  17. [25]

    CARLA: An Open Urban Driving Simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “CARLA: An Open Urban Driving Simulator,” inConference on robot learning (CoLR). PMLR, 2017, pp. 1–16

    2017

  18. [26]

    Proxi- mal Policy Optimization Algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proxi- mal Policy Optimization Algorithms,”arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  19. [27]

    S. P. Boyd and L. Vandenberghe,Convex Optimization. Cambridge, UK ; New York: Cambridge University Press, 2004

    2004

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.