Generalizable Pareto-Optimal Offloading with Reinforcement Learning in Mobile Edge Computing

Junrui Wen; Meng Zhang; Ming Tang; Ning Yang

REVIEW 4 major objections 6 minor 39 references

Generalizable Pareto-Optimal Offloading with Reinforcement Learning in Mobile Edge Computing

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A single preference-conditioned RL policy can reproduce, within 0.3% hypervolume, a 101-policy Pareto front for MEC offloading and generalize to unseen CPU frequencies and server counts.

desk verdict Solid empirical MORL-for-MEC paper with a weak theory section; the Pareto front and generalization results are plausible and useful, but Theorem 2 is false as stated and the T_off < Δt assumption is unverified. read the letter →

arxiv 2509.10474 v1 pith:D3UEGCFS submitted 2025-08-27 eess.SY cs.SY

Ning Yang , Junrui Wen , Meng Zhang , Ming Tang This is my paper

classification eess.SYcs.SY

keywords mobileedgecomputingmulti-objectivereinforcementlearningtaskoffloadingParetofrontpreference-conditionedpolicydomainrandomizationdiscretesoftactor-criticgeneralization

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

This paper tries to show that one reinforcement-learning policy can replace many. For a mobile edge computing system where tasks must be offloaded to servers under competing objectives (delay vs. energy), it trains a single Discrete-SAC model that takes the user's preference weighting as an input, and it randomizes CPU frequencies and the number of edge servers during training. The result is a policy that traces a Pareto front across preferences, and in simulation its hypervolume is 64.1 versus 64.3 for 101 separately trained policies (a 0.3% gap), while beating the LinUCB, simulated-annealing, NSGA-II, and random baselines by 10.7% to 121.0%. The same policy also handles CPU frequencies and edge-server counts outside its training range, with a 1.7% hypervolume error against a policy trained directly on the test distribution. If the result transfers outside simulation, the practical gain is a single deployable model instead of a per-preference, per-system set of policies.

What carries the argument

The load-bearing mechanism is a preference-conditioned contextual MOMDP solved by Discrete-SAC, with three supports: a histogram-based state encoding that summarizes each server's unfinished workload; a masked action space expanded to the maximum possible number of servers, so a fixed-size network can output a valid action for any number of edge servers; and a delay reward (Eq. 29) that accounts for how a new offload decision shifts the residual delays of tasks already running on that server. The masked softmax operator is what allows one policy to be reused at different server counts, and the domain-randomized training over contexts is what makes CPU-frequency generalization possible.

What would settle it

Run the Section 5 comparison with the interference-corrected rate C'_{u,e} (Eq. 4) or with a step duration Δt short enough that offloading delay exceeds it. If the GMORL hypervolume advantage over LinUCB collapses or the 0.3% gap to the 101-policy upper bound widens substantially, the delay-reward model in Eq. (29) is what carried the result.

Watch

Extended reading notes

Core claim

The central discovery is that multi-objective offloading does not require a separate policy per preference or per system configuration. Given a preference vector ω (the weight placed on delay vs energy), the GMORL policy picks a server for each arriving task, and the same weights also enter the reward that trains the actor and critics. By sampling preferences, CPU frequencies, and up to eight edge servers during training, and by masking the action distribution to however many servers are actually present, the agent learns a near-Pareto-optimal mapping from preference to offloading decision. In the paper's simulations, the single policy's delay-energy Pareto front reaches hypervolume 64.1, wi

Load-bearing premise

The delay estimates that train the policy assume offloading takes less than one time step and that the uplink is interference-free; in dense real networks both can fail, which would bias the reward and the reported Pareto fronts.

Editorial extensions

If this is right

Deployment cost drops from one trained policy per objective weighting to one policy whose input includes the weighting; runtime preference changes need no retraining, only a new ω.
A fixed neural network can schedule systems with any number of servers up to its trained maximum, because the action mask zeroes out unavailable servers.
Out-of-distribution generalization to unseen CPU frequencies is measurable and small in aggregate: 1.7% hypervolume error in the paper's test setup.
In the paper's simulations, the single policy dominates contextual-bandit and heuristic baselines (10.7-121.0% hypervolume improvement) and is nearly indistinguishable from the multi-policy upper bound (0.3% gap).
The provided convergence and performance-difference bounds give a stability guarantee for the repeated policy updates, assuming the reward model is accurate.

Reading between the lines

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

Because preference ω is an input to the policy, online preference switching should work without retraining; the paper only evaluates fixed preferences per episode, so this is an untested extension.
The masked histogram architecture is resource-count agnostic and could be transferred to other variable-size scheduling domains (e.g., cloud container placement), but the paper's evidence is only for MEC.
The 1.7% generalization error is an aggregate over the tested contexts; worst-case contexts are not reported and could be larger than the average.
If the uplink interference correction of Eq. (4) is added, the framework's structure does not change, so the method is plausibly extensible to denser, non-ideal channels; the paper does not test this.

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

First, the practical core: single policy, preference-conditioned Discrete-SAC with a masked action space and histogram state encoding, trained with domain randomization, and tested over varied CPU frequencies and server counts. The headline numbers are close to the multi-policy upper bound (64.1 vs 64.3 hypervolume, 0.3% gap), and the generalization error versus a reference policy is 1.7%. If those numbers hold across seeds, this is a genuinely useful engineering result: one model instead of 101, deployable to systems not seen in training. The paper's Table 1 is also honest: no prior work covers multi-preference, system-parameter, and server-quantity generalization together.

The theory section, by contrast, is not in the same shape. Theorem 2 is not a theorem: the step from Eq. (A8) to Eq. (A9) is a jump, the L1 norm of policies is never defined, and the claimed positive lower bound contradicts the possibility that a policy update makes things worse, which is routine in RL. The bound can be negative. This should be removed or reworked. Theorem 1's convergence proof is the tabular Bellman contraction argument applied to SAC with function approximation; it does not handle function approximation error or the policy network, so the claim is stronger than the proof supports.

The reward model also rests on an unverified assumption: T_off < Δt, stated before Eq. (28). The state vector has no placeholder for tasks still in transit, so if offloading delay exceeds the step duration, the MDP becomes partially observable and the delay reward is biased. The paper does not report the distribution of T_off or the channel parameters needed to compute it. This might be fine in the simulated regime, but the paper has to show it.

Experimental reporting: no error bars or seed information anywhere. For claims that are about small gaps, that is a real weakness, not a cosmetic one.

All that said, the empirical design is coherent, and the authors appear to understand the prior work. The GitHub link is a plus if the repo actually ships.

Bottom line: worth peer review. The engineering result is real, the central empirical claim is well supported, and the flaws are fixable by deleting Theorem 2, softening Theorem 1, adding variance, and checking the T_off condition. I would not desk-reject this.

Referee Report

4 major / 6 minor

Summary. The paper studies multi-objective task offloading in mobile edge computing under unknown user preferences and heterogeneous MEC system parameters. It proposes GMORL, a single-policy multi-objective deep RL framework based on Discrete-SAC, with a contextual MOMDP formulation, a histogram-based state encoding for multiple servers, a scalarized delay/energy reward, and a masked-action neural architecture that supports varying numbers of edge servers. The main empirical claims are that GMORL achieves a Pareto front close to a 101-policy MORL upper bound (hypervolume 64.1 vs. 64.3, a 0.3% gap), outperforms several benchmarks by up to 121.0% in hypervolume, and generalizes to unseen CPU frequencies and to server quantities outside the training range (e.g., a 1.7% hypervolume gap versus a reference policy trained on the test distribution). The paper also states convergence and performance-difference theorems in Section 4 and proves them in the appendix.

Significance. If the empirical results are reliable, the paper makes a useful engineering contribution: it shows that a single preference-conditioned policy with domain randomization, histogram state encoding, and masked action heads can approximate the Pareto front of a multi-policy MORL approach while generalizing across system parameters and server counts. This is a practically relevant result for MEC offloading, and the authors provide a public code repository, which is a strength. However, the theoretical guarantees in Section 4 are not established at the claimed level: Theorem 2's proof has a gap, and Theorem 1 is a generic contraction argument that does not cover the actual deep-RL, domain-randomized procedure. The empirical validity also rests on an unverified assumption about offloading delay relative to the decision-step duration. These issues do not necessarily invalidate the central simulation-based contribution, but they require substantial revision before the paper can be accepted.

major comments (4)

[§4.3 / Appendix D.3 (Theorem 2, Eq. (48), Eqs. (A5)-(A9))] The proof of Theorem 2 does not establish the claimed lower bound. Eq. (A8) is a weighted sum of offloading-decision differences; for an arbitrary adjacent pair of policies this expression can be negative, whereas A||πt − πt+1||1 is nonnegative. In fact, if πt+1 is worse than πt on the scalarized objective, then ∆J < 0, making the claimed positive lower bound false. Moreover, ||πt − πt+1||1 is never defined for policies in the contextual MOMDP; if it is meant as an L1 distance on decision vectors, Eq. (A8) does not imply it. The constant A also mixes quantities with different units (Φm,e is an energy, ωT is dimensionless), so the bound is dimensionally inconsistent. I recommend either removing Theorem 2 or restating it as a conditional bound under an explicit improvement assumption that makes ∆J ≥ 0.
[§3.1, Eqs. (18), (28)-(29)] The reward and state design rely on the assumption, stated just before Eq. (28), that the offloading delay satisfies T_off_m < ∆t. If T_off_m ≥ ∆t for a non-negligible fraction of tasks, the task is still in flight at the next decision epoch, but the state st,e in Eq. (18) — including the histogram Be, which only tracks executed residual sizes — contains no information about in-flight tasks. The environment then becomes partially observable / non-Markovian, and the expression in Eq. (28) incorrectly inserts the task into the server queue at τ′_t before the next state is observed. The paper does not report the distribution of T_off_m/∆t in its simulations, nor the channel parameters (noise power σ², path loss, distribution of |h|²) needed to evaluate Eq. (2). Since Eq. (2) also assumes an ideal interference-free uplink, the learned delay reward may be biased relative to realistic channels
[§4.2 / Appendix D.1 (Theorem 1)] Theorem 1 states that GMORL converges to the optimal policy π* and Q* as the number of epochs and update rounds grow. The proof in Appendix D.1 applies the Banach fixed-point theorem to a Bellman operator with a fixed scalarized reward. This does not cover the actual algorithm: the policy and Q-functions are deep neural networks updated by stochastic gradients, the context is resampled from C on every episode, and the entropy temperature αH is adapted. The phrase 'sufficiently diverse action-state space' is not defined, and no argument is given for convergence of the non-tabular, contextualized objective. I recommend either weakening Theorem 1 to an empirical convergence statement or proving convergence only for a tabular or linear MDP variant with the exact update rules used.
[§5.3.2, Fig. 7] The generalization claim for server quantities is load-bearing for the paper's central message, but Fig. 7 reports only a qualitative 'certain gap' and 'moderate level of concordance' for E = 9, which is outside the training context space CE = {1,...,8}. No hypervolume or other error metric is reported for this case, in contrast to the CPU-frequency generalization experiment where a 1.7% gap is given. Please quantify the E = 9 result and, ideally, report standard errors across random seeds for all generalization numbers so the reader can assess the stability of the claimed 0.3% and 1.7% gaps.

minor comments (6)

[Eq. (30) and Eq. (42)] The scalarized reward is written as rω(st, at) = ωT × (αT rT, αE rE), which is a vector-valued expression and omits ωE. It should be ωT αT rT + ωE αE rE (and similarly in Eq. (42)). Please correct the notation throughout.
[Appendix C.1] The 'Task Delay' definition repeats the 'Energy Consumption' definition verbatim ('The total energy consumption ...'); it should state the total delay and delay per Mbit.
[Abstract] The sentence 'Our code are avavilable at ...' contains typos; should be 'Our code is available at ...'.
[§5.3.2, Fig. 5] The text says the training edge CPU frequency range is [1.75, 2.25] GHz and the testing range is [2.00, 2.50] GHz, but Table A1 lists the testing edge CPU range as [1.5, 2.5] GHz. Please clarify which range is used in Fig. 5.
[Table 2] The simulation setup lists bandwidth and offloading power but omits the AWGN noise power σ², path-loss model, and channel coefficient distribution needed to instantiate Eq. (2). These parameters should be reported so the delay model is reproducible.
[Eq. (4)] The interference term in Eq. (3) sums over all users u′ including u itself, which makes Eq. (4) include the user's own signal as self-interference. Usually the self-interference term should be excluded, i.e., u′ ≠ u. Please correct or clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training reward is the optimized objective by design, and the self-cited baseline [24] is re-run empirically rather than used as a load-bearing derivation.

full rationale

GMORL trains a Discrete-SAC policy to maximize the scalarized reward rω = ωT(αT rT, αE rE) (Eq. 30) and is then evaluated on the same delay/energy model (Eqs. 5-11). This is standard RL objective-evaluation alignment, not a hidden fit: no parameter is fitted to the evaluation metric and then renamed as a prediction. The reward-shaping identity RT = -Σ Tm (Eq. 25) is the algorithm's design target, not an independent forecast. The main empirical claims (0.3% gap to the multi-policy MORL upper bound, 1.7% generalization error) compare GMORL against a reference policy and against the authors' prior multi-policy scheme [24]; both are re-implemented and run in this paper's simulation, so [24] is an empirical baseline rather than an unverified self-citation used to prove the central claim. The paper's assumptions (e.g., T_off < Δt before Eq. 28; interference-free uplink Eq. 2) are correctness and deployment risks, and the proof of Theorem 2 contains questionable inequality directions (Appendix D.3), but these are not cases where an output reduces by construction to an input or where a self-citation supplies the derivation. No circular step was found.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central empirical claim rests on a set of domain assumptions (exponential tasks, Poisson arrivals, ideal uplink, binary offloading, fast offloading) and several unreported or hand-tuned parameters (reward scaling α_T, α_E, histogram bin count N). The theoretical guarantees in Section 4 are not needed to reproduce the simulations, but Theorem 2's proof has an unproven step.

free parameters (5)

Reward scaling coefficients α_T and α_E = Not reported
Used in Eq. (30) to scale delay and energy rewards to the same order of magnitude; values are not given in the paper, making the scalarized reward and thus the learned policy not fully specified.
Histogram bin count N = Not reported
The histogram state encoding Be(τt) in Eq. (19) discretizes residual task sizes into N bins; N is not specified, which affects state representation and generalization.
Entropy temperature α_H = 0.05
Fixed in Table II with learning rate λα=0, so it is a hand-set hyperparameter controlling exploration.
Training context ranges (Cf0, CfE', CE) = Cf0=[3.5,4.5] GHz, CfE'=[1.75,2.25] GHz, CE={1,...,8}
The generalization claim is measured against these training context spaces; the choice of ranges and the number of preferences Ng=64 are experimental design decisions that affect the reported generalization gap.
Hypervolume reference point = Max delay and energy across all Pareto fronts
The reference point in Eq. (16) is chosen post hoc, which affects all hypervolume numbers (64.1, 57.9, etc.) and the reported 121.0% improvement.

assumptions (6)

domain assumption Tasks arrive according to independent Poisson processes with exponential sizes; channel gains are Rayleigh fading with AWGN noise (Eq. (2)).
Section 2.1 uses these distributions for arrivals, task sizes, and rates; simulation relies on them.
domain assumption Each task is either fully offloaded to exactly one server or executed locally (binary offloading).
Constraint (12b)-(12c) in Problem (12), stated in Section 2.1.
domain assumption Offloading delay T_off^m is shorter than the step duration Δt.
Stated before Eq. (28) in Section 3.1: 'We consider a high-speed communication system that offloading delay T_off^m is shorter than the duration of one step Δt'. The delay reward Eq. (29) depends on this.
domain assumption The uplink is interference-free and only AWGN is present (Eq. (2)), not the interference-limited rate of Eq. (4).
Section 2.1 says 'To simplify the analysis of the initial model, it is assumed here that the uplink is in an ideal interference-free communication environment'. The reward and experiments appear to use this simplified model.
standard math The Bellman operator is a contraction even with function approximation and stochastic gradient updates, so Q-learning converges (Appendix D.1).
Theorem 1's proof applies the Banach fixed-point theorem to the tabular Bellman operator and asserts convergence of the neural network policy; this ignores approximation error.
ad hoc to paper The lower bound in Theorem 2, Eq. (A9) ΔJ ≥ A||πt - πt+1||1, follows from Eq. (A8).
This step is asserted without a definition of the L1 policy norm or a derivation connecting the per-task offloading decision differences to the norm; it does not follow from (A8) when some offloading decisions move in the worse direction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Generalizable Pareto-Optimal Offloading with Reinforcement Learning in Mobile Edge Computing." pith.science (2026). https://pith.science/paper/D3UEGCFS

@misc{pith2026250910474,
  author       = {Pith},
  title        = {Pith review of: Generalizable Pareto-Optimal Offloading with Reinforcement Learning in Mobile Edge Computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3UEGCFS}},
  note         = {Machine review of arXiv:2509.10474}
}

abstract

Mobile edge computing (MEC) is essential for next-generation mobile network applications that prioritize various performance metrics, including delays and energy efficiency. However, conventional single-objective scheduling solutions cannot be directly applied to practical systems in which the preferences (i.e., the weights of different objectives) are often unknown or challenging to specify in advance. In this study, we formulate a multi-objective offloading problem for MEC with multiple edges to minimize the sum of expected long-term energy consumption and delay while considering unknown preferences. To address the challenge of unknown preferences and the potentially diverse MEC systems, we propose a generalizable multi-objective (deep) reinforcement learning (GMORL)-based tasks offloading framework, which employs the Discrete Soft Actor-Critic (Discrete-SAC) method. Our method uses a single policy model to efficiently schedule tasks based on varying preferences and adapt to heterogeneous MEC systems with different CPU frequencies and server quantities. Under the proposed framework, we introduce a histogram-based state encoding method for constructing features for multiple edges in MEC systems, a sophisticated reward function for accurately computing the utilities of delay and energy consumption, and a novel neural network architecture for improving generalization. Simulation results demonstrate that our proposed GMORL scheme enhances the hypervolume of the Pareto front by up to $121.0\%$ compared to benchmarks. Our code are avavilable at https://github.com/gracefulning/Generalizable-Pareto-Optimal-Offloading-with-Reinforcement-Learning-in-Mobile-Edge-Computing

Figures

Figures reproduced from arXiv: 2509.10474 by the authors.

**Figure 2.** Pareto fronts of the proposed GMORL algorithm and benchmark algorithms. [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗

**Figure 3.** Pareto fronts of the proposed GMORL algorithm. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Total task delay and energy consumption with different preferences. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

**Figure 5.** Figure 5: Pareto fronts of GMORL policy and reference policy when [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

**Figure 6.** Figure 6: CPU frequency generalization experiment when [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

**Figure 7.** Figure 7: Pareto fronts of GMORL policy and reference policy when [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages

[1]

Farhan Pervez, Ajmery Sultana, Cungang Yang, and Lian Zhao. Energy and latency efficient joint communication and computation optimization in a multi-uav-assisted mec network.IEEE Transactions on Wireless Communications, 23(3):1728–1741, 2024

work page 2024
[2]

Deep reinforcement learning based computa- tion offloading and resource allocation for mec

Ji Li, Hui Gao, Tiejun Lv, and Yueming Lu. Deep reinforcement learning based computa- tion offloading and resource allocation for mec. In 2018 IEEE Wireless Communications and Networking Conference (WCNC), pages 1–6. IEEE, 2018

work page 2018
[3]

Optimal Task Assignment and Power Allocation for NOMA Mobile-Edge Computing Networks

Fang Fang, Yanqing Xu, Zhiguo Ding, Chao Shen, Mugen Peng, and George K Karagiannidis. Optimal task assignment and power allocation for noma mobile-edge computing networks. arXiv preprint arXiv:1904.12389, 2019

work page Pith review arXiv 1904
[4]

Joint task offloading and resource allocation for multi-server mobile-edge computing networks

Tuyen X Tran and Dario Pompili. Joint task offloading and resource allocation for multi-server mobile-edge computing networks. IEEE Transactions on Vehicular Technology, 68(1):856– 868, 2018

work page 2018
[5]

Latency and energy optimization for mec enhanced sat-iot networks

Gaofeng Cui, Xiaoyao Li, Lexi Xu, and Weidong Wang. Latency and energy optimization for mec enhanced sat-iot networks. IEEE Access, 8:55915–55926, 2020

work page 2020
[6]

Multiuser resource control with deep reinforcement learning in iot edge computing

Lei Lei, Huijuan Xu, Xiong Xiong, Kan Zheng, Wei Xiang, and Xianbin Wang. Multiuser resource control with deep reinforcement learning in iot edge computing. IEEE Internet of Things J., 6(6):10119–10133, 2019

work page 2019
[7]

Stacked autoencoder-based deep reinforcement learning for online resource scheduling in large-scale mec networks.IEEE Internet of Things J., 7(10):9278–9290, 2020

Feibo Jiang, Kezhi Wang, Li Dong, Cunhua Pan, and Kun Yang. Stacked autoencoder-based deep reinforcement learning for online resource scheduling in large-scale mec networks.IEEE Internet of Things J., 7(10):9278–9290, 2020

work page 2020
[8]

A survey of multi-objective sequential decision-making

Diederik M Roijers, Peter Vamplew, Shimon Whiteson, and Richard Dazeley. A survey of multi-objective sequential decision-making. Journal of Artificial Intelligence Research, 48:67– 113, 2013

work page 2013

Show all 39 references

[9]

A generalized algorithm for multi- objective reinforcement learning and policy adaptation

Runzhe Yang, Xingyuan Sun, and Karthik Narasimhan. A generalized algorithm for multi- objective reinforcement learning and policy adaptation. Advances in neural information pro- cessing systems, 32, 2019

2019
[10]

Offloading and resource allocation with gen- eral task graph in mobile edge computing: A deep reinforcement learning approach

Jia Yan, Suzhi Bi, and Ying Jun Angela Zhang. Offloading and resource allocation with gen- eral task graph in mobile edge computing: A deep reinforcement learning approach. IEEE Transactions on Wireless Communications, 19(8):5404–5419, 2020

2020
[11]

Gasto: A fast adaptive graph learning framework for edge computing empowered task offloading

Yinong Li, Jianbo Li, Zhiqiang Lv, Haoran Li, Yue Wang, and Zhihao Xu. Gasto: A fast adaptive graph learning framework for edge computing empowered task offloading. IEEE Transactions on Network and Service Management, 2023

2023
[12]

Fast adaptive task offloading and resource allocation in large-scale mec systems via multi-agent graph reinforcement learning.IEEE Internet of Things Journal, 2023

Zhen Gao, Lei Yang, and Yu Dai. Fast adaptive task offloading and resource allocation in large-scale mec systems via multi-agent graph reinforcement learning.IEEE Internet of Things Journal, 2023

2023
[13]

Tao Ren, Jianwei Niu, and Yuan Qiu. Enhancing generalization of computation offloading policies in novel mobile edge computing environments by exploiting experience utility.Journal of Systems Architecture, 125:102444, 2022

2022
[14]

Deep reinforcement learning for online computation offloading in wireless powered mobile-edge computing networks.IEEE Transac- tions on Mobile Computing, 19(11):2581–2593, 2019

Liang Huang, Suzhi Bi, and Ying-Jun Angela Zhang. Deep reinforcement learning for online computation offloading in wireless powered mobile-edge computing networks.IEEE Transac- tions on Mobile Computing, 19(11):2581–2593, 2019

2019
[15]

Domain randomization for transferring deep neural networks from simulation to the real world

Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages 23–...

2017
[16]

A practical guide to multi-objective reinforcement learning and planning

Conor F Hayes, Roxana R ˘adulescu, Eugenio Bargiacchi, Johan K ¨allstr¨om, Matthew Macfar- lane, Mathieu Reymond, Timothy Verstraeten, Luisa M Zintgraf, Richard Dazeley, Fredrik Heintz, et al. A practical guide to multi-objective reinforcement learning and planning. Au- tonomo...

2022
[17]

A survey of generalisa- tion in deep reinforcement learning

Robert Kirk, Amy Zhang, Edward Grefenstette, and Tim Rockt ¨aschel. A survey of generalisa- tion in deep reinforcement learning. arXiv preprint arXiv:2111.09794, 2021

2021 arXiv
[18]

Why generalization in rl is difficult: Epistemic pomdps and implicit partial observability

Dibya Ghosh, Jad Rahme, Aviral Kumar, Amy Zhang, Ryan P Adams, and Sergey Levine. Why generalization in rl is difficult: Epistemic pomdps and implicit partial observability. Ad- vances in Neural Information Processing Systems, 34:25502–25515, 2021

2021
[19]

Deep reinforcement learning for collaborative offloading in heterogeneous edge networks

Dinh C Nguyen, Pubudu N Pathirana, Ming Ding, and Aruna Seneviratne. Deep reinforcement learning for collaborative offloading in heterogeneous edge networks. In2021 IEEE/ACM 21st International Symposium on Cluster, Cloud and Internet Computing (CCGrid), pages 297–303. IEEE, 2021

2021
[20]

Distributed resource schedul- ing for large-scale mec systems: A multiagent ensemble deep reinforcement learning with imitation acceleration

Feibo Jiang, Li Dong, Kezhi Wang, Kun Yang, and Cunhua Pan. Distributed resource schedul- ing for large-scale mec systems: A multiagent ensemble deep reinforcement learning with imitation acceleration. IEEE Internet of Things Journal, 9(9):6597–6610, 2021

2021
[21]

Fast adaptive task offloading in edge computing based on meta reinforcement learning

Jin Wang, Jia Hu, Geyong Min, Albert Y Zomaya, and Nektarios Georgalas. Fast adaptive task offloading in edge computing based on meta reinforcement learning. IEEE Transactions on Parallel and Distributed Systems, 32(1):242–253, 2020

2020
[22]

A scalable com- putation offloading scheme for mec based on graph neural networks

Tuan Wu, Wenpeng Jing, Xiangming Wen, Zhaoming Lu, and Shuyue Zhao. A scalable com- putation offloading scheme for mec based on graph neural networks. In 2021 IEEE Globecom Workshops (GC Wkshps), pages 1–6. IEEE, 2021

2021
[23]

Achieving fast environment adap- tation of drl-based computation offloading in mobile edge computing

Zheyuan Hu, Jianwei Niu, Tao Ren, and Mohsen Guizani. Achieving fast environment adap- tation of drl-based computation offloading in mobile edge computing. IEEE Transactions on Mobile Computing, 2023

2023
[24]

Multi-objective deep reinforcement learning for mobile edge computing

Ning Yang, Junrui Wen, Meng Zhang, and Ming Tang. Multi-objective deep reinforcement learning for mobile edge computing. In 2023 21st international symposium on modeling and optimization in mobile, ad hoc, and wireless networks (WiOpt), pages 1–8. IEEE, 2023

2023
[25]

Attention-based deep reinforcement learning for edge user allocation

Jiaxin Chang, Jian Wang, Bing Li, Yuqi Zhao, and Duantengchuan Li. Attention-based deep reinforcement learning for edge user allocation. IEEE Transactions on Network and Service Management, 2023

2023
[26]

Joint computation offloading and multiuser scheduling using approximate dynamic programming in nb-iot edge computing system

Lei Lei, Huijuan Xu, Xiong Xiong, Kan Zheng, and Wei Xiang. Joint computation offloading and multiuser scheduling using approximate dynamic programming in nb-iot edge computing system. IEEE Internet of Things J., 6(3):5345–5362, 2019

2019
[27]

K. Wang, F. Fang, Dbd Costa, and Z. Ding. Sub-channel scheduling, task assignment, and power allocation for oma-based and noma-based mec systems. IEEE Trans. Commun., PP(99):1–1, 2020

2020
[28]

Policy gradient methods for reinforcement learning with function approximation

Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. Advances in neural infor- mation processing systems, 12, 1999

1999
[29]

Soft actor-critic for discrete action settings

Petros Christodoulou. Soft actor-critic for discrete action settings. arXiv preprint arXiv:1910.07207, 2019

1910 arXiv
[30]

Policy gradient approaches for multi-objective sequential decision making

Simone Parisi, Matteo Pirotta, Nicola Smacchia, Luca Bascetta, and Marcello Restelli. Policy gradient approaches for multi-objective sequential decision making. In 2014 International Joint Conference on Neural Networks (IJCNN), pages 2323–2330. IEEE, 2014

2014
[31]

IEEE Std 802.11a-1999 , pages 1–102, 1999

Ieee standard for telecommunications and information exchange between systems - lan/man specific requirements - part 11: Wireless medium access control (mac) and physical layer (phy) specifications: High speed physical layer in the 5 ghz band. IEEE Std 802.11a-1999 , pages 1–102, 1999

1999
[32]

A contextual-bandit approach to personalized news article recommendation

Lihong Li, Wei Chu, John Langford, and Robert E Schapire. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international confer- ence on World wide web, pages 661–670, 2010

2010
[33]

Task replication for vehicular cloud: Contextual combinatorial bandit with delayed feedback

Lixing Chen and Jie Xu. Task replication for vehicular cloud: Contextual combinatorial bandit with delayed feedback. In IEEE INFOCOM 2019-IEEE Conference on Computer Communi- cations, pages 748–756. IEEE, 2019

2019
[34]

Collaboration-aware relay selec- tion for auv in internet of underwater network: Evolving contextual bandit learning approach

Haihong Zhao, Xinbin Li, Song Han, Lei Yan, and Junzhi Yu. Collaboration-aware relay selec- tion for auv in internet of underwater network: Evolving contextual bandit learning approach. IEEE Internet of Things Journal, 2022

2022
[35]

Computation rate maximization for wireless powered mobile- edge computing with binary computation offloading

Suzhi Bi and Ying Jun Zhang. Computation rate maximization for wireless powered mobile- edge computing with binary computation offloading. IEEE Transactions on Wireless Commu- nications, 17(6):4177–4190, 2018

2018
[36]

A fast and elitist multiobjective genetic algorithm: Nsga-ii

Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation , 6(2):182–197, 2002

2002
[37]

A comprehensive survey on nsga-ii for multi-objective optimization and applications

Haiping Ma, Yajing Zhang, Shengyi Sun, Ting Liu, and Yu Shan. A comprehensive survey on nsga-ii for multi-objective optimization and applications. Artificial Intelligence Review , 56(12):15217–15270, 2023

2023
[38]

Multi-objective reinforcement learning using sets of pareto dominating policies

Kristof Van Moffaert and Ann Now ´e. Multi-objective reinforcement learning using sets of pareto dominating policies. The Journal of Machine Learning Research , 15(1):3483–3512, 2014

2014
[39]

X m∈M γm(ωT Tm(πt) + ωEEm(πt)) # −

Sriraam Natarajan and Prasad Tadepalli. Dynamic preferences in multi-criteria reinforcement learning. In Proceedings of the 22nd International Conference on Machine learning , pages 601–608, 2005. Appendix A Differences in Generalization Compared to Related Works Many studies ...

2005

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Farhan Pervez, Ajmery Sultana, Cungang Yang, and Lian Zhao. Energy and latency efficient joint communication and computation optimization in a multi-uav-assisted mec network.IEEE Transactions on Wireless Communications, 23(3):1728–1741, 2024

work page 2024

[2] [2]

Deep reinforcement learning based computa- tion offloading and resource allocation for mec

Ji Li, Hui Gao, Tiejun Lv, and Yueming Lu. Deep reinforcement learning based computa- tion offloading and resource allocation for mec. In 2018 IEEE Wireless Communications and Networking Conference (WCNC), pages 1–6. IEEE, 2018

work page 2018

[3] [3]

Optimal Task Assignment and Power Allocation for NOMA Mobile-Edge Computing Networks

Fang Fang, Yanqing Xu, Zhiguo Ding, Chao Shen, Mugen Peng, and George K Karagiannidis. Optimal task assignment and power allocation for noma mobile-edge computing networks. arXiv preprint arXiv:1904.12389, 2019

work page Pith review arXiv 1904

[4] [4]

Joint task offloading and resource allocation for multi-server mobile-edge computing networks

Tuyen X Tran and Dario Pompili. Joint task offloading and resource allocation for multi-server mobile-edge computing networks. IEEE Transactions on Vehicular Technology, 68(1):856– 868, 2018

work page 2018

[5] [5]

Latency and energy optimization for mec enhanced sat-iot networks

Gaofeng Cui, Xiaoyao Li, Lexi Xu, and Weidong Wang. Latency and energy optimization for mec enhanced sat-iot networks. IEEE Access, 8:55915–55926, 2020

work page 2020

[6] [6]

Multiuser resource control with deep reinforcement learning in iot edge computing

Lei Lei, Huijuan Xu, Xiong Xiong, Kan Zheng, Wei Xiang, and Xianbin Wang. Multiuser resource control with deep reinforcement learning in iot edge computing. IEEE Internet of Things J., 6(6):10119–10133, 2019

work page 2019

[7] [7]

Stacked autoencoder-based deep reinforcement learning for online resource scheduling in large-scale mec networks.IEEE Internet of Things J., 7(10):9278–9290, 2020

Feibo Jiang, Kezhi Wang, Li Dong, Cunhua Pan, and Kun Yang. Stacked autoencoder-based deep reinforcement learning for online resource scheduling in large-scale mec networks.IEEE Internet of Things J., 7(10):9278–9290, 2020

work page 2020

[8] [8]

A survey of multi-objective sequential decision-making

Diederik M Roijers, Peter Vamplew, Shimon Whiteson, and Richard Dazeley. A survey of multi-objective sequential decision-making. Journal of Artificial Intelligence Research, 48:67– 113, 2013

work page 2013