A Novel Deep Reinforcement Learning Method for Computation Offloading in Multi-User Mobile Edge Computing with Decentralization

Nguyen Chi Long; Nguyen Ngoc Hai Dang; Ta Hai Tung; Trinh Van Chien; Trong-Minh Hoang; Van Son Nguyen

REVIEW 5 major objections 5 minor 17 references

A Novel Deep Reinforcement Learning Method for Computation Offloading in Multi-User Mobile Edge Computing with Decentralization

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Twin Delayed DDPG outperforms DDPG as a decentralized offloading policy in multi-user MEC.

desk verdict A routine TD3-vs-DDPG benchmark in a borrowed MEC model, with a plausible reward claim that the reported single unseeded run does not actually support. read the letter →

arxiv 2506.02458 v1 pith:WOPO6J2L submitted 2025-06-03 cs.IT cs.LGmath.IT

Nguyen Chi Long , Trinh Van Chien , Ta Hai Tung , Van Son Nguyen , Trong-Minh Hoang , Nguyen Ngoc Hai Dang This is my paper

classification cs.ITcs.LGmath.IT

keywords mobileedgecomputingcomputationoffloadingdeepreinforcementlearningTwinDelayedDDPGDeterministicPolicyGradientdecentralizedcontrolpowerallocationportableusers

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 argues that a specific deep reinforcement learning algorithm, Twin Delayed DDPG (TD3), can replace Deep Deterministic Policy Gradient (DDPG) as the per-user learning engine in a decentralized mobile edge computing (MEC) system. In the setup, each mobile user independently decides how much power to spend on local computation and how much to spend on offloading tasks to a base station-attached MEC server, using only local observations such as buffer queue length, channel vector, and a power-ratio estimate. The paper simulates three portable users whose distances to the base station change randomly, and reports that TD3-learned policies achieve higher average reward for all three users compared with DDPG, with the overall performance read as superior to the DDPG-based strategy. If the comparison holds, the result matters because it suggests that standard improvements to actor-critic reinforcement learning transfer to decentralized offloading, making MEC systems more scalable and energy-efficient without requiring global state sharing.

What carries the argument

The central object is the Twin Delayed DDPG (TD3) algorithm, an actor-critic deep reinforcement learning method that maintains two Q-value networks and uses the smaller of their target values when computing Bellman errors, updates the policy less frequently than the critics, and adds clipped noise to target actions. It operates within a per-user reward $r_{m,t} = -w_{m,1}(p_{l,m}(t)+p_{o,m}(t)) - w_{m,2}B_m(t)$ and a state $s_{m,t} = [B_m(t), \phi_m(t-1), h_m(t)]$, where $B_m(t)$ is the task buffer queue length, $\phi_m(t-1)$ is the estimated power ratio after zero-forcing detection, and $h_m(t)$ is the user's channel vector. This machinery carries the argument by giving each user a more stable critic target, mitigating the overestimation that the paper identifies as DDPG's weakness, and thereby producing the reported reward gains in the portable-user simulations.

What would settle it

Run the exact Table I setup (three users, $w_m = 0.8$, 2000 episodes, 200 steps per episode) with ten or more independent random seeds for both algorithms, and compare distributions of per-user average reward, power, and delay. If TD3's per-user reward advantage does not persist with non-overlapping confidence intervals, or if average power is not lower for the majority of users, the paper's superiority claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that in the decentralized multi-user MEC framework introduced in [7], replacing the DDPG agent with TD3 yields better learned computation offloading policies, especially when mobile users are portable. Concretely, the reported per-user average rewards improve from -9.52 to -3.09, from -14.08 to -3.71, and from -19.49 to -3.89 while users move; average power decreases for two of the three users and average delay decreases for two of the three users. The paper attributes the improvement to TD3's remedies for DDPG's Q-value overestimation, namely clipped double-Q learning and delayed policy updates, which stabilize training and lead to higher-reward policies.

Load-bearing premise

The paper's conclusion that TD3 is better rests on a single set of simulation runs with one reward weighting, one movement pattern, and unstated random seeds; if the reported gaps shrink or flip under different seeds or hyperparameters, the superiority claim collapses.

Editorial extensions

If this is right

Each portable user can learn an offloading policy independently from its own local state, without knowing the total number of users or the statistics of task arrivals and channels.
Replacing DDPG with TD3 improves the average reward per episode for all three simulated users, and the paper interprets the overall results as showing that TD3 exceeds the DDPG-based power control strategy.
The learned policies operate with finite feedback: each user observes only its buffer queue, the estimated power ratio after zero-forcing detection, and its own channel vector.
Because TD3 uses two Q-networks and delayed policy updates, its runtime stays close to DDPG's while reducing Q-value overestimation, making it a practical drop-in replacement in this simulator.

Reading between the lines

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

If the reported advantage is robust to random seeds and hyperparameter choices, the same TD3 recipe (clipped double-Q targets and delayed policy updates) could be applied to other decentralized resource-allocation problems in MEC, such as channel assignment or split inference, without changing the local-state architecture.
The reward weight $w_m$ is fixed at 0.8 for all users; varying it would trace the energy-delay Pareto frontier, and TD3's edge over DDPG may depend on the operating point, which the paper does not explore.
A natural extension is to test the same decentralized learning rule with more than three users, since the scalability argument in the paper is not yet backed by simulations with larger $M$.

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

The one-sentence take: this is a routine empirical comparison of TD3 against DDPG in the MEC setup of [7], with the only new elements being a random-walk mobility model and a claim that TD3 wins. That claim may be true, but the evidence as reported is not enough to establish it.

What the paper does well: the system model and the two algorithms are described clearly, and the match to [7] is explicit, so a reader can reconstruct the setup without going back to the original. The result that TD3 gives a better average reward for all three users is at least directionally consistent with what TD3 is designed to do. The paper also honestly says it is building on prior work; it does not oversell the algorithmic novelty.

The soft spots are real and load-bearing. Section IV-B states that TD3 achieved better average power and better average delay, but Table I contradicts that for User 3 (power 0.30 vs 0.20, worse for TD3) and for User 1 (delay 4.67 vs 1.95, worse for TD3). If the intended claim is aggregate-over-users, the paper never says so. More importantly, the results are from a single unseeded run with one hand-chosen reward weight (w_m=0.8). No standard deviations, no seeds, no confidence intervals. Hyperparameters are partially taken from [7] but the learning rates, minibatch size, target update delay, and layer sizes are not given. The 'average delay' metric in Table I is never defined; the reward uses queue length, not delay. With one run, the observed reward gap (about -3.6 vs -14.4) cannot be separated from run-to-run variance or from the particular weight choice.

The novelty is also thin: TD3 is a standard algorithm, the system model is from [7], and the simulation has only three users. So this is not a paper that opens a new direction. It is a small empirical data point.

Who is this for? A reader who wants a quick, informal check of whether TD3 might work in this decentralized offloading setting. That reader should treat the comparison as preliminary.

My recommendation: engage with it, but require revision. A serious referee should ask for multiple seeds, error bars, a defined delay metric, full hyperparameter disclosure, a corrected Table I or corrected text, and ideally code. With those fixes the paper could be a modest but honest contribution. Send it to peer review; it deserves referee time to force the experimental rigor.

Referee Report

5 major / 5 minor

Summary. The paper applies Twin Delayed DDPG (TD3) to the decentralized multi-user mobile edge computing (MEC) problem studied by [7], with the added feature that users are portable and their distances to the base station change over time. Each user is an independent DRL agent with state comprising queue length, estimated zero-forcing power ratio, and channel vector; the action is the continuous local-execution and offloading power pair; and the reward is a negatively weighted sum of power consumption and queue length. The paper reports a simulation with three users over 2000 episodes, comparing TD3 against DDPG and claiming that TD3 achieves better average reward, average power, and average delay. The contribution is entirely empirical: no theoretical analysis is offered, and the central claim rests on the reported experimental comparison.

Significance. If the reported advantage were actually established, the paper would offer a simple, practical improvement over DDPG for decentralized MEC with moving users, and the system model in Section II is coherent and closely follows a known benchmark. Credit is due for providing explicit pseudocode for both DDPG and TD3 and for specifying the simulation setup in some detail. However, the significance is severely limited by the fact that the only evidence is a single unseeded run, the numerical table contradicts the qualitative text, and key metrics and hyperparameters are left undefined. As it stands, the central claim is not supported to the standard expected for an empirical comparison.

major comments (5)

[Section IV-A and IV-B] The paper's central claim that TD3 'has much better results than DDPG' (Abstract) and 'exceeded the conventional DDPG-based power control strategy' is based on a single unseeded run per algorithm. There are no standard deviations, random seeds, confidence intervals, or statistical tests. With only three users and one realization of the random channel, task, and mobility processes, the observed reward gap (mean about -3.6 versus -14.4) cannot be separated from run-to-run variance. The authors should report multiple independent seeds with per-user and aggregate means and standard deviations, and ideally a paired statistical test, or temper the superiority claim accordingly.
[Table I and Section IV-B] Table I contradicts the qualitative statements in Section IV-B. For User 3, the average power for TD3 is 0.30 versus 0.20 for DDPG, meaning DDPG is better on this metric, yet the text states that 'regarding average power per episode, TD3 also achieved better results compared with DDPG.' Similarly, for User 1, the average delay for TD3 is 4.67 versus 1.95 for DDPG, while the surrounding text implies better delay performance. If the intended claim is only about the average over users or only about some users, that is not stated. The table and the qualitative summary must be reconciled.
[Table I and Section IV-A] Table I reports 'Average delay' but no definition of this metric is given in Section IV-A or elsewhere. The reward function (17) uses the queue length B_m(t), not delay; Section III-A mentions Little's theorem, but the exact empirical quantity reported in the table (mean queue length over time, steady-state queue length, or an actual delay) is never specified. Without this definition, the delay column is uninterpretable and cannot be used to support the superiority claim.
[Algorithm 2, line 14] The TD3 pseudocode computes the target next action as clip(μ(s'|θ^μ) + clip(ε,-c,c), a_low, a_high), using the current actor parameters θ^μ. In standard TD3, the target action must be computed with the target actor parameters θ^{μ'}, which are defined in line 1 but are never used in line 14. If the implementation follows this pseudocode, the algorithm is not TD3, and the reported improvement over DDPG could be an artifact of a different exploration and update scheme. This must be corrected, and if the experiments were run with the incorrect version, they should be rerun.
[Section IV-A] The simulation setup is incomplete with respect to the DRL hyperparameters. The text says the networks use 'almost the same settings as in [7]' and cites [17] for initialization, but it omits the learning rates α_Q and α_μ, the minibatch size, the soft-update rate τ, the policy-delay parameter update_every, the number of gradient updates per step, and the layer sizes, widths, and activations. The mobility model is also only sketched: distances are 'added by a value drawn randomly from the standard normal distribution, such that the absolute value of the cumulative sum of these distances is lower than 10.' These omissions prevent reproduction and make it impossible to judge whether the result is robust to hyperparameter choices, which is critical because the whole claim rests on one comparison.

minor comments (5)

[Equation (1)] The summation index is j, but the channel, power, and data symbol inside the sum are all indexed by m; in addition, the noise term n(t) appears inside the summation. This should be h_j(t), p_{o,j}(t), and s_j(t), with n(t) outside the sum.
[Section II-A] The text 'variances 2 = 1' is garbled; it should state, for example, that E[|s_m(t)|^2] = 1.
[Section III] The symbol γ is used both for SINR in (4) and for the discount factor in (15) and in the algorithms. These two uses should be distinguished to avoid confusion.
[Algorithms 1 and 2] The symbol B is used both for the task buffer queue length in (5) and for the experience replay buffer in Algorithms 1 and 2. The notation should be changed.
[Section IV-B and Figures 3-5] The figures are described in terms of orange and black lines, but no legend or axis-label description is provided in the text; including legends and defining what is averaged (over episodes or over users) would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the TD3-vs-DDPG comparison is an independent empirical benchmark, not a derivation that presupposes its conclusion.

full rationale

The paper's central claim is an empirical superiority of TD3 over DDPG in a specific MEC simulator. The MEC model, state/action/reward definitions, and DRL algorithms are either taken from prior work (Chen-Wang [7] and Lillicrap et al. [17]) or stated directly in the paper, and TD3 is a standard, externally described algorithm. No parameter is fitted to a subset of data and then reported as a prediction of a closely related quantity: the reward weights (w_m = 0.8) are chosen once for both algorithms, and the comparison uses the same environment, episode count, and initialization. The single-seed/no-error-bar issue is an empirical robustness concern, not circularity. There are no load-bearing self-citations: the authors cite [7] for the framework and compare against it, but the DDPG baseline and environment are reproducible from the cited description, and the TD3 implementation is standard. No uniqueness theorem or prior result by the same authors is invoked to forbid alternatives. The observed performance gap is an empirical outcome of the simulator, not an identity, a renamed known result, or a consequence of the definitions. Therefore the derivation chain is self-contained with respect to circularity.

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

The paper's central comparison rests on several hand-set simulation choices: reward weights w_m=0.8, random-walk mobility, exploration noise, replay buffer size, and training length. It also inherits domain assumptions from [7], including abundant edge resources, perfect local CSI, and negligible feedback delay. No new physical or algorithmic entities are introduced. The main uncertainty is whether the DDPG baseline was tuned to the same standard as TD3 and whether the results survive multiple seeds.

free parameters (6)

Reward balance factor w_m = 0.8
Chosen identically for all users in Section IV-A; it sets w_{m,1}=10^{0.8} and w_{m,2}=0.2, directly determining the objective and therefore which policy is called better.
Movement noise standard deviation = 1 (standard normal)
Each distance step is drawn from the standard normal distribution in Section IV-A; the portability scenario and all mobility results depend on this choice.
Ornstein-Uhlenbeck exploration noise parameters = theta=0.15, sigma=0.12
Fixed in Section IV-A for exploration; the comparison can depend on them, and no sensitivity analysis is reported.
Experience replay buffer size = 2.5e5
Set in Section IV-A; affects learning stability. No sensitivity analysis is provided.
Training episodes and episode length = Kmax=2000, Tmax=200
Stopping rule for training, set in Section IV-A; there is no convergence criterion and no multiple-seed averaging.
Unspecified DRL hyperparameters = not specified
Network sizes, learning rates, and TD3 policy delay are needed by Algorithm 2 but are not reported in Section IV-A; the paper says settings are 'almost same' as [7].

assumptions (5)

domain assumption The MEC server has abundant computation resources, so edge processing latency and feedback delay are negligible; all offloaded bits are served in the same slot.
Section II-B.2 makes this assumption so that the offloading rate equals the Shannon rate in Eq. (9). If edge latency or feedback delay were nonzero, the queue dynamics and delay results would change.
domain assumption Each user has perfect knowledge of its own channel vector h_m(t) and queue length B_m(t) at each slot without communication overhead.
The state in Eq. (10) includes h_m(t); the decentralized learning claim relies on local observation, but obtaining an exact channel vector normally requires estimation and feedback.
domain assumption Users move according to a random walk with standard-normal step sizes, truncated so that cumulative absolute displacement stays below 10 m.
Section IV-A describes this portability model; the central comparison between DDPG and TD3 is made under this specific mobility process.
domain assumption The task arrival processes are i.i.d. with means lambda_m, and the zero-forcing detector behaves as modeled in Eq. (4).
Sections II-A and II-B inherit these modeling choices from [7]; the simulation and conclusions are valid only within this model.
domain assumption The DDPG baseline from [7] is implemented faithfully and tuned fairly relative to TD3.
Section IV-A says the neural network settings are 'almost same' as [7], but the paper does not show that both algorithms received equal tuning. The fairness of the comparison is assumed, not demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Novel Deep Reinforcement Learning Method for Computation Offloading in Multi-User Mobile Edge Computing with Decentralization." pith.science (2026). https://pith.science/paper/WOPO6J2L

@misc{pith2026250602458,
  author       = {Pith},
  title        = {Pith review of: A Novel Deep Reinforcement Learning Method for Computation Offloading in Multi-User Mobile Edge Computing with Decentralization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WOPO6J2L}},
  note         = {Machine review of arXiv:2506.02458}
}

read the original abstract

Mobile edge computing (MEC) allows appliances to offload workloads to neighboring MEC servers that have the potential for computation-intensive tasks with limited computational capabilities. This paper studied how deep reinforcement learning (DRL) algorithms are used in an MEC system to find feasible decentralized dynamic computation offloading strategies, which leads to the construction of an extensible MEC system that operates effectively with finite feedback. Even though the Deep Deterministic Policy Gradient (DDPG) algorithm, subject to their knowledge of the MEC system, can be used to allocate powers of both computation offloading and local execution, to learn a computation offloading policy for each user independently, we realized that this solution still has some inherent weaknesses. Hence, we introduced a new approach for this problem based on the Twin Delayed DDPG algorithm, which enables us to overcome this proneness and investigate cases where mobile users are portable. Numerical results showed that individual users can autonomously learn adequate policies through the proposed approach. Besides, the performance of the suggested solution exceeded the conventional DDPG-based power control strategy.

Figures

Figures reproduced from arXiv: 2506.02458 by the authors.

**Figure 1.** MEC architecture that come with computational resources and accomplish computation tasks competently. So, researchers have generally and closely examined computation offloading strategies for MEC in the literature to get greater energy potency or computational performance. Research by [3] and [4] has demonstrated that reinforcement learning (RL) techniques can effectively optimize dynamic computational offloading p… view at source ↗

**Figure 2.** DDPG Algorithm functioning as the critic, estimates the Q-value network, represented as Q(s, a|θ Q). Detailed visualization of this algorithm is shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Average reward per episode [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Average power per episode CN 0, h0 d0 d α IN as the error vector. Moreover, we use a system bandwidth W of 1MHz, and we set the noise power σ 2 R = 10−9W and the maximum transmission power Po,m = 2W. Conversely, for local processing, we set the maximum local exe…

**Figure 5.** Figure 5: Average delay per episode TABLE I: RESULTS Average reward (↑) Average power (↓) Average delay (↓) DDPG TD3 DDPG TD3 DDPG TD3 User 1 -9.52 -3.09 0.70 0.21 1.95 4.67 User 2 -14.08 -3.71 0.27 0.24 59.47 8.91 User 3 -19.49 -3.89 0.20 0.30 89.30 5.29 and critic network. For…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

[7]

Decentralized computation offloading for multi-user mobile edge computing: a deep reinforcement learning approach.EURASIP J

Zhao Chen and Xiaodong Wang. Decentralized computation offloading for multi-user mobile edge computing: a deep reinforcement learning approach.EURASIP J. Wirel. Commun. Netw., 2020(1), sep 2020

work page 2020
[17]

Lillicrap, Jonathan J

Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Manfred Otto Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning.CoRR, abs/1509.02971, 2015

arXiv 2015
[1]

Toward a unified elastic computing platform for smartphones with cloud support.IEEE Network, 27(5):34–40, 2013

Weiwen Zhang, Yonggang Wen, Jun Wu, and Hui Li. Toward a unified elastic computing platform for smartphones with cloud support.IEEE Network, 27(5):34–40, 2013

work page 2013
[2]

The emergence of edge computing.Com- puter, 50(1):30–39, 2017

Mahadev Satyanarayanan. The emergence of edge computing.Com- puter, 50(1):30–39, 2017

work page 2017
[3]

Online learning for offloading and autoscaling in energy harvesting mobile edge computing, 2017

Jie Xu, Lixing Chen, and Shaolei Ren. Online learning for offloading and autoscaling in energy harvesting mobile edge computing, 2017

work page 2017
[4]

Thinh Quang Dinh, Quang Duy La, Tony Q. S. Quek, and Hyundong Shin. Learning for computation offloading in mobile edge computing. IEEE Transactions on Communications, 66(12):6353–6367, 2018

work page 2018
[5]

Richard S Sutton and Andrew G Barto.Reinforcement learning: An introduction. 2018

work page 2018
[6]

Rusu, Joel Veness, Marc G

V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Kirkeby Fidjeland, Georg Ostrovski, Stig Petersen, Charlie Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Ku- maran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep rein...

2015

Show all 17 references

[8]

Mobile edge computing: A survey on architecture and computation offloading.CoRR, abs/1702.05309, 2017

Pavel Mach and Zdenek Becvar. Mobile edge computing: A survey on architecture and computation offloading.CoRR, abs/1702.05309, 2017

2017 arXiv
[9]

Power- delay tradeoff in multi-user mobile-edge computing systems.2016 IEEE Global Communications Conference (GLOBECOM), pages 1–6, 2016

Yuyi Mao, Jun Zhang, Shenghui Song, and Khaled Ben Letaief. Power- delay tradeoff in multi-user mobile-edge computing systems.2016 IEEE Global Communications Conference (GLOBECOM), pages 1–6, 2016

2016
[10]

Zhao Chen, Lei Zhang, Yukui Pei, Chunxiao Jiang, and Liuguo Yin. Noma-based multi-user mobile edge computation offloading via coop- erative multi-agent deep reinforcement learning.IEEE Transactions on Cognitive Communications and Networking, 8(1):350–364, 2022

2022
[11]

Prasobh Sankar and Sundeep Prabhakar Chepuri

R.S. Prasobh Sankar and Sundeep Prabhakar Chepuri. Channel esti- mation in mimo systems with one-bit spatial sigma-delta adcs.IEEE Transactions on Signal Processing, 70:4681–4696, 2022

2022
[12]

Larsson, and Thomas L

Hien Quoc Ngo, Erik G. Larsson, and Thomas L. Marzetta. Energy and spectral efficiency of very large multiuser mimo systems.IEEE Transactions on Communications, 61(4):1436–1449, 2013

2013
[13]

Dream: Dynamic resource and task allocation for energy minimization in mobile cloud systems.IEEE Journal on Selected Areas in Communications, 33(12):2510–2523, 2015

Jeongho Kwak, Yeongjin Kim, Joohyun Lee, and Song Chong. Dream: Dynamic resource and task allocation for energy minimization in mobile cloud systems.IEEE Journal on Selected Areas in Communications, 33(12):2510–2523, 2015

2015
[14]

Burd and Robert W

Thomas D. Burd and Robert W. Brodersen. Processor design for portable systems.Journal of VLSI signal processing systems for signal, image and video technology, 13:203–221, 1996

1996
[15]

Miettinen and Jukka K

Antti P. Miettinen and Jukka K. Nurminen. Energy efficiency of mobile clients in cloud computing. InProceedings of the 2nd USENIX Conference on Hot Topics in Cloud Computing, HotCloud’10, page 4, USA, 2010. USENIX Association

2010
[16]

Shortle, James M

Donald Gross, John F. Shortle, James M. Thompson, and Carl M. Harris.Fundamentals of Queueing Theory. Wiley-Interscience, USA, 4th edition, 2008

2008

Pith tools

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

[1] [7]

Decentralized computation offloading for multi-user mobile edge computing: a deep reinforcement learning approach.EURASIP J

Zhao Chen and Xiaodong Wang. Decentralized computation offloading for multi-user mobile edge computing: a deep reinforcement learning approach.EURASIP J. Wirel. Commun. Netw., 2020(1), sep 2020

work page 2020

[2] [17]

Lillicrap, Jonathan J

Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Manfred Otto Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning.CoRR, abs/1509.02971, 2015

arXiv 2015

[3] [1]

Toward a unified elastic computing platform for smartphones with cloud support.IEEE Network, 27(5):34–40, 2013

Weiwen Zhang, Yonggang Wen, Jun Wu, and Hui Li. Toward a unified elastic computing platform for smartphones with cloud support.IEEE Network, 27(5):34–40, 2013

work page 2013

[4] [2]

The emergence of edge computing.Com- puter, 50(1):30–39, 2017

Mahadev Satyanarayanan. The emergence of edge computing.Com- puter, 50(1):30–39, 2017

work page 2017

[5] [3]

Online learning for offloading and autoscaling in energy harvesting mobile edge computing, 2017

Jie Xu, Lixing Chen, and Shaolei Ren. Online learning for offloading and autoscaling in energy harvesting mobile edge computing, 2017

work page 2017

[6] [4]

Thinh Quang Dinh, Quang Duy La, Tony Q. S. Quek, and Hyundong Shin. Learning for computation offloading in mobile edge computing. IEEE Transactions on Communications, 66(12):6353–6367, 2018

work page 2018

[7] [5]

Richard S Sutton and Andrew G Barto.Reinforcement learning: An introduction. 2018

work page 2018

[8] [6]

Rusu, Joel Veness, Marc G

V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Kirkeby Fidjeland, Georg Ostrovski, Stig Petersen, Charlie Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Ku- maran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep rein...

2015