pith. sign in

arxiv: 2605.14649 · v1 · pith:VSWAVVWNnew · submitted 2026-05-14 · 💻 cs.DC

Multi-objective application placement in fog computing using graph neural network-based reinforcement learning

Isaac Lera , Carlos Guerrero This is my paper

Pith reviewed 2026-06-30 20:30 UTC · model grok-4.3

classification 💻 cs.DC
keywords fog computingapplication placementgraph neural networksreinforcement learningmulti-objective optimizationPareto frontservice dependenciesdeep reinforcement learning
0
0 comments X

The pith

A graph neural network with dual actor-critics enables real-time multi-objective application placement in fog computing while matching the quality of slower optimization methods.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a deep reinforcement learning model for placing applications in fog computing environments to optimize multiple objectives simultaneously. It uses a graph neural network to capture dependencies between services in an application, giving priority to more interconnected ones. Two actor-critic components handle the learning, allowing the model to make placement decisions quickly after training. Experiments show that this approach produces Pareto-optimal solutions comparable to genetic algorithms or integer programming but executes in milliseconds instead of hours.

Core claim

The framework employs a graph neural network combined with two actor-critics in a deep reinforcement learning setup to model service relationships and generate placement decisions that balance multiple objectives, achieving execution times in milliseconds while producing Pareto sets similar to those from traditional methods.

What carries the argument

Graph neural network integrated with two actor-critics in a deep reinforcement learning model that incorporates service interdependencies to prioritize placement decisions.

Load-bearing premise

The assumption that a model trained on specific instances can effectively handle new but similar placement problems in real time without further training or significant performance loss.

What would settle it

Running the trained model on a fog network topology or service dependency graph that differs substantially from the training set and checking whether solution quality falls below that of genetic algorithms or integer programming.

read the original abstract

We propose a framework designed to tackle a multi-objective optimization challenge related to the placement of applications in fog computing, employing a deep reinforcement learning (DRL) approach. Unlike other optimization techniques, such as integer linear programming or genetic algorithms, DRL models are applied in real time to solve similar problem situations after training. Our model comprises a learning process featuring a graph neural network and two actor-critics, providing a holistic perspective on the priorities concerning interconnected services that constitute an application. The learning model incorporates the relationships between services as a crucial factor in placement decisions: Services with higher dependencies take precedence in location selection. Our experimental investigation involves illustrative cases where we compare our results with baseline strategies and genetic algorithms. We observed a comparable Pareto set with negligible execution times, measured in the order of milliseconds, in contrast to the hours required by alternative approaches.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes a GNN-based DRL framework with two actor-critics for multi-objective application placement in fog computing. It models service dependencies to prioritize placement decisions and claims that, after training, the policy solves similar instances in real time, producing Pareto fronts comparable to ILP or GA but with millisecond inference times versus hours, as demonstrated on illustrative cases.

Significance. If the generalization and performance claims hold, the work could enable practical real-time multi-objective optimization for dynamic fog environments, where traditional solvers scale poorly. The explicit use of graph structure to capture service interdependencies is a technical strength that aligns with the problem's natural representation.

major comments (2)
  1. [Abstract] Abstract: the central claim that the approach yields 'a comparable Pareto set with negligible execution times, measured in the order of milliseconds, in contrast to the hours required by alternative approaches' supplies no experimental details, baseline definitions, statistical tests, or validation protocol. This renders the performance advantage uninspectable and is load-bearing for the main contribution.
  2. [Abstract] Abstract (paragraph on the learning process): the assertion that 'DRL models are applied in real time to solve similar problem situations after training' rests on the unverified premise that a single trained GNN-RL policy generalizes across varying service graphs, fog topologies, and objective weightings. No train/test splits, out-of-distribution evaluation, or scaling analysis with graph size are provided, directly undermining the real-time applicability claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments highlighting areas where the abstract could better support the claims. We address each major comment below with clarifications drawn directly from the manuscript's experimental investigation on illustrative cases.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the approach yields 'a comparable Pareto set with negligible execution times, measured in the order of milliseconds, in contrast to the hours required by alternative approaches' supplies no experimental details, baseline definitions, statistical tests, or validation protocol. This renders the performance advantage uninspectable and is load-bearing for the main contribution.

    Authors: The abstract summarizes results from the experimental investigation on illustrative cases comparing against baseline strategies and genetic algorithms. The full manuscript details these comparisons and the observed time differences. We agree the abstract would benefit from added context on the evaluation protocol to make the claims more inspectable. We will revise the abstract to briefly reference the illustrative cases used for validation. revision: yes

  2. Referee: [Abstract] Abstract (paragraph on the learning process): the assertion that 'DRL models are applied in real time to solve similar problem situations after training' rests on the unverified premise that a single trained GNN-RL policy generalizes across varying service graphs, fog topologies, and objective weightings. No train/test splits, out-of-distribution evaluation, or scaling analysis with graph size are provided, directly undermining the real-time applicability claim.

    Authors: The manuscript evaluates the trained policy on multiple illustrative cases with varying service graphs and fog topologies, demonstrating real-time application to similar problem instances. While the current version does not include explicit train/test splits or formal out-of-distribution analysis, the results on these cases support the stated applicability. We will revise the abstract to clarify the demonstrated scope of generalization without overstating the evaluation. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents an empirical DRL framework using GNN and actor-critics for multi-objective fog placement optimization. Its central claim of comparable Pareto fronts at millisecond inference (versus hours for ILP/GA) rests on post-training experimental comparisons on illustrative cases, not on any mathematical derivation, fitted parameter renamed as prediction, or self-citation chain. No equations, uniqueness theorems, or ansatzes are referenced that reduce the reported performance to inputs by construction. Standard RL training-then-inference workflow is described without self-referential reduction, making the result self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the work relies on standard DRL training assumptions that are not enumerated.

pith-pipeline@v0.9.1-grok · 5668 in / 1109 out tokens · 35436 ms · 2026-06-30T20:30:24.896720+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 28 canonical work pages · 1 internal anchor

  1. [1]

    Computer49(8), 112–116 (2016) https://doi.org/10.1109/MC

    Dastjerdi, A.V., Buyya, R.: Fog Computing: Helping the Internet of Things Real- ize Its Potential. Computer49(8), 112–116 (2016) https://doi.org/10.1109/MC. 2016.245

    work page doi:10.1109/mc 2016

  2. [2]

    Coffman, E.G., Garey, M.R., Johnson, D.S.: Approximation algorithms for bin packing: a survey, pp. 46–93. PWS Publishing Co., USA (1996)

    1996

  3. [3]

    IEEE Transactions on Network and Service Management 17(2), 1026–1039 (2020) https://doi.org/10.1109/TNSM.2019.2963643

    Sami, H., Mourad, A.: Dynamic on-demand fog formation offering on-the-fly IoT service deployment. IEEE Transactions on Network and Service Management 17(2), 1026–1039 (2020) https://doi.org/10.1109/TNSM.2019.2963643

    work page doi:10.1109/tnsm.2019.2963643 2020

  4. [4]

    Software: Practice and Experience50(5), 719–740 (2020) https://doi.org/10.1002/spe.2766

    Brogi, A., Forti, S., Guerrero, C., Lera, I.: How to Place Your Apps in the Fog: State of the Art and Open Challenges. Software: Practice and Experience50(5), 719–740 (2020) https://doi.org/10.1002/spe.2766

    work page doi:10.1002/spe.2766 2020

  5. [5]

    ACM Computing Surveys53(3) (2020) https://doi

    Salaht, F.A., Desprez, F., Lebre, A.: An Overview of Service Placement Problem in Fog and Edge Computing. ACM Computing Surveys53(3) (2020) https://doi. org/10.1145/3391196

    work page doi:10.1145/3391196 2020

  6. [6]

    Multimedia Tools and Appli- cations83(8), 23019–23045 (2024) https://doi.org/10.1007/s11042-023-16399-2

    Fahimullah, M., Ahvar, S., Agarwal, M., Trocan, M.: Machine learning-based solutions for resource management in fog computing. Multimedia Tools and Appli- cations83(8), 23019–23045 (2024) https://doi.org/10.1007/s11042-023-16399-2

    work page doi:10.1007/s11042-023-16399-2 2024

  7. [7]

    Li, C., Han, S., Zeng, S., Yang, S.: Multi-objective Optimization, pp. 181–202. Springer, Singapore (2024). https://doi.org/10.1007/978-981-97-3286-9 9

    work page doi:10.1007/978-981-97-3286-9 2024

  8. [8]

    Cogent Engineering5(1), 1502242 (2018) https://doi.org/10.1080/ 20 23311916.2018.1502242

    Gunantara, N.: A review of multi-objective optimization: Methods and its applications. Cogent Engineering5(1), 1502242 (2018) https://doi.org/10.1080/ 20 23311916.2018.1502242

    work page arXiv 2018

  9. [9]

    Springer, Singapore (2020)

    Dong, H., Ding, Z., Zhang, S.: Deep Reinforcement Learning Fundamentals, Research and Applications: Fundamentals, Research and Applications. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-4095-0

    work page doi:10.1007/978-981-15-4095-0 2020

  10. [10]

    In: McIlraith, S.A., Weinberger, K.Q

    Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., Meger, D.: Deep Reinforcement Learning That Matters. In: McIlraith, S.A., Weinberger, K.Q. (eds.) AAAI Press, pp. 3207–3214 (2018)

    2018

  11. [11]

    Cluster Computing (2024) https://doi.org/10.1007/s10586-024-04518-z

    Allaoui, T., Gasmi, K., Ezzedine, T.: Reinforcement learning based task offloading of IoT applications in fog computing: algorithms and optimization techniques. Cluster Computing (2024) https://doi.org/10.1007/s10586-024-04518-z

    work page doi:10.1007/s10586-024-04518-z 2024

  12. [12]

    IEEE Transactions on Mobile Computing (2021) https://doi.org/10

    Goudarzi, M., Palaniswami, M., Buyya, R.: A Distributed Deep Reinforcement Learning Technique for Application Placement in Edge and Fog Computing Envi- ronments. IEEE Transactions on Mobile Computing (2021) https://doi.org/10. 1109/TMC.2021.3123165

    work page arXiv 2021

  13. [13]

    IEEE Transactions on Parallel and Distributed Systems32, 242–253 (2020) https:// doi.org/10.1109/TPDS.2020.3014896

    Wang, J., Hu, J., Min, G., Zomaya, A.Y., Georgalas, N.: Fast Adaptive Task Offloading in Edge Computing Based on Meta Reinforcement Learning. IEEE Transactions on Parallel and Distributed Systems32, 242–253 (2020) https:// doi.org/10.1109/TPDS.2020.3014896

    work page doi:10.1109/tpds.2020.3014896 2020

  14. [14]

    https://openreview.net/forum?id=ryGs6iA5Km

    Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How Powerful are Graph Neural Networks? In: International Conference on Learning Representations (2019). https://openreview.net/forum?id=ryGs6iA5Km

    2019

  15. [15]

    Proximal Policy Optimization Algorithms

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal Policy Optimization Algorithms. ArXivabs/1707.06347(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  16. [16]

    IEEE Transactions on Systems, Man, and Cybernetics: Systems 45(3), 385–398 (2015) https://doi.org/10.1109/TSMC.2014.2358639

    Liu, C., Xu, X., Hu, D.: Multiobjective Reinforcement Learning: A Comprehen- sive Overview. IEEE Transactions on Systems, Man, and Cybernetics: Systems 45(3), 385–398 (2015) https://doi.org/10.1109/TSMC.2014.2358639

    work page doi:10.1109/tsmc.2014.2358639 2015

  17. [17]

    Addison Wesley series in artificial intelligence

    Genetic Algorithms in Search, Optimization, and Machine Learning. Addison Wesley series in artificial intelligence. Addison-Wesley (1989). https://books. google.es/books?id=2IIJAAAACAAJ

    1989

  18. [18]

    Artificial Intelligence Review57(5), 124 (2024) https://doi.org/10

    Zhou, G., Tian, W., Buyya, R., Xue, R., Song, L.: Deep reinforcement learning- based methods for resource scheduling in cloud computing: a review and future directions. Artificial Intelligence Review57(5), 124 (2024) https://doi.org/10. 1007/s10462-024-10756-9

    2024

  19. [19]

    Internet of Things21, 100674 (2023) https://doi.org/10.1016/j.iot.2022.100674

    Iftikhar, S., Gill, S.S., Song, C., Xu, M., Aslanpour, M.S., Toosi, A.N., Du, J., Wu, H., Ghosh, S., Chowdhury, D., Golec, M., Kumar, M., Abdelmoniem, A.M., Cuadrado, F., Varghese, B., Rana, O., Dustdar, S., Uhlig, S.: AI-based fog and 21 edge computing: A systematic review, taxonomy and future directions. Internet of Things21, 100674 (2023) https://doi.o...

    work page doi:10.1016/j.iot.2022.100674 2023

  20. [20]

    Journal of Supercomputing76, 388–410 (2020) https://doi.org/10.1007/s11227-019-03032-z

    Farhat, P., Sami, H., Mourad, A.: Reinforcement R-learning model for time scheduling of on-demand fog placement. Journal of Supercomputing76, 388–410 (2020) https://doi.org/10.1007/s11227-019-03032-z

    work page doi:10.1007/s11227-019-03032-z 2020

  21. [21]

    IEEE Access7, 128014–128025 (2019) https://doi.org/10.1109/ACCESS.2019.2939735

    Nassar, A., Yilmaz, Y.: Reinforcement Learning for Adaptive Resource Allocation in Fog RAN for IoT With Heterogeneous Latency Requirements. IEEE Access7, 128014–128025 (2019) https://doi.org/10.1109/ACCESS.2019.2939735

    work page doi:10.1109/access.2019.2939735 2019

  22. [22]

    Journal of Cloud Computing11(1) (2022) https://doi.org/10.1186/s13677-021-00276-0

    Zheng, T., Wan, J., Zhang, J., Jiang, C.: Deep Reinforcement Learning-Based Workload Scheduling for Edge Computing. Journal of Cloud Computing11(1) (2022) https://doi.org/10.1186/s13677-021-00276-0

    work page doi:10.1186/s13677-021-00276-0 2022

  23. [23]

    In: 2019 IEEE 8th International Confer- ence on Cloud Networking (CloudNet), pp

    Mseddi, A., Jaafar, W., Elbiaze, H., Ajib, W.: Intelligent Resource Allocation in Dynamic Fog Computing Environments. In: 2019 IEEE 8th International Confer- ence on Cloud Networking (CloudNet), pp. 1–7 (2019). https://doi.org/10.1109/ CloudNet47604.2019.9064110

    work page arXiv 2019

  24. [24]

    Wireless Communications and Mobile Computing2020, 1–16 (2020) https://doi.org/10.1155/2020/8863865

    Li, X., Qin, Y., Zhou, H., Chen, D., Yang, S., Zhang, Z.: An intelligent adaptive algorithm for servers balancing and tasks scheduling over mobile fog computing networks. Wireless Communications and Mobile Computing2020, 1–16 (2020) https://doi.org/10.1155/2020/8863865

    work page doi:10.1155/2020/8863865 2020

  25. [25]

    ACM Transactions on Internet Technology19(2) (2019) https://doi.org/10.1145/3234463

    Li, H., Ota, K., Dong, M.: Deep Reinforcement Scheduling for Mobile Crowdsens- ing in Fog Computing. ACM Transactions on Internet Technology19(2) (2019) https://doi.org/10.1145/3234463

    work page doi:10.1145/3234463 2019

  26. [26]

    In: IM, pp

    Poltronieri, F., Tortonesi, M., Stefanelli, C., Suri, N.: Reinforcement learning for value-based placement of fog services. In: IM, pp. 466–472 (2021)

    2021

  27. [27]

    IEEE Transactions on Emerging Topics in Computing10(4), 1810–1820 (2022) https://doi.org/10.1109/TETC.2021.3115793

    Zhou, X., Liu, Z., Guo, M., Zhao, J., Wang, J.: SACC: A Size Adaptive Content Caching Algorithm in Fog/Edge Computing Using Deep Reinforcement Learning. IEEE Transactions on Emerging Topics in Computing10(4), 1810–1820 (2022) https://doi.org/10.1109/TETC.2021.3115793

    work page doi:10.1109/tetc.2021.3115793 2022

  28. [28]

    IEEE Transactions on Intelligent Transporta- tion Systems, 1–14 (2022) https://doi.org/10.1109/TITS.2022.3169421

    Gao, H., Huang, W., Liu, T., Yin, Y., Li, Y.: PPO2: Location Privacy-Oriented Task Offloading to Edge Computing Using Reinforcement Learning for Intelligent Autonomous Transport Systems. IEEE Transactions on Intelligent Transporta- tion Systems, 1–14 (2022) https://doi.org/10.1109/TITS.2022.3169421

    work page doi:10.1109/tits.2022.3169421 2022

  29. [29]

    Journal of Grid Computing 22(1), 18 (2024) https://doi.org/10.1007/s10723-023-09729-z

    Zhang, Z., Gu, K., Xu, Z.: DRL-based Task and Computational Offloading for Internet of Vehicles in Decentralized Computing. Journal of Grid Computing 22(1), 18 (2024) https://doi.org/10.1007/s10723-023-09729-z

    work page doi:10.1007/s10723-023-09729-z 2024

  30. [30]

    In: 2021 IEEE 12th International Confer- ence on Software Engineering and Service Science (ICSESS), pp

    Bai, W., Qian, C.: Deep Reinforcement Learning for Joint Offloading and 22 Resource Allocation in Fog Computing. In: 2021 IEEE 12th International Confer- ence on Software Engineering and Service Science (ICSESS), pp. 131–134 (2021). https://doi.org/10.1109/ICSESS52187.2021.9522334

    work page doi:10.1109/icsess52187.2021.9522334 2021

  31. [31]

    IEEE Internet of Things Journal6(2), 3641–3651 (2019) https://doi.org/10.1109/JIOT.2018.2889511

    Lera, I., Guerrero, C., Juiz, C.: Availability-aware service placement policy in fog computing based on graph partitions. IEEE Internet of Things Journal6(2), 3641–3651 (2019) https://doi.org/10.1109/JIOT.2018.2889511

    work page doi:10.1109/jiot.2018.2889511 2019

  32. [32]

    INFORMS Journal on Computing3, 149–156 (1991)

    Applegate, D.L., Cook, W.J.: A Computational Study of the Job-Shop Scheduling Problem. INFORMS Journal on Computing3, 149–156 (1991)

    1991

  33. [33]

    IEEE Transactions on Cybernetics51(6), 3103–3114 (2021) https: //doi.org/10.1109/TCYB.2020.2977661

    Li, K., Zhang, T., Wang, R.: Deep Reinforcement Learning for Multiobjective Optimization. IEEE Transactions on Cybernetics51(6), 3103–3114 (2021) https: //doi.org/10.1109/TCYB.2020.2977661

    work page doi:10.1109/tcyb.2020.2977661 2021

  34. [34]

    Curran Associates Inc., Red Hook, NY, USA (2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., K¨ opf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: an imperative style, high-performance deep learning library. Curran Associates Inc., Red H...

    2019

  35. [35]

    IEEE Transactions on Evolutionary Computation6(2), 182–197 (2002).https://doi.org/10.1109/4235.996017

    Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 6(2), 182–197 (2002) https://doi.org/10.1109/4235.996017

    work page doi:10.1109/4235.996017 2002

  36. [36]

    IEEE Access 8, 89497–89509 (2020)

    Blank, J., Deb, K.: pymoo: Multi-Objective Optimization in Python. IEEE Access 8, 89497–89509 (2020)

    2020

  37. [37]

    In: Proceedings of the IEEE/ACM 16th International Conference on Utility and Cloud Computing

    Viv´ o, S., Lera, I., Guerrero, C.: Comparing Evolutionary Optimization Algo- rithms for the Fog Service Placement Problem. In: Proceedings of the IEEE/ACM 16th International Conference on Utility and Cloud Computing. UCC ’23. Asso- ciation for Computing Machinery, New York, NY, USA (2024). https://doi.org/ 10.1145/3603166.3632547 23

    work page doi:10.1145/3603166.3632547 2024