Recurrent Adversarial Service Times

Bogdan Georgiev; C\'esar Ojeda; Christian Bauckhage; Jannis Schuecker; Kostadin Cvejosky; Rams\'es J. S\'anchez

arxiv: 1906.09808 · v1 · pith:MUCEEZFBnew · submitted 2019-06-24 · 📊 stat.ML · cs.LG

Recurrent Adversarial Service Times

C\'esar Ojeda , Kostadin Cvejosky , Rams\'es J. S\'anchez , Jannis Schuecker , Bogdan Georgiev , Christian Bauckhage This is my paper

Pith reviewed 2026-05-25 17:12 UTC · model grok-4.3

classification 📊 stat.ML cs.LG

keywords queuing theoryrecurrent neural networksgenerative adversarial networksservice time distributionarrival processdata-driven modelingempirical service systems

0 comments

The pith

A recurrent neural network paired with a recurrent GAN models queuing arrivals and service times directly from data.

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

The paper proposes a neural solution to queuing systems that avoids the usual parametric assumptions about arrival and service time distributions. A recurrent neural network captures the point process of customer arrivals while a recurrent generative adversarial network learns and samples from the empirical service time distribution. This combination is evaluated on real datasets from blockchain transactions, GitHub activity, Stackoverflow posts, and New York taxi rides. The approach lets service dynamics be learned end-to-end from observations rather than imposed by fixed families of distributions. If the method works, queuing models can adapt to complex, non-standard patterns observed in actual service systems.

Core claim

The combination of a recurrent neural network for the arrival process and a recurrent generative adversarial network for the service time distribution provides a data-driven solution to queuing problems that does not rely on parametric assumptions about inter-event time distributions.

What carries the argument

Recurrent generative adversarial network that learns and generates service time distributions, used together with a recurrent neural network for the arrival process.

If this is right

Queuing models can be fitted to any observed arrival and service time sequences without choosing a parametric family in advance.
The same architecture applies across internet service logs and physical mobility traces.
Queue performance predictions become possible by simulating forward from the learned arrival and service processes.
The method can incorporate long-range temporal dependencies in both arrivals and service times through the recurrent components.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The learned service time generator could be swapped into existing queue simulators to test staffing or capacity changes on historical patterns.
If the model is updated incrementally on streaming data, it could track non-stationary service behavior in live systems.
Extensions to multi-server or network queues would require composing multiple such recurrent GAN modules while preserving the arrival process model.

Load-bearing premise

The recurrent GAN must reliably capture and reproduce the full empirical service time distribution without mode collapse or overfitting to the training datasets.

What would settle it

Generate service times on a held-out portion of one of the datasets and check whether the statistical properties of the generated times, such as the distribution of inter-service intervals or resulting queue length statistics, match those measured directly from the held-out data.

Figures

Figures reproduced from arXiv: 1906.09808 by Bogdan Georgiev, C\'esar Ojeda, Christian Bauckhage, Jannis Schuecker, Kostadin Cvejosky, Rams\'es J. S\'anchez.

**Figure 2.** Figure 2: Mempool Specifics. Panel (a) depicts the raw Mempoo [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison between the synthetic simulated data d [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison between the empirical data distributi [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Comparison of distributions for the mempool datas [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Comparison of the probability distributions obta [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

read the original abstract

Service system dynamics occur at the interplay between customer behaviour and a service provider's response. This kind of dynamics can effectively be modeled within the framework of queuing theory where customers' arrivals are described by point process models. However, these approaches are limited by parametric assumptions as to, for example, inter-event time distributions. In this paper, we address these limitations and propose a novel, deep neural network solution to the queuing problem. Our solution combines a recurrent neural network that models the arrival process with a recurrent generative adversarial network which models the service time distribution. We evaluate our methodology on various empirical datasets ranging from internet services (Blockchain, GitHub, Stackoverflow) to mobility service systems (New York taxi cab).

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.

Desk Editor's Note private letter to a colleague

The paper pairs an RNN arrival model with a recurrent GAN for service times to sidestep parametric assumptions in queuing, but the abstract contains no numbers so we cannot tell if the idea actually works.

read the letter

The main contribution here is the specific pairing of a recurrent neural net for customer arrivals with a recurrent generative adversarial network to generate service times. This is positioned as a way around the usual parametric assumptions in queuing theory, like exponential or other standard distributions for inter-event times. It applies the idea to several real-world datasets: blockchain transactions, GitHub activity, StackOverflow, and New York taxi data. That choice of domains makes sense because those systems have complex, non-standard timing patterns that parametric models often struggle with. The approach itself is straightforward in concept: use the RNN to handle the point process of arrivals, then the recurrent GAN to learn and sample from the service time distribution. If it works, it could give more flexible simulations for operations and logistics. The big gap is the complete lack of quantitative evidence in the abstract. There are no error metrics, no baseline comparisons, no checks on whether the generated service times match the real distributions in ways that matter for queuing (like tail behavior or quantiles). Recurrent GANs are known to have trouble with mode collapse, and service times in these datasets are likely to be heavy-tailed or multimodal. If the model under-samples long service events, the simulated queues would look less congested than they really are, undercutting the whole point. The stress-test concern about the GAN failing to capture the full distribution seems to apply directly here, since nothing in the description shows how they avoided it. Without those checks, the non-parametric advantage stays theoretical. This kind of work is aimed at researchers bridging machine learning and stochastic modeling, or practitioners in service system design who want data-driven alternatives. A reader looking for new architectures might find the recurrent GAN idea worth exploring, but anyone needing validated performance would have to wait for the details. I'd recommend sending it for peer review if the full paper includes solid experiments and comparisons. The idea is clear enough to be worth referee input, even if revisions are needed.

Referee Report

2 major / 1 minor

Summary. The paper claims that limitations of parametric assumptions in queuing theory (e.g., inter-event time distributions) can be addressed by a deep neural network solution that combines a recurrent neural network to model the arrival process with a recurrent generative adversarial network to model the service time distribution. The approach is evaluated on empirical datasets from internet services (Blockchain, GitHub, Stackoverflow) and mobility systems (New York taxi cab).

Significance. If validated with quantitative evidence that the recurrent GAN faithfully reproduces empirical service time distributions (including tails), the work would offer a flexible, non-parametric alternative for modeling complex service dynamics where standard distributions fail. The integration of RNN arrival modeling with recurrent GAN service modeling represents a novel application of generative deep learning to queuing problems, with potential for improved predictions of waiting times and utilization across the cited application domains.

major comments (2)

[Abstract] Abstract: the claim that the proposed RNN + recurrent GAN combination 'addresses these limitations' of parametric queuing models is not supported by any quantitative results, baselines, error metrics (e.g., KS statistic, quantile matching, or end-to-end queuing simulation error), or validation details. Without such evidence it is impossible to determine whether the recurrent GAN avoids mode collapse or under-generation of heavy-tailed service events, which is load-bearing for the central claim that the non-parametric advantage materializes.
[Abstract] Abstract (evaluation paragraph): the description of the recurrent GAN for service times provides no architecture, training procedure, or loss details, leaving open the known risk that recurrent GANs can fail to capture multi-modal or heavy-tailed distributions common in the cited datasets (taxi, blockchain). This directly affects whether downstream queuing metrics improve over parametric baselines.

minor comments (1)

[Abstract] The phrase 'as to, for example,' in the abstract is slightly awkward and could be rephrased for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We address each major point below and agree that revisions to the abstract are warranted to better support the central claims with references to the paper's quantitative evaluations.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the proposed RNN + recurrent GAN combination 'addresses these limitations' of parametric queuing models is not supported by any quantitative results, baselines, error metrics (e.g., KS statistic, quantile matching, or end-to-end queuing simulation error), or validation details. Without such evidence it is impossible to determine whether the recurrent GAN avoids mode collapse or under-generation of heavy-tailed service events, which is load-bearing for the central claim that the non-parametric advantage materializes.

Authors: We agree that the abstract would be strengthened by explicitly referencing the quantitative results from the full manuscript. The paper evaluates the approach on the Blockchain, GitHub, Stackoverflow, and New York taxi datasets, including comparisons against parametric baselines and metrics for service time distribution fidelity as well as end-to-end queuing performance (waiting times and utilization). We will revise the abstract to summarize these findings and note the training stabilizations used to address mode collapse and heavy-tail capture. This directly supports the non-parametric advantage claim. revision: yes
Referee: [Abstract] Abstract (evaluation paragraph): the description of the recurrent GAN for service times provides no architecture, training procedure, or loss details, leaving open the known risk that recurrent GANs can fail to capture multi-modal or heavy-tailed distributions common in the cited datasets (taxi, blockchain). This directly affects whether downstream queuing metrics improve over parametric baselines.

Authors: The abstract's brevity limits full architectural disclosure, but we accept that a high-level mention would clarify the approach. We will revise the abstract to include a concise reference to the recurrent architecture, adversarial training with gradient penalty for stability, and the specific loss used to better capture multi-modal and heavy-tailed service times. Complete architecture, training procedure, and loss details remain in the methods section of the manuscript, where they support the reported queuing metric improvements. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical neural modeling with no derivation chain

full rationale

The paper proposes an RNN arrival model combined with a recurrent GAN service model as a data-driven replacement for parametric queuing assumptions. No equations, first-principles derivations, or predictions are presented that reduce to fitted inputs by construction. Evaluation is performed directly on external empirical datasets (Blockchain, GitHub, taxi, etc.), with no self-citation load-bearing steps or uniqueness theorems invoked. The approach is self-contained as a modeling methodology rather than a closed mathematical reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central claim rests on the untested assumption that neural networks can capture arbitrary service time distributions from finite event logs without additional regularization or validation.

invented entities (1)

recurrent generative adversarial network for service times no independent evidence
purpose: Model non-parametric service time distributions in queuing systems
Core novel component introduced to replace parametric assumptions

pith-pipeline@v0.9.0 · 5663 in / 918 out tokens · 22364 ms · 2026-05-25T17:12:18.813020+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 4 internal anchors

[1]

Martin Arjovsky, Soumith Chintala, and Léon Bottou. 201 7. Wasserstein Genera- tive Adversarial Networks. In Proceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research), Doina Precup and Yee Whye Teh (Eds.), Vol. 70. PMLR, International Conventio n Centre, Sydney, Australia, 214–223

work page
[2]

Søren Asmussen. 2008. Applied probability and queues. Vol. 51. Springer Science & Business Media

work page 2008
[3]

Paidamoyo Chapfuwa, Chenyang Tao, Chunyuan Li, Courtne y Page, Benjamin Goldstein, Lawrence Carin, and Ricardo Henao. 2018. Advers arial Time-to- Event Modeling. In ICML

work page 2018
[4]

Kyunghyun Cho, Bart Van Merriënboer, Dzmitry Bahdanau, and Yoshua Ben- gio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014
[5]

Daryl J Daley and David Vere-Jones. 2007. An introduction to the theory of point processes: volume II: general theory and structure . Springer Science & Business Media

work page 2007
[7]

Nan Du, Hanjun Dai, Rakshit Trivedi, Utkarsh Upadhyay, M anuel Gomez- Rodriguez, and Le Song. 2016. Recurrent marked temporal poi nt processes: Em- bedding event history to vector. In Proceedings of the 22nd ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining. ACM, 1555–1564

work page 2016
[8]

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Nets. In Advances in Neural Information Processing Systems 27 , Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (E ds.). Curran As- sociates, Inc., 2672–2680

work page 2014
[9]

Alex Graves. 2013. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850 (2013)

work page internal anchor Pith review Pith/arXiv arXiv 2013
[10]

Courville

Ishaan Gulrajani, Faruk Ahmed, Martín Arjovsky, Vince nt Dumoulin, and Aaron C. Courville. 2017. Improved Training of Wasserstein GANs. In NIPS

work page 2017
[11]

Alan G Hawkes and David Oakes. 1974. A cluster process re presentation of a self-exciting process. Journal of Applied Probability 11, 3 (1974), 493–503

work page 1974
[12]

Stephanie Hyland, CristÃşbal Esteban, and Gunnar RÃďt sch. 2018. Real-valued (Medical) Time Series Generation with Recurrent Condition al GANs

work page 2018
[13]

How Jing and Alexander J Smola. 2017. Neural survival re commender. In Pro- ceedings of the Tenth ACM International Conference on Web Search and Data Min- ing. ACM, 515–524

work page 2017
[14]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for s tochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014
[15]

John Frank Charles Kingman. 1993. Poisson processes. Wiley Online Library

work page 1993
[16]

Leonard Kleinrock. 1976. Queueing systems, volume 2: Computer applications . Vol. 66. wiley New York

work page 1976
[17]

Hongyuan Mei and Jason M Eisner. 2017. The neural hawkes process: A neurally self-modulating multivariate point process. In Advances in Neural Information Ojeda, et al. Processing Systems. 6754–6764

work page 2017
[18]

Hongyuan Mei and Jason M Eisner. 2017. The neural hawkes process: A neurally self-modulating multivariate point process. In Advances in Neural Information Processing Systems. 6738–6748

work page 2017
[19]

Olof Mogren. 2016. C-RNN-GAN: Continuous recurrent ne ural networks with adversarial training. CoRR abs/1611.09904 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[20]

Satoshi Nakamoto. 2008. Bitcoin: A peer-to-peer elect ronic cash system. (2008)

work page 2008
[21]

Henning Petzka, Asja Fischer, and Denis Lukovnikov. 20 18. On the regulariza- tion of Wasserstein GANs. In International Conference on Learning Representa- tions

work page
[22]

Charles Sutton and Michael I Jordan. 2011. Bayesian inf erence for queueing networks and modeling of internet services. The Annals of Applied Statistics (2011), 254–282

work page 2011
[23]

Cédric Villani. 2009. Optimal Transport: Old and New . Springer

work page 2009
[24]

Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J Smola , and How Jing

work page
[25]

In Proceedings of the tenth ACM inter- national conference on web search and data mining

Recurrent recommender networks. In Proceedings of the tenth ACM inter- national conference on web search and data mining . ACM, 495–503

work page
[26]

Tengfei Zhou, Hui Qian, Zebang Shen, Chao Zhang, Chengw ei Wang, Shichen Liu, and Wenwu Ou. 2018. JUMP: a joint predictor for user clic k and dwell time. In Proceedings of the 27th International Joint Conference on Ar tiﬁcial Intelligence. AAAI Press, 3704–3710

work page 2018

[1] [1]

Martin Arjovsky, Soumith Chintala, and Léon Bottou. 201 7. Wasserstein Genera- tive Adversarial Networks. In Proceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research), Doina Precup and Yee Whye Teh (Eds.), Vol. 70. PMLR, International Conventio n Centre, Sydney, Australia, 214–223

work page

[2] [2]

Søren Asmussen. 2008. Applied probability and queues. Vol. 51. Springer Science & Business Media

work page 2008

[3] [3]

Paidamoyo Chapfuwa, Chenyang Tao, Chunyuan Li, Courtne y Page, Benjamin Goldstein, Lawrence Carin, and Ricardo Henao. 2018. Advers arial Time-to- Event Modeling. In ICML

work page 2018

[4] [4]

Kyunghyun Cho, Bart Van Merriënboer, Dzmitry Bahdanau, and Yoshua Ben- gio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014

[5] [5]

Daryl J Daley and David Vere-Jones. 2007. An introduction to the theory of point processes: volume II: general theory and structure . Springer Science & Business Media

work page 2007

[6] [7]

Nan Du, Hanjun Dai, Rakshit Trivedi, Utkarsh Upadhyay, M anuel Gomez- Rodriguez, and Le Song. 2016. Recurrent marked temporal poi nt processes: Em- bedding event history to vector. In Proceedings of the 22nd ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining. ACM, 1555–1564

work page 2016

[7] [8]

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Nets. In Advances in Neural Information Processing Systems 27 , Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (E ds.). Curran As- sociates, Inc., 2672–2680

work page 2014

[8] [9]

Alex Graves. 2013. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850 (2013)

work page internal anchor Pith review Pith/arXiv arXiv 2013

[9] [10]

Courville

Ishaan Gulrajani, Faruk Ahmed, Martín Arjovsky, Vince nt Dumoulin, and Aaron C. Courville. 2017. Improved Training of Wasserstein GANs. In NIPS

work page 2017

[10] [11]

Alan G Hawkes and David Oakes. 1974. A cluster process re presentation of a self-exciting process. Journal of Applied Probability 11, 3 (1974), 493–503

work page 1974

[11] [12]

Stephanie Hyland, CristÃşbal Esteban, and Gunnar RÃďt sch. 2018. Real-valued (Medical) Time Series Generation with Recurrent Condition al GANs

work page 2018

[12] [13]

How Jing and Alexander J Smola. 2017. Neural survival re commender. In Pro- ceedings of the Tenth ACM International Conference on Web Search and Data Min- ing. ACM, 515–524

work page 2017

[13] [14]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for s tochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

work page internal anchor Pith review Pith/arXiv arXiv 2014

[14] [15]

John Frank Charles Kingman. 1993. Poisson processes. Wiley Online Library

work page 1993

[15] [16]

Leonard Kleinrock. 1976. Queueing systems, volume 2: Computer applications . Vol. 66. wiley New York

work page 1976

[16] [17]

Hongyuan Mei and Jason M Eisner. 2017. The neural hawkes process: A neurally self-modulating multivariate point process. In Advances in Neural Information Ojeda, et al. Processing Systems. 6754–6764

work page 2017

[17] [18]

Hongyuan Mei and Jason M Eisner. 2017. The neural hawkes process: A neurally self-modulating multivariate point process. In Advances in Neural Information Processing Systems. 6738–6748

work page 2017

[18] [19]

Olof Mogren. 2016. C-RNN-GAN: Continuous recurrent ne ural networks with adversarial training. CoRR abs/1611.09904 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[19] [20]

Satoshi Nakamoto. 2008. Bitcoin: A peer-to-peer elect ronic cash system. (2008)

work page 2008

[20] [21]

Henning Petzka, Asja Fischer, and Denis Lukovnikov. 20 18. On the regulariza- tion of Wasserstein GANs. In International Conference on Learning Representa- tions

work page

[21] [22]

Charles Sutton and Michael I Jordan. 2011. Bayesian inf erence for queueing networks and modeling of internet services. The Annals of Applied Statistics (2011), 254–282

work page 2011

[22] [23]

Cédric Villani. 2009. Optimal Transport: Old and New . Springer

work page 2009

[23] [24]

Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J Smola , and How Jing

work page

[24] [25]

In Proceedings of the tenth ACM inter- national conference on web search and data mining

Recurrent recommender networks. In Proceedings of the tenth ACM inter- national conference on web search and data mining . ACM, 495–503

work page

[25] [26]

Tengfei Zhou, Hui Qian, Zebang Shen, Chao Zhang, Chengw ei Wang, Shichen Liu, and Wenwu Ou. 2018. JUMP: a joint predictor for user clic k and dwell time. In Proceedings of the 27th International Joint Conference on Ar tiﬁcial Intelligence. AAAI Press, 3704–3710

work page 2018