Supply Chain Optimization via Generative Simulation and Iterative Decision Policies

Haifeng Chen; Haoyue Bai; Haoyu Wang; Nanxu Gong; Wangyang Ying; Xinyuan Wang; Yanjie Fu

REVIEW 4 major objections 6 minor 12 references

Supply Chain Optimization via Generative Simulation and Iterative Decision Policies

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A closed-loop simulator and policy network picks shipping modes that raise on-time delivery and profit.

desk verdict A plausible simulator-decision framework whose real-world gains are unproven because training and evaluation both depend on unvalidated counterfactual predictions. read the letter →

arxiv 2507.07355 v1 pith:LCAWFASA submitted 2025-07-10 cs.AI

Haoyue Bai , Haoyu Wang , Nanxu Gong , Xinyuan Wang , Wangyang Ying , Haifeng Chen , Yanjie Fu This is my paper

classification cs.AI

keywords Sim-to-DecsupplychainoptimizationshippingmodeselectiongenerativesimulationautoregressiveLSTMpolicynetworkcounterfactualpredictioncontextualbandit

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that shipping-mode decisions in supply chains can be optimized without hand-built rules by coupling a generative simulator with an iterative policy network. The simulator learns transportation dynamics from historical orders and predicts, for a chosen mode, how an order's delay risk, delivery time, and on-time status evolve. The policy network learns to pick modes by blending historical reward expectations with future reward estimates produced through simulator feedback. The authors report that on three real-world datasets this combined framework improves timely delivery rates and profit and that the simulator keeps its accuracy when the test distribution shifts. A sympathetic reader would care because a data-driven, closed-loop testbed would let logistics operators compare strategies before deployment at low risk.

What carries the argument

The load-bearing mechanism is the autoregressive generative simulator: an LSTM encoder fuses four order-attribute groups (product, customer, shipping, order) and a pooled global batch representation with an embedding of the candidate shipping mode, and an LSTM decoder sequentially produces each evolutionary attribute's latent embedding, with a dedicated predictor mapping each to a value. Because the decoder generates delay risk, delivery time, and on-time status one after another conditioning on earlier states, it captures interdependencies among them. The second mechanism is the policy network, trained with a combined loss that rewards historically successful modes and, through a value network, rewards modes whose simulator-evaluated batch outcomes—timely delivery rate plus profit—are high, with Gumbel-Softmax making the discrete mode choice differentiable.

What would settle it

Run the decision policy on a test set where some orders were deliberately assigned modes different from the historical policy, then compare the simulator's predicted on-time status and profit for those orders against the realized results; if switched-mode prediction accuracy is close to chance, the Table 2 improvements would not transfer to actual operations.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that treating supply chain simulation as an autoregressive generative task and coupling it tightly with a differentiable decision network yields a framework that outperforms both rule-based simulation and isolated decision methods. The simulator encodes order attributes and the selected shipping mode into a fused LSTM representation, then autoregressively decodes the order's evolutionary states, training against observed outcomes by squared error. The decision network scores candidate modes; a Gumbel-Softmax sample selects one, the frozen simulator produces batch-level timely delivery rate and profit, and a combined loss—historical expected reward plus future-reward prediction—refines the policy. The reported experiments show Sim-to-Dec leading on simulator accuracy and on the combined timely-delivery-profit objective, and the authors link that success to the fine-grained, interdependent state generation and to the tight simulation–decision loop.

Load-bearing premise

The framework's promised gains rest on the assumption that the generative simulator, trained only on observed order–mode–outcome triples, can correctly predict what would have happened under a shipping mode that was not actually used for that order, because both policy training and the reported evaluation rely on those simulated switched-mode outcomes.

Editorial extensions

If this is right

A logistics operator could apply the framework to a new dataset by retraining on that dataset's orders, with no domain-specific simulation rules to encode.
The single balancing weight λ lets an operator shift the policy between profit-oriented and delivery-oriented behavior without redesigning the optimizer.
The same autoregressive state-generation idea extends to other sequential shipping decisions, such as carrier assignment or route selection, whenever order states evolve in stages.
Because decision evaluation takes seconds on the tested datasets, iterative strategy comparison becomes feasible in everyday planning loops.
Under distribution shift, the simulator reproduces the delivery-time distribution shape better than Markov and one-shot generation baselines, supporting its use when demand patterns change.

Reading between the lines

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

If the simulator's counterfactual predictions are biased—never checked against realized outcomes for switched modes—the reported decision gains in Table 2 could be artifacts of the simulator rather than real logistics improvements; a held-out test with actually switched modes would settle this.
The framework optimizes aggregate batch profit and on-time rate, so order-level constraints like contractual delivery windows may need to be added before deployment in regulated or customer-facing settings.
Because the global pooled batch representation is part of the simulator input, the model implicitly assumes order dependence within a batch; extending it to online, streaming order arrival would require reformulating the pooling step.
The contextual-bandit view of the decision maker suggests the same policy training could be applied to any counterfactual evaluation environment where rewards are non-differentiable, not only supply chain simulators.

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 paper is a decent engineering attempt, but the headline result is not established. What is new is the specific combination: an autoregressive LSTM simulator that sequentially generates order status attributes, coupled with a policy network trained against simulated batch rewards and a historical per-mode average loss. The simulator itself does quite well on factual prediction—Table 1 shows large gains over Markov and prediction baselines on three public datasets, and the distribution-shift experiment is a reasonable sanity check that the simulator generalizes beyond its training distribution. That part is worth taking seriously.

The load-bearing problem is counterfactual validity. The policy is trained (Eq. 14, Eq. 16) and evaluated (Table 2) using the simulator's predicted outcomes for the chosen shipping mode. But the simulator is only trained on factual outcomes—the mode that was actually used for each order. For any order where the policy picks a different mode, the simulator is extrapolating, and there is no ground-truth data to check that extrapolation. The paper never validates switched-mode outcomes, so the improvements over 'Real' in Table 2 could simply be optimistic bias in the simulator's counterfactual predictions. The historical branch doesn't rescue this, because it uses only per-mode averages that are constant across orders, so it cannot provide order-level counterfactual signal.

The evaluation protocol has a secondary issue: the value network regresses each order's probability output to a batch-level reward (Eq. 16), which is scale-mismatched and uses the same target for every order in a mode group. That makes the learning signal coarse, and it is unclear what the decision network is actually learning beyond a mode classifier that picks the historically best average mode.

There are also smaller inaccuracies. The contributions claim validation on a live transportation system, but no such experiment appears anywhere in the body. That assertion needs to be removed or substantiated. The text also says all necessary outcomes are available from real datasets, which is misleading for non-chosen modes.

Overall, the framework is coherent and the simulator may be useful for factual prediction, but the central claim of improving real-world delivery and profit is unsupported as presented. The authors could fix this by reframing the contribution as simulator-based policy optimization, or by validating counterfactual predictions through a controlled experiment or an offline evaluation method that makes assumptions explicit.

Send it to peer review—the methodological issue is important and the authors deserve the chance to address it. A serious referee can help them either establish counterfactual validity or correct the claims.

Referee Report

4 major / 6 minor

Summary. The paper proposes Sim-to-Dec, a framework for shipping-mode selection in supply chain transportation. It couples a generative autoregressive simulator (LSTM encoder-decoder) that predicts order status attributes (delay risk, delivery time, on-time status) conditioned on order features and a chosen shipping mode, with a decision network that is trained by combining historical per-mode expected rewards and simulated future batch rewards. The authors evaluate the simulator's factual prediction accuracy on three public datasets (DataCo, GlobalStore, OAS), test its robustness under a distribution shift, and report decision-quality metrics (timely delivery rate and profit) for Sim-to-Dec against LP, RL, and LLM baselines. The central claim is that Sim-to-Dec significantly improves timely delivery rates and profit on real-world datasets and was validated on a live transportation system.

Significance. If the central claim were supported, the paper would offer a useful template for data-driven, rule-free simulation of transportation dynamics and for coupling that simulator with policy optimization. The strengths of the manuscript include a clearly described autoregressive simulator architecture, public code availability, experiments on three real-world datasets, and ablation/sensitivity analyses. However, the claimed real-world decision improvements are not established by the evidence presented: the decision policy is trained and evaluated on simulator outputs whose counterfactual validity for unobserved shipping modes is never checked, and the claimed live-system validation is absent from the body of the paper. These issues affect the paper's main empirical contribution, so the significance of the result as currently evidenced is limited.

major comments (4)

[§4.2, Eqs. (14)–(16); §5.2, Table 2] The central claim that Sim-to-Dec improves real-world timely delivery and profit is not supported because the policy is both trained and evaluated on outcomes produced by the simulator S, while S is trained only on observed (features, chosen mode, outcome) triples via Eq. (9). For any mode not historically selected for an order, S must extrapolate counterfactually, and the paper provides no validation of such counterfactual predictions. Table 1 validates only factual simulation accuracy, and the distribution-shift experiment in Figure 4 still evaluates factual prediction of fE_time under covariate shift, not mode switching. Since every non-Real row of Table 2 (including Sim-to-Dec) is scored using S's predicted outcomes, the reported gains over Real are compatible with an optimistically biased simulator and do not demonstrate realized logistics improvements. A concrete remedy would require a deployment or A/B experiment, or at minimum an offline counterfactual validation strategy (e.g., policy evaluation on logged bandit data), none of which appears.
[§4.2, Eq. (16)] The future-estimation loss Lf is not a valid regression target as written. R_batch^d is a batch-level reward computed from orders grouped by their sampled decisions, whereas \hat R_n^d is the per-order probability output for mode d; the squared difference between these quantities is dimensionally mismatched, and the target changes with the current policy's grouping. The paper does not describe how R_batch^d is assigned as a target for individual orders, so the claimed value-network training is underspecified and may not learn a meaningful reward prediction. This undermines the decision-training loop that Eq. (14) is supposed to support.
[§1 and §5] The contributions state that the framework is validated on a live transportation system, but no live experiment, protocol, or results appear anywhere in the manuscript. The experiments in Section 5 use only static historical datasets (DataCo, GlobalStore, OAS). Either the live validation must be reported and analyzed, or this claim must be removed; as written, the claim of real-world validation is unsupported.
[§5.2, Table 2] The paper acknowledges that LP enumerates all possible strategies and their associated rewards on the test set, which means the LP row in Table 2 is not a fair learned-policy comparison. This does not affect the main counterfactual-validity concern, but it should be addressed by using an LP policy trained on training-set statistics or by explicitly reporting the comparison as an oracle upper bound rather than as a baseline policy.

minor comments (6)

[§3] The sentence 'in our experiments, we has four shipping modes' contains a grammatical error and should be revised.
[§5.2] The text contains typos: 'ClobalStore' should be 'GlobalStore' and 'oveall' should be 'overall'.
[§2 and §5.2] The definitions of T_Time and T_Profit are never given as equations; the paper should specify how timely delivery rate and profit are computed from order attributes and simulated statuses, including how profit is normalized.
[§4.2, Eq. (17)] The hyperparameter lambda and other network hyperparameters (LSTM hidden size, number of layers, embedding dimension, learning rate, batch size) are not reported, which limits reproducibility despite the provided code.
[References] The reference list contains duplicated entries for Daroń 2022a and 2022b, and the DataCo dataset citation is malformed in the author list.
[§4.1, Eq. (3)] The notation for the global order group and the pooling operation is imprecise; the paper should clarify how Pool aggregates heterogeneous attribute representations across orders in a batch.

Circularity Check

1 steps flagged · score 8.0 of 10

Decision-policy 'improvements' are computed from the same unvalidated generative simulator used to train the policy, so Table 2's gains reduce to the simulator's counterfactual outputs rather than measured logistics outcomes.

fitted input called prediction [Section 4.2, Eqs. (14)-(16); evaluation in Section 5.2 Table 2]
"The simulator predicts statuses f^E_e,n based on the inherent attributes F^I_n and the selected decision f^D_d,n: f^E_n = S(F^I_n, f^D_d,n). The simulator is pre-trained, with its parameters frozen during decision-making. ... The orders are grouped according to the selected shipping mode f^D_d,n. We calculate the timely delivery rate and profit for each shipping mode using the necessary information, including order information, selected shipping mode, and the simulated order status."

The simulator S is trained only on realized (order features, historical mode, actual outcome) triples via Eq. (9). During decision-making, Eq. (14) uses the same S to produce the order status for whatever mode the policy selects, Eq. (15) computes batch rewards from those simulated statuses, and Eq. (16) trains the policy on that reward. For test-set orders whose selected mode differs from the historical mode, no ground-truth counterfactual outcome exists, so Table 2's T_Time and T_Profit for Sim-to-Dec are by construction the simulator's own predicted outcomes. Any optimistic bias in S for non-historical modes is therefore reported as a real logistics improvement.

full rationale

The central claim—that Sim-to-Dec significantly improves timely delivery and profit on real-world datasets—is not independently validated. The decision policy is trained against batch rewards computed by the generative simulator (Eqs. 14-16), and the same simulator's predicted statuses are used to compute the reported decision metrics in Table 2. Because counterfactual outcomes for unobserved shipping modes are never compared to any held-out ground truth, the 'improvements' are functions of the simulator's fitted outputs rather than evidence about real logistics outcomes. The paper's simulator-accuracy study (Table 1) only measures factual prediction accuracy on the historical mode, and the robustness check tests factual prediction under distribution shift, not switched-mode counterfactuals. The claim in the contributions that the framework is validated on a live transportation system is unsupported by any experimental protocol in the body, which further weakens the empirical grounding. No significant self-citation chain or imported uniqueness theorem is involved; the circularity is in the evaluation procedure itself. The abstract's promise of generalization across supply chain settings therefore rests on an unvalidated counterfactual simulator, and the reported gains reduce, by the paper's own equations, to the simulator's predictions.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the unverified counterfactual fidelity of the simulator and on several modeling choices that are either constant across orders or unreported. No new entities are introduced.

free parameters (2)

lambda (λ) in Eq. 17 = tuned via sensitivity analysis over 0.1 to 10
Balances the historical-experience loss and the future-estimation loss; no principled criterion for its value is given.
Neural network hyperparameters (LSTM hidden size, number of layers, embedding dimension, learning rate, batch size) = not reported
These numbers are chosen by hand or fit to validation data, but the paper omits them entirely, preventing reproduction and assessment of their influence.

assumptions (4)

domain assumption The simulator, trained only on observed (order features, chosen mode, outcome) triples, can accurately predict outcomes for shipping modes not observed in the training data.
This counterfactual validity is required by Eq. 14 for training the decision policy and by the evaluation in Table 2, but it is never validated against real switched-mode outcomes.
domain assumption The three order status attributes (delay risk, delivery time, on-time status) sufficiently capture the dynamics that determine timely delivery and profit.
The simulator and reward computation use only these three predicted attributes; any dynamics outside them are ignored. Stated in Section 3 and used throughout.
domain assumption A random 8:1:1 split of the data provides a valid evaluation without temporal leakage or distribution shift.
The paper claims random splitting mitigates temporal biases, but random splits can place later orders in training and earlier orders in test, allowing the model to look ahead. The distribution-shift experiment repartitions by f_E_time but does not use a temporal split.
standard math The MSE loss in Eq. 9 is an appropriate objective for simulator fidelity.
MSE is a standard regression loss, but for binary and categorical attributes it may not be the optimal choice; the paper does not justify it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Supply Chain Optimization via Generative Simulation and Iterative Decision Policies." pith.science (2026). https://pith.science/paper/LCAWFASA

@misc{pith2026250707355,
  author       = {Pith},
  title        = {Pith review of: Supply Chain Optimization via Generative Simulation and Iterative Decision Policies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCAWFASA}},
  note         = {Machine review of arXiv:2507.07355}
}

read the original abstract

High responsiveness and economic efficiency are critical objectives in supply chain transportation, both of which are influenced by strategic decisions on shipping mode. An integrated framework combining an efficient simulator with an intelligent decision-making algorithm can provide an observable, low-risk environment for transportation strategy design. An ideal simulation-decision framework must (1) generalize effectively across various settings, (2) reflect fine-grained transportation dynamics, (3) integrate historical experience with predictive insights, and (4) maintain tight integration between simulation feedback and policy refinement. We propose Sim-to-Dec framework to satisfy these requirements. Specifically, Sim-to-Dec consists of a generative simulation module, which leverages autoregressive modeling to simulate continuous state changes, reducing dependence on handcrafted domain-specific rules and enhancing robustness against data fluctuations; and a history-future dual-aware decision model, refined iteratively through end-to-end optimization with simulator interactions. Extensive experiments conducted on three real-world datasets demonstrate that Sim-to-Dec significantly improves timely delivery rates and profit.

Figures

Figures reproduced from arXiv: 2507.07355 by the authors.

**Figure 1.** Proposed Framework To meet these criteria and support adaptive, efficient transportation strategy design, we propose Simulation-toDecision (Sim-to-Dec), a novel framework that unifies generative simulation with end-to-end decision optimization. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Simulator 4.1 Generative Simulator via Deep Generative AI Traditional simulation approaches in supply chain shipping mode management heavily rely on expert-driven models with predefined heuristics, making them rigid and difficult to adapt to evolving market conditions. To address these limitations, we propose a generative simulator that models simulation as a generative AI task by learning directly from offline hist… view at source ↗

**Figure 3.** Decision-Maker The simulator captures system patterns from real data and can flexibly reflect the state changes of the system under decision-making. However, simple observation is not enough to support the need to reduce decision risks in a traceable way in the supply chain, so we propose a decision-maker model in this section, which elevates the observation of the data set to the level of intervention. By modifying… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Simulation under distribution shift on DataCo dataset indicators to the [0, 1] range, and report two aggregate metrics: the absolute difference between the two objectives (Diff), and their sum (Overall), to comprehensively assess decision quality. Baseline Algorithms. …

**Figure 5.** Figure 5: Ablation Study and Sensitivity Analysis Robustness Check of Generative Simulator against Distribution Shift. The dynamic complexity of the real world means that the environment in which the system operates is constantly changing. Mining potential patterns from data and…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 7 canonical work pages

[4]

Language Models are Few-Shot Learners

“Language Models are Few-Shot Learners”. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , edited by H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin. Caruana, R

work page 2020
[7]

Is a Good Representation Sufficient for Sample Efficient Reinforcement Learning?

“Is a Good Representation Sufficient for Sample Efficient Reinforcement Learning?”. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020: OpenReview.net. Forrester, J. W

work page 2020
[10]

Assessing supply chain responsiveness, resilience and robustness (Triple-R) by computer simulation: a systematic review of the literature

“Assessing supply chain responsiveness, resilience and robustness (Triple-R) by computer simulation: a systematic review of the literature”. Int. J. Prod. Res. 62(4):1458–1488 https://doi.org/10.1080/ 00207543.2023.2180302. Sutton, R. S., and A. G. Barto

arXiv 2023
[12]

Challenges and opportunities in integration of simulation and optimization in maritime logistics

“Challenges and opportunities in integration of simulation and optimization in maritime logistics”. In 2018 Winter Simulation Conference (WSC) , 2897–2908. IEEE. AUTHOR BIOGRAPHIES HAOYUE BAI is a PhD student at Arizona State University. His email address is baihaoyue621@gmail.com. HAOYU W ANGis a researcher in the DSSS department at NEC Labs America. His...

work page 2018
[1998]

Reinforcement Learning: An Introduction

“Reinforcement Learning: An Introduction”. IEEE Trans. Neural Networks 9(5):1054– 1054 https://doi.org/10.1109/TNN.1998.712192. Vinay34

arXiv 1998
[2018]

Non-Autoregressive Neural Machine Translation

“Non-Autoregressive Neural Machine Translation”. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings: OpenReview.net. Hosseini, S., D. Ivanov, and A. Dolgui

work page 2018
[2019]

DataCo SMART SUPPLY CHAIN FOR BIG DATA ANALYSIS

“DataCo SMART SUPPLY CHAIN FOR BIG DATA ANALYSIS” https://doi.org/10.17632/8gx2fvg2k6.5. Correa-Martinez, Y . C., and M. Seck

work page doi:10.17632/8gx2fvg2k6.5
[2020]

Survey on Applications of Multi-Armed and Contextual Bandits

“Survey on Applications of Multi-Armed and Contextual Bandits”. In IEEE Congress on Evolutionary Computation, CEC 2020, Glasgow, United Kingdom, July 19-24, 2020 , 1–8: IEEE https: //doi.org/10.1109/CEC48606.2020.9185782. Brailsford, S. C., T. Eldabi, M. Kunc, N. Mustafee, and A. F. Osorio

arXiv 2020

Show all 12 references

[2022]

A Model-based Multi-agent Framework to Enable an Agile Response to Supply Chain Disruptions *

“A Model-based Multi-agent Framework to Enable an Agile Response to Supply Chain Disruptions *”. In 18th IEEE International Conference on Automation Science and Engineering, CASE 2022, Mexico City, Mexico, August 20-24, 2022 , 235–241: IEEE https://doi.org/10.1109/CASE49997.20...

2022
[2023]

A review on reinforcement learning algorithms and applications in supply chain management

“A review on reinforcement learning algorithms and applications in supply chain management”. Int. J. Prod. Res. 61(20):7151–7179 https://doi.org/10.1080/00207543.2022. 2140221. Routroy, S., and A. Behera

2022
[2024]

Exploration of an Intelligent Decision-making System for International Freight Forwarding Based on Simulation Optimization

“Exploration of an Intelligent Decision-making System for International Freight Forwarding Based on Simulation Optimization”. In Proceedings of the 2024 9th International Conference on Intelligent Information Processing , 251–256. Churchman, C. W., R. L. Ackoff, and E. L. Arnoff

2024
[2025]

Brownian Bridge Augmented Surrogate Simulation and Injection Planning for Geological CO _2 Storage

“Brownian Bridge Augmented Surrogate Simulation and Injection Planning for Geological CO _2 Storage”. arXiv preprint arXiv:2505.18204 . Barykin, S. Y ., A. A. Bochkarev, O. V . Kalinina, and V . K. Yadykin

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [4]

Language Models are Few-Shot Learners

“Language Models are Few-Shot Learners”. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , edited by H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin. Caruana, R

work page 2020

[2] [7]

Is a Good Representation Sufficient for Sample Efficient Reinforcement Learning?

“Is a Good Representation Sufficient for Sample Efficient Reinforcement Learning?”. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020: OpenReview.net. Forrester, J. W

work page 2020

[3] [10]

Assessing supply chain responsiveness, resilience and robustness (Triple-R) by computer simulation: a systematic review of the literature

“Assessing supply chain responsiveness, resilience and robustness (Triple-R) by computer simulation: a systematic review of the literature”. Int. J. Prod. Res. 62(4):1458–1488 https://doi.org/10.1080/ 00207543.2023.2180302. Sutton, R. S., and A. G. Barto

arXiv 2023

[4] [12]

Challenges and opportunities in integration of simulation and optimization in maritime logistics

“Challenges and opportunities in integration of simulation and optimization in maritime logistics”. In 2018 Winter Simulation Conference (WSC) , 2897–2908. IEEE. AUTHOR BIOGRAPHIES HAOYUE BAI is a PhD student at Arizona State University. His email address is baihaoyue621@gmail.com. HAOYU W ANGis a researcher in the DSSS department at NEC Labs America. His...

work page 2018

[5] [1998]

Reinforcement Learning: An Introduction

“Reinforcement Learning: An Introduction”. IEEE Trans. Neural Networks 9(5):1054– 1054 https://doi.org/10.1109/TNN.1998.712192. Vinay34

arXiv 1998

[6] [2018]

Non-Autoregressive Neural Machine Translation

“Non-Autoregressive Neural Machine Translation”. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings: OpenReview.net. Hosseini, S., D. Ivanov, and A. Dolgui

work page 2018

[7] [2019]

DataCo SMART SUPPLY CHAIN FOR BIG DATA ANALYSIS

“DataCo SMART SUPPLY CHAIN FOR BIG DATA ANALYSIS” https://doi.org/10.17632/8gx2fvg2k6.5. Correa-Martinez, Y . C., and M. Seck

work page doi:10.17632/8gx2fvg2k6.5

[8] [2020]

Survey on Applications of Multi-Armed and Contextual Bandits

“Survey on Applications of Multi-Armed and Contextual Bandits”. In IEEE Congress on Evolutionary Computation, CEC 2020, Glasgow, United Kingdom, July 19-24, 2020 , 1–8: IEEE https: //doi.org/10.1109/CEC48606.2020.9185782. Brailsford, S. C., T. Eldabi, M. Kunc, N. Mustafee, and A. F. Osorio

arXiv 2020