Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Decision-Focused Continual Learning for Seaport Power-Logistics Scheduling: Generalization across Varying Tasks

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A Fisher-information penalty keeps a port's forecast-then-optimize model decision-optimal as vessel-arrival tasks evolve, without replaying data or retraining per task.

desk verdict A sensible EWC-plus-differentiable-KNN combination for continual decision-focused learning, but the headline regret numbers rest on an underspecified evaluation oracle. read the letter →

arxiv 2511.07938 v4 pith:W2ZKWKJL submitted 2025-11-11 cs.LG cs.SYeess.SY

Chuanqing Pu , Feilong Fan , Nengling Tai , Yan Xu , Wentao Huang , Honglin Wen This is my paper
classification cs.LGcs.SYeess.SY
keywords decision-focusedlearningcontinualpredict-then-optimizeseaportpower-logisticsschedulingelasticweightconsolidationFisherinformationdifferentiableoptimizationcatastrophicforgetting
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

Seaport scheduling follows a predict-then-optimize pipeline: forecasts of price and net load feed a scheduling optimizer whose costs are asymmetric, so statistical accuracy is not the same as decision quality. Decision-focused learning (DFL) trains forecasts on the downstream regret, but a DFL model tuned to one vessel-arrival configuration loses its edge (regret rises 2.20% to 8.67%) when the task stream changes. The paper's claim is that a Fisher-information penalty, applied as a quadratic regularizer between tasks, lets a single forecasting model stay aligned with decision costs across an evolving stream of scheduling tasks, without replaying past data or retraining per task. On six tasks calibrated to Jurong Port, this decision-focused continual learning (DFCL) achieves the lowest cumulative regret after each task, cutting regret by up to 2.5% and forgetting by up to 27.7% relative to naive online fine-tuning, with training time that stays roughly flat. If right, ports could keep one cost-aligned forecast model running indefinitely as vessel arrivals evolve, instead of rebuilding models for every new schedule.

What carries the argument

Two components carry the argument. First, the Fisher-information regularizer: after each task, the empirical Fisher matrix of the regret loss, i.e., weighted outer products of per-sample gradients, is treated as the precision matrix of a Laplace-approximated posterior, converting 'keep what earlier tasks needed' into a quadratic penalty on parameter drift, the Elastic-Weight-Consolidation idea. Second, the differentiable convex surrogate: a KNN regressor with a smooth soft top-k operator maps forecasts to the discrete quay-crane and berth assignments (P_QC, V), fixing them so the remaining scheduling problem is convex; implicit differentiation of its KKT system yields the decision gradient f

What would settle it

Re-run the six-task continual-learning comparison with test-time regret computed from the original mixed-integer scheduling problem (Eq. 2) instead of the convexified set (Eq. 16); if DFCL's margin over naive fine-tuning shrinks or vanishes, the reported gains are an artifact of the relaxation. As a second check, compare the surrogate-fixed (P_QC, V) with the MILP-optimal crane and berth assignments on the test set; the cost gap between the two decisions bounds how much of the claimed savings is real.

Watch

Extended reading notes

Core claim

The paper's central claim is that a port's forecasting model can stay decision-optimal across an unbounded stream of scheduling tasks, without replaying past data or retraining per task. Each new task is trained on its own regret plus a quadratic penalty built from the empirical Fisher information of earlier tasks, anchoring the parameters those tasks depended on. Because the embedded scheduling problem is mixed-integer, the discrete quay-crane and vessel-berth decisions are produced by a differentiable KNN surrogate with a soft top-k operator, leaving a convex subproblem that is differentiated through its KKT conditions. Across six Jurong-Port-calibrated tasks, the method achieves the lowes

Load-bearing premise

Everything rests on the KNN surrogate that fixes the discrete quay-crane and berth decisions in the convexified feasible set: if its choices diverge from the true mixed-integer schedule, the reported regrets measure a proxy problem rather than real port operating cost, and the paper never states whether test costs come from the original MILP or the relaxed problem.

Editorial extensions

If this is right

  • A single forecasting model can serve a port across an unbounded sequence of vessel-arrival configurations, with memory and cumulative training time staying roughly flat instead of growing with each new task.
  • Adding the Fisher-information penalty to decision-focused fine-tuning lowers cumulative regret by up to 2.5% and the forgetting measure by up to 27.7% over six tasks, relative to naive Adam-based continual DFL.
  • DFL models trained on a single task degrade on later tasks by 2.20% to 8.67% in regret, so ports that keep retraining on the latest data alone lose decision quality on earlier conditions.
  • End-to-end DFL reduces operating cost even when point-forecast MAE is slightly worse than statistical learning, because regret depends on the direction and timing of errors, not just their magnitude.
  • Task-wise and multi-task DFL retraining become unnecessary: DFCL's regularization updates replace joint retraining over accumulated historical data.

Reading between the lines

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

  • The reported dollar savings are meaningful only if test-time costs come from the true mixed-integer schedule; the paper never states whether evaluation uses the original MILP (Eq. 2) or the convexified relaxation (Eq. 16), so a reader should treat the magnitude of the gains as provisional until that is clarified.
  • The Fisher matrix is built from regret-gradient outer products, but Lemma 1 requires a correctly specified statistical model; since regret is not a likelihood, the Bayesian justification is heuristic, and the paper itself leaves the theory of task relatedness open for future work.
  • Appendix E.3 explicitly labels its illustrative single-day analysis as case-specific; the quantitative case rests on the six-task results, so that one-day narrative should not be generalized on its own.
  • The same recipe, EWC-style consolidation plus a surrogate for the discrete decisions, transfers naturally to other predict-then-optimize settings with shifting exogenous parameters, such as renewable trading under changing market rules.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes Decision-Focused Continual Learning (DFCL), an EWC-style regularized training scheme for predict-then-optimize pipelines in seaport power-logistics scheduling. The goal is to maintain decision quality across a stream of tasks with varying vessel arrivals, without replay or retraining. The method uses Fisher-information-based quadratic regularization to protect parameters important for past tasks, and a differentiable KNN-based convex surrogate to handle the non-convex mixed-integer logistics subproblem. Experiments calibrated to Jurong Port data over six tasks report that DFCL(EWC) achieves lower cumulative regret and less forgetting than a naive DFL(Adam) baseline, with stable training time and bounded memory.

Significance. If the evaluation oracle is faithful to the real scheduling problem, the paper addresses an important and underexplored problem: continual decision-focused learning over varying optimization tasks. The authors contribute a clear problem formulation, a concrete regularizer construction, a differentiable surrogate with an explicit code listing, and a realistic port-scheduling case study. The computational-cost comparison against task-wise and multi-task retraining is useful. However, the central empirical claim rests on an unspecified evaluation oracle and on single-seed comparisons, making the reported 2.5% regret and 27.7% forgetting improvements unverifiable from the manuscript as written.

major comments (4)
  1. [§4.2.1–4.2.2, Eq. (14)–(16), Tables 2–3] The paper never states whether the cost f in L_reg (Eq. 4) is evaluated by solving the original mixed-integer problem (2) or the convexified problem (16) with (P_QC,V) fixed by the KNN surrogate. The pipeline description in §4.2.1–4.2.2 suggests both training and evaluation use the surrogate: “the PLS scheduling problem is solved based on the forecasting results and the surrogate model” (§4, p.9). If evaluation uses (16), the reported regrets are not necessarily true port operating costs: the KNN outputs are weighted averages of historical decisions and need not satisfy the berth/crane constraints (A.2)–(A.4) for the current vessel set. Moreover, the KNN query u=[p_hat, pi_hat] (Eq. 14) omits the dynamic task parameters ξ_d, so the surrogate cannot distinguish tasks with similar forecasts but different vessel fleets. The headline effects are small (up to 2.5% regret, 27.7% forgetting), s
  2. [§5.4, Table 3] All regret, RR, and FM values in Table 3 are presented as single scalars with no error bars, no multiple seeds, and no statistical significance tests. The differences between DFCL(EWC) and DFL(Adam) are small relative to the task-to-task variation (e.g., cumulative regret 3164.40 vs 3240.01 after task 6, a 2.3% difference). Without repeated experiments, one cannot rule out that the ordering is due to initialization or optimization noise. The paper should report means and standard deviations over at least 5 seeds, and ideally a paired test across the task stream.
  3. [§3.2, Eq. (13) and Appendix B] The derivation of the precision matrix relies on Lemma 1 (the Fisher identity), which holds for the Hessian of a log-likelihood under a correctly specified statistical model. The training objective L_reg is a regret loss, not a log-likelihood. Appendix B acknowledges this and heuristically replaces the Hessian with the outer product of regret gradients (Eq. B.5), but no justification is given for why this 'Fisher-like' curvature is the right consolidation weight for a non-probabilistic decision loss. This weakens the theoretical grounding of the method. At minimum, the paper should either derive the regularizer from a proper probabilistic model of the decision loss or provide empirical evidence that the empirical-Fisher approximation is a reliable proxy for Hessian curvature in this setting.
  4. [§5.4, baselines] The continual-learning comparison is limited to DFL(Adam), DFL(Adam-F), and SBL. EWC is only compared against its absence (naive fine-tuning) and a simple layer-freezing variant. The abstract claims superiority over 'existing methods,' but the paper does not compare against established continual-learning regularizers such as online EWC, Synaptic Intelligence, MAS, or replay-based methods with bounded memory. Since the paper's contribution is specifically the use of Fisher regularization in a decision-focused setting, a comparison against at least one other regularization baseline is needed to establish that the benefit comes from the Fisher term rather than from any regularization.
minor comments (5)
  1. [Eq. (25)] The definition of FM_k is unclear: the term max_{i<k} E_Di(Lreg(Θ∗_j)) uses index i in the expectation but the task index j in the model parameters; it should likely be max_{i<j} E_Dj(Lreg(Θ∗_i)) or similar. Please clarify the notation.
  2. [§2.0.2, Eq. (2a)] The imbalance cost term is written as ρ+π⊤∆P+ − ρ−π⊤∆P−. Depending on the convention for positive/negative deviations, the sign of the second term might need a plus; please verify consistency with the imbalance settlement model.
  3. [Appendix C] The differentiable KNN implementation is a welcome reproducibility artifact, but it is not integrated with the PyTorch example in a runnable script (e.g., no training loop, no data loading). A full repository link would aid reproducibility.
  4. [§5.1.3] The datasets are public, but the synthetic scaling to port scale and the vessel-arrival configuration are not; please include a data-generation script or a detailed description of the scaling factors so that the experiments can be reproduced.
  5. [Abstract] The abstract says 'superior decision performance and cross-task generalization over existing methods' but the paper compares only against one gradient-free DFL, one surrogate-based DFL, and two continual-learning baselines; the claim is stronger than the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regularizer is an approximation and all methods share the same evaluation protocol.

full rationale

The derivation of the DFCL objective (Eqs. 7–13) is a Bayesian/Laplace approximation in which the precision matrix is approximated by an empirical Fisher of the regret loss (Eqs. 12–13; Appendix B). This is an acknowledged approximation (B.5), not an equation that identifies the predicted regret with the regularizer's inputs; the regularizer penalizes parameter movement, and the reported regret improvements in Table 3 are empirical, not forced by construction. The KNN surrogate (Eqs. 14–16) is used to convexify the embedded MILP and provide gradients; if test-time regret is also computed on the relaxed problem, the absolute cost numbers may not reflect true port operating cost, but this is an evaluation-oracle validity concern, not circularity, because the same surrogate is applied to all compared methods. Self-citations ([3], [4], [7], [14]) are to related work, baselines, or a standard big-M linearization technique; none is invoked as a uniqueness theorem or as the sole justification of the central claim. The empirical comparison is self-contained against external datasets and a shared protocol.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. It does introduce a KNN memory set and a soft-top-k operator, but these are computational modules, not postulated phenomena. The free parameters listed are the unstated knobs that control the regularizer and the surrogate.

free parameters (5)
  • beta_t (Fisher scale factors) = unspecified
    In Eq. (12), beta_t is described as 'interpreted as the number of samples in task tau_t', but no concrete schedule or tuning values are given. The effective strength of the EWC regularizer therefore has a hidden degree of freedom.
  • EWC regularization weight = not reported
    The quadratic term in Eq. (11) is added to the regret loss; no coefficient or schedule is specified in the paper, leaving the effective regularization strength a free parameter.
  • K (nearest neighbors in surrogate) = 5 (code default)
    The DifferentiableKNN module uses k=5 by default (Listing 1); no sensitivity analysis is reported.
  • Memory set size M = unspecified
    The size of the KNN memory set M in Eq. (14) is not reported; it affects the surrogate's accuracy and computational cost.
  • Tikhonov regularization epsilon = small (unspecified)
    Section 4.2.2 introduces epsilon||x||^2/2 to make the KKT system well-conditioned; the value of epsilon is not given and can affect gradient quality.
assumptions (4)
  • standard math Laplace approximation of the posterior as a Gaussian at its mode
    Used to derive Eq. (9)-(11); assumes the posterior over parameters is well-approximated by a Gaussian, which is a strong assumption for neural networks.
  • ad hoc to paper Fisher identity holds for the regret loss
    Appendix B replaces the Hessian of L_reg with an outer product of gradients 'to preserve Fisher-like curvature'; L_reg is not a negative log-likelihood, so the Fisher identity does not apply. This is a heuristic, not a derivation.
  • domain assumption Surrogate fidelity
    Fixing (P_QC,V) from KNN and optimizing over the convex set X_convex in Eq. (16) is assumed to preserve the operational economics of the original MILP; no validation against exact solutions is given on the test tasks.
  • domain assumption The convexified relaxation captures relevant port costs
    The non-convex constraints (A.2)-(A.4) (berth conflicts, crane assignment) are dropped by fixing V and P_QC; the paper assumes the remaining power-balance constraints dominate the cost signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decision-Focused Continual Learning for Seaport Power-Logistics Scheduling: Generalization across Varying Tasks." pith.science (2026). https://pith.science/paper/W2ZKWKJL

@misc{pith2026251107938,
  author       = {Pith},
  title        = {Pith review of: Decision-Focused Continual Learning for Seaport Power-Logistics Scheduling: Generalization across Varying Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W2ZKWKJL}},
  note         = {Machine review of arXiv:2511.07938}
}
read the original abstract

Power-logistics scheduling in modern seaports typically follows a predict-then-optimize pipeline. To enhance the decision quality of predictions, decision-focused learning has been proposed, which aligns the training of forecasting models with downstream decision outcomes. However, this end-to-end design inherently restricts the value of forecasting models to a specific task structure and therefore generalizes poorly to evolving tasks induced by varying vessel arrivals. We address this gap with a decision-focused continual learning framework that adapts online to a stream of scheduling tasks. Specifically, we introduce Fisher-information-based regularization to enhance cross-task generalization by preserving parameters critical to prior tasks. A differentiable convex surrogate is also developed to stabilize gradient backpropagation. The proposed approach enables learning a decision-aligned forecasting model across a varying task stream with sustainable long-term computational and memory requirements. Experiments calibrated to Jurong Port show improved decision performance and cross-task generalization over existing methods, together with reduced computational cost and a bounded memory footprint.

Figures

Figures reproduced from arXiv: 2511.07938 by the authors.

Figure 1
Figure 1. Illustration of different learning paradigms for predict-then-optimize pipelines. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the loss landscapes for SBL vs. DFL (left) and DFL vs. DFCL (right). The x-axes and y-axes represent [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. End-to-end training pipeline for DFCL in PLS. Each module represents a differentiable computational unit. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Loss heat maps of a DFL model trained on Task 1 and evaluated on Tasks 1 to 6. The x-axis represents the load [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Regret trends on Tasks 1 to 5 and average regret across all tasks after learning each task. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Training time comparison of different methods in continual learning setting. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith

  1. [1]

    Toward future green maritime transportation: An overview of seaport microgrids and all-electric ships,

    S. Fang, Y. Wang, B. Gou, and Y. Xu, “Toward future green maritime transportation: An overview of seaport microgrids and all-electric ships,”IEEE Trans. Veh. Technol., vol. 69, no. 1, pp. 207–219, 2019

    2019

  2. [2]

    Optimal energy-logistics dispatch for seaport green vehicles: Leveraging reversible solid oxide cell energy system for sustainable energy,

    Y. Lu, S. Fang, T. Niu, G. Chen, and R. Liao, “Optimal energy-logistics dispatch for seaport green vehicles: Leveraging reversible solid oxide cell energy system for sustainable energy,”IEEE Trans. Ind. Appl., early access, 2025

    2025

  3. [3]

    Collaborative scheduling of seaport integrated energy, logistics, and vessels: A bi-level Nash–Stackelberg–Nash game approach,

    Y. Zhou, Y. Li, W. Huang, S. Chen, H. Zang, and N. Tai, “Collaborative scheduling of seaport integrated energy, logistics, and vessels: A bi-level Nash–Stackelberg–Nash game approach,”Appl. Energy, vol. 400, Art. no. 126598, 2025

    2025

  4. [4]

    An adaptive MARL large model for dispatch strategy generation in logistics-energy spatiotemporal coordination of container seaports,

    Y. Huang, W. Huang, R. Li, T. Huang, C. Li, and N. Tai, “An adaptive MARL large model for dispatch strategy generation in logistics-energy spatiotemporal coordination of container seaports,”IEEE Trans. Smart Grid, early access, 2025

    2025

  5. [5]

    Smart “predict, then optimize

    A. N. Elmachtoub and P. Grigas, “Smart “predict, then optimize”,”Manage. Sci., vol. 68, no. 1, pp. 9–26, 2022. 18

    2022

  6. [6]

    Decision-focused learning for power system decision-making under uncertainty,

    H. Zhang, R. Li, Q. Du, J. Tao, S. Pineda, G. Kariniotakis, S. Camal, C. B. Monroc, M. Sun, C. Wan,et al., “Decision-focused learning for power system decision-making under uncertainty,”IEEE Trans. Power Syst., early access, 2025

    2025

  7. [7]

    Distributed stochastic operation of low-carbon port energy-logistics systems via learning to warm-start,

    F. Fan, C. Pu, N. Tai, J. Luo, Z. Chen, and X. Zhang, “Distributed stochastic operation of low-carbon port energy-logistics systems via learning to warm-start,”IEEE Trans. Ind. Appl., early access, 2025

    2025

  8. [8]

    Decision-focused learning: Foundations, state of the art, benchmark and future opportunities,

    J. Mandi, J. Kotary, S. Berden, M. Mulamba, V. Bucarey, T. Guns, and F. Fioretto, “Decision-focused learning: Foundations, state of the art, benchmark and future opportunities,”J. Artif. Intell. Res., vol. 80, pp. 1623–1701, 2024

    2024

Show all 32 references
  1. [9]

    Toward value-oriented renewable energy forecasting: An iterative learning approach,

    Y. Zhang, M. Jia, H. Wen, Y. Bian, and Y. Shi, “Toward value-oriented renewable energy forecasting: An iterative learning approach,”IEEE Trans. Smart Grid, early access, 2024

    2024

  2. [10]

    Feature-driven economic improvement for network-constrained unit commitment: A closed-loop predict-and-optimize framework,

    X. Chen, Y. Yang, Y. Liu, and L. Wu, “Feature-driven economic improvement for network-constrained unit commitment: A closed-loop predict-and-optimize framework,”IEEE Trans. Power Syst., vol. 37, no. 4, pp. 3104–3118, 2021

    2021

  3. [11]

    Predict-and-optimize robust unit commitment with statistical guar- antees via weight combination,

    R. Xie, Y. Chen, and P. Pinson, “Predict-and-optimize robust unit commitment with statistical guar- antees via weight combination,”IEEE Trans. Power Syst., early access, pp. 1–15, 2025, doi: 10.1109/T- PWRS.2025.3613360

    2025

  4. [12]

    A task-based day-ahead load forecasting model for stochastic economic dispatch,

    J. Han, L. Yan, and Z. Li, “A task-based day-ahead load forecasting model for stochastic economic dispatch,”IEEE Trans. Power Syst., vol. 36, no. 6, pp. 5294–5304, 2021

    2021

  5. [13]

    Effective end-to-end learning framework for economic dispatch,

    C. Lu, W. Jiang, and C. Wu, “Effective end-to-end learning framework for economic dispatch,”IEEE Trans. Netw. Sci. Eng., vol. 9, no. 4, pp. 2673–2683, 2022

    2022

  6. [14]

    End-to-end forecasting towards economic operation of microgrid using derivative-free learning,

    J. Yu, C. Pu, Z. Dou, C. Zhang, J. Li, and F. Fan, “End-to-end forecasting towards economic operation of microgrid using derivative-free learning,” inProc. IEEE Sustain. Power Energy Conf. (iSPEC), 2024, pp. 399–403

    2024

  7. [15]

    An integrated approach for value-oriented energy forecasting and data-driven decision-making application to renewable energy trading,

    T. Carriere and G. Kariniotakis, “An integrated approach for value-oriented energy forecasting and data-driven decision-making application to renewable energy trading,”IEEE Trans. Smart Grid, vol. 10, no. 6, pp. 6933–6944, 2019

    2019

  8. [16]

    Prescriptivetreesforintegratedforecasting and optimization applied in trading of renewable energy,

    A.Stratigakos, S.Camal, A.Michiorri, andG.Kariniotakis, “Prescriptivetreesforintegratedforecasting and optimization applied in trading of renewable energy,”IEEE Trans. Power Syst., vol. 37, no. 6, pp. 4696–4708, 2022. 19

    2022

  9. [17]

    Cost-oriented prediction intervals: On bridging the gap between forecasting and decision,

    C. Zhao, C. Wan, and Y. Song, “Cost-oriented prediction intervals: On bridging the gap between forecasting and decision,”IEEE Trans. Power Syst., vol. 37, no. 4, pp. 3048–3062, 2021

    2021

  10. [18]

    Electricity price prediction for energy storage system arbitrage: A decision-focused approach,

    L. Sang, Y. Xu, H. Long, Q. Hu, and H. Sun, “Electricity price prediction for energy storage system arbitrage: A decision-focused approach,”IEEE Trans. Smart Grid, vol. 13, no. 4, pp. 2822–2832, 2022

    2022

  11. [19]

    Perturbed decision-focused learning for modeling strategic energy storage,

    M. Yi, S. Alghumayjan, and B. Xu, “Perturbed decision-focused learning for modeling strategic energy storage,”IEEE Trans. Smart Grid, early access, 2025

    2025

  12. [20]

    Load data valuation in multi-energy systems: An end-to-end approach,

    Y. Zhou, Q. Wen, J. Song, X. Cui, and Y. Wang, “Load data valuation in multi-energy systems: An end-to-end approach,”IEEE Trans. Smart Grid, vol. 15, no. 5, pp. 4564–4575, 2024

    2024

  13. [21]

    Multi-task predict-then-optimize,

    B. Tang and E. B. Khalil, “Multi-task predict-then-optimize,” inProc. Int. Conf. Learn. Intell. Optim., 2023, pp. 506–522

    2023

  14. [22]

    End-to-end learning for fair multiobjective optimization under uncertainty,

    M. H. Dinh, J. Kotary, and F. Fioretto, “End-to-end learning for fair multiobjective optimization under uncertainty,”arXiv preprintarXiv:2402.07772, 2024

    2024 arXiv

  15. [23]

    Adaptive decision-objective loss for forecast-then-optimize in power systems,

    H. Zhang, R. Li, M. Sun, and T. Fei, “Adaptive decision-objective loss for forecast-then-optimize in power systems,”arXiv preprintarXiv:2312.13501, 2023

    2023 arXiv

  16. [24]

    Same accuracy, twice as fast: Continuous training surpasses retraining from scratch,

    E. Verwimp, G. Hacohen, and T. Tuytelaars, “Same accuracy, twice as fast: Continuous training surpasses retraining from scratch,”arXiv preprintarXiv:2502.21147, 2025

    2025 arXiv

  17. [25]

    OptNet: Differentiable optimization as a layer in neural networks,

    B. Amos and J. Z. Kolter, “OptNet: Differentiable optimization as a layer in neural networks,” inProc. Int. Conf. Mach. Learn., 2017, pp. 136–145

    2017

  18. [26]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatricket al., “Overcoming catastrophic forgetting in neural networks,”Proc. Natl. Acad. Sci. USA, vol. 114, pp. 3521–3526, 2017

    2017

  19. [27]

    Overcoming catastrophic forgetting in graph neural networks with experience replay,

    F. Zhou and C. Cao, “Overcoming catastrophic forgetting in graph neural networks with experience replay,” inProc. AAAI Conf. Artif. Intell., vol. 35, no. 5, pp. 4714–4722, 2021

    2021

  20. [28]

    Softmax Backpropagation: A Smooth Approximation for Top-kSelection,

    J. Su, “Softmax Backpropagation: A Smooth Approximation for Top-kSelection,”Science Space Blog, Sep. 2024. [Online]. Available:https://spaces.ac.cn/archives/10373

    2024

  21. [29]

    Differentiable convex opti- mization layers,

    A. Agrawal, B. Amos, S. Barratt, S. Boyd, S. Diamond, and J. Z. Kolter, “Differentiable convex opti- mization layers,”Adv. Neural Inf. Process. Syst., vol. 32, 2019

    2019

  22. [30]

    D. J. C. MacKay,Information Theory, Inference and Learning Algorithms. Cambridge, U.K.: Cam- bridge Univ. Press, 2003. 20

    2003

  23. [31]

    The hybrid renewable energy forecasting and trading competition 2024,

    J. Browell, D. van der Meer, H. Kälvegren, S. Haglund, E. Simioni, R. J. Bessa, and Y. Wang, “The hybrid renewable energy forecasting and trading competition 2024,”arXiv preprintarXiv:2507.01579, 2025

    2024

  24. [32]

    Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond,

    T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, and R. J. Hyndman, “Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond,”Int. J. Forecast., vol. 32, no. 3, pp. 896–913, 2016. 21 Appendix A. The constraints of ESS and logistics system ...

    2014

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.