Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Smart Surrogate Losses for Contextual Stochastic Linear Optimization with Robust Constraints

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A convex surrogate loss, SPO-RC+, is Fisher consistent with the SPO-RC loss in robust-constrained contextual linear optimization, and training on truncated, importance-reweighted data improves decisions while preserving feasibility.

desk verdict A sensible extension of SPO to uncertain constraints, but the headline Fisher consistency claim is not proven as stated. read the letter →

arxiv 2505.22881 v1 pith:J4HEXTVQ submitted 2025-05-28 cs.LG math.OCstat.ML

Hyungki Im , Wyame Benslimane , Paul Grigas This is my paper
classification cs.LGmath.OCstat.ML MSC 90C1590C25
keywords contextualstochasticlinearoptimizationpredict-then-optimizedecision-focusedlearningrobustconformalpredictionimportancereweightingSPOlossFisherconsistency
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

Contextual stochastic linear optimization usually assumes uncertainty only in the objective; this paper tackles the harder case where constraints also depend on uncertain parameters predicted from context. It introduces SPO-RC, a loss that measures the decision error of a predicted cost vector under a robust feasibility set, and SPO-RC+, a convex surrogate, and proves that the two losses are Fisher consistent: under stated assumptions and an unrestricted hypothesis class, minimizing the surrogate recovers the conditional expectation $E[c|x]$. Because SPO-RC+ is only a valid upper bound when the true constraint parameter lies inside the uncertainty set, the authors train on truncated data and correct the induced selection bias with importance reweighting. On fractional knapsack and alloy production problems, SPO-RC+ with truncation and reweighting yields near-zero infeasibility and lower normalized decision error than predict-then-optimize or MSE training, especially as cost-to-feature complexity grows.

What carries the argument

The central object is the SPO-RC+/cost+ surrogate: for a fixed uncertainty set $U$, $\mathrm{cost}^+(\hat c,c;U)=\max_{w\in S}(c-2\hat c)^\top w + 2\hat c^\top w^*(c,U)$, a convex function of $\hat c$ whose subgradient is $2(w^*(c,U)-w^*(2\hat c-c,U))$. Because $U(x)$ is fixed once $x$ is fixed, the Fisher-consistency proof reduces to the classical SPO+ argument, with the SPO-RC loss adding the penalty $\Delta_S(C)$ for out-of-set realizations. The second mechanism is covariate-shift correction: training data are truncated to the feasibility-guaranteed region $a\in U(x)$, and KMM estimates the importance weight $\beta(x)=P_D(x)/P_{\tilde D}(x)$, which Lemma 3.3 shows preserves the conditional law of $c$ given $x$ when $c$ and $a$ are conditionally independent.

What would settle it

Generate data where $c$ and $a$ share a latent variable $z$ given $x$ (e.g., $a=z+\epsilon_a$, $c=z+\epsilon_c$), construct $U(x)$ by split conformal prediction, truncate to $a\in U(x)$, train SPO-RC+ with KMM reweighting, and compare the decisions against an oracle that knows $E[c|x]$; if the reweighted rule remains as biased as the unweighted truncated rule in the heavily truncated region, the conditional-independence premise is doing the work and the reweighting claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that decision-focused learning extends cleanly to robust-constrained linear problems. For a fixed uncertainty set $U(x)$, the cost+ metric inherited from SPO remains a convex upper bound on the true cost of the robust-optimal decision, with subgradient $2(w^*(c,U)-w^*(2\hat c-c,U))$, and Theorem 2.4 proves that its expected minimizer coincides with that of the true cost metric and with $E[c|x]$ under uniqueness of the robust optimum, symmetric continuous conditional cost distributions, and nonempty interior. The SPO-RC loss adds a feasibility-sensitive term: when the realized constraint parameter $a$ falls outside $U(x)$, the loss is set to an upper bound $\Delta_S(C)$ rather than allowing the negative regret that would otherwise occur. The paper further claims that truncating the training set to $a\in U(x)$ and reweighting by $\beta(x)=P_D(x)/P_{\tilde D}(x)$, estimated by KMM, leaves the conditional distribution of $c$ given $x$ unchanged under conditional independence of $c$ and $a$, so the reweighted truncated objective still targets $E[c|x]$. Experiments on fractional knapsack and alloy production show the method maintains feasibility—infeasibility around 0.02% versus 45% for direct predict-then-optimize—and improves decision error as cost complexity rises.

Load-bearing premise

The load-bearing premise is Assumption 3.2: conditioned on the context $x$, the cost vector $c$ and the constraint vector $a$ are independent, so truncating on $a$ does not change the conditional law of $c$ and only the $x$-marginal shift needs reweighting.

Editorial extensions

If this is right

  • Under the assumptions of Theorem 2.4, training with the convex SPO-RC+ surrogate instead of the discontinuous SPO-RC loss yields the same optimal predictor, namely $E[c|x]$.
  • In the fractional knapsack experiments, every method that solves the robust formulation with an uncertainty set keeps test infeasibility near 0.02%, whereas direct predict-then-optimize with the predicted constraint parameter is infeasible in about 45% of test instances.
  • As the polynomial degree relating features to costs increases, SPO-RC+ models consistently outperform MSE-trained linear models on normalized SPO-RC test error, while MSE can be competitive or better at low complexity.
  • Combining SPO-RC+ with solution caching reduces training time dramatically with only slight degradation in decision quality across the datasets considered.
  • The generalization bounds in Appendix B extend prior predict-then-optimize bounds to context-dependent feasibility sets and to importance-reweighted truncated training data.

Reading between the lines

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

  • The conditional-independence assumption (Assumption 3.2) is untested in the experiments; when costs and constraint coefficients share latent drivers, truncation distorts $c|x$ and KMM corrects only the $x$-marginal shift. A natural extension is to reweight using an estimate of $P(a\in U(x)\mid x)$ without imposing independence.
  • Fisher consistency is proven for the cost metrics and then transferred to SPO-RC losses; the transfer silently requires $P(a\in U(x)\mid x)>0$. If coverage is zero on a region, SPO-RC degenerates to the constant penalty $\Delta_S(C)$ there and the surrogate is no longer an upper bound.
  • The paper's synthetic experiments generate $c$ and $a$ with independent noise, so they satisfy Assumption 3.2 by construction; applying the method to real data where this fails is the key untested regime.
  • A similar surrogate could likely be derived for nonlinear objectives, but the current SPO-RC+ relies on the linearity of expectation in the objective, so extending it would require a different convexification.
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

3 major / 4 minor

Summary. The paper studies contextual stochastic linear optimization with uncertain parameters in both the objective and the constraints. It constructs contextual uncertainty sets via conformal prediction, defines a robust predict-then-optimize decision rule, and introduces two loss functions: the SPO-RC loss, which compares the cost of the robust solution with the hindsight-optimal cost, and its convex surrogate SPO-RC+. The authors prove Fisher consistency of the cost and cost+ metrics under symmetry conditions, claim that this consistency transfers to the SPO-RC/SPO-RC+ loss pair, and propose training on the truncated dataset where a is in U(x), with KMM importance reweighting to correct the induced covariate shift. Experiments on fractional knapsack and alloy production problems compare models trained on original, truncated, and reweighted data, and report that SPO-RC+ with reweighting yields low SPO-RC test loss and near-zero infeasibility.

Significance. If the consistency claim can be properly repaired, the paper makes a useful contribution: it extends the SPO/SPO+ framework to settings where the feasible set itself is uncertain and context-dependent, and it offers a practical truncation-plus-reweighting scheme. The convexity of cost+ and its subgradient formula (Proposition 2.2) are correct and directly usable; the proof of cost/cost+ Fisher consistency is legitimate and borrows parameter-free propositions from prior work rather than introducing circular arguments; and the authors provide generalization bounds in Appendix B as well as reproducible experiments built on PyEPO. The main risks are the under-specified transfer from cost metrics to the SPO-RC losses and the reliance on conditional independence and positivity in the reweighting argument, neither of which is stated precisely or tested experimentally.

major comments (3)
  1. [Appendix A.1, Section 2.3] The final transfer step in the proof of Theorem 2.4 asserts that the minimizers f*_cost and f*_cost+ 'remain the same' when cost/cost+ are replaced by the SPO-RC/SPO-RC+ losses. This is not true on the original distribution D as stated: taking expectations, E_D[ell_SPO-RC] = E_D[cost(c_hat,c;U) 1{a in U}] + constant, whereas E_D[ell_SPO-RC+] = E_D[cost+(c_hat,c;U)] + constant. The indicator 1{a in U} only factors out of the first expectation when c and a are conditionally independent given x (Assumption 3.2) and P(a in U(x)|x) > 0 almost surely; neither condition appears in Section 2.3 or in Appendix A.1. Since the abstract and contribution list claim Fisher consistency between the SPO-RC and SPO-RC+ losses, this is a load-bearing gap; please either add these assumptions to the consistency theorem and prove the factorization, or restrict the consistency claim to the truncated distribution and re-derive the reweighting objective accordingly.
  2. [Section 3.2, Lemma 3.3] Lemma 3.3 and the importance-reweighting argument require, in addition to Assumption 3.2, the strict positivity condition P_D(a in U(x)|x) > 0 for (almost) every x in the support of D_x. Without it, the truncated distribution has zero mass at some x, P_{\tilde D}(c|x) is undefined, and the weight beta(x)=P_D(x)/P_{\tilde D}(x) is not finite. This condition is never stated. Moreover, if Assumption 3.2 fails, truncation changes the conditional law of c given x, and the reweighted objective targets E[c|x,a in U(x)] rather than E[c|x]; the synthetic generators in Appendices C.2 and C.3 draw the cost noise and constraint noise independently given x, so the experiments satisfy the assumption by construction and cannot reveal this failure mode. Please state the positivity condition and include a discussion of the dependent case.
  3. [Appendix B.1, Lemma B.3 and Proposition B.4] Lemma B.3 and Proposition B.4 assume that the importance weights beta(x) are known exactly (and bounded by B), whereas Algorithm 1 uses KMM-estimated weights. The generalization bound therefore does not cover the actual training procedure; it needs an additional term controlling the estimation error of the KMM weights, or an explicit assumption that the KMM tolerance makes the bias negligible. Please state this limitation or provide the missing estimation-error analysis.
minor comments (4)
  1. [Section 2.1] The feasible set of the robust problem P(c_hat,U) is denoted both S and S(x) without a clear definition; please fix the notation and define the feasible set explicitly as {w in S : h(w;a) <= 0 for all a in U(x)}.
  2. [Section 4, NormSPORCTest] The test metric uses ell_SPO-RC, but the text then says 'we set the loss ell_SPO-RC+ ... to the numerator value'; this appears to be a typo, and the exact infeasibility penalty used in the reported numbers should be stated clearly.
  3. [Appendix B, Theorem B.1] The theorem uses Omega_S(C) as though the cost metric were nonnegative and bounded by Omega; since cost can be negative, the Hoeffding and Rademacher constants should be stated in terms of the actual range of |cost|.
  4. [Appendix C.1.2] The reported 'almost perfect' NormSPORCTest value of 0.2% for DIR is evaluated only on the region x < 0.5; please state this restriction in the main text or figure caption to avoid overclaiming.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proof reuses prior published propositions as external lemmas, and no fitted quantity is relabeled as a prediction.

full rationale

The derivation chain is self-contained in the relevant sense. The only author-overlap citations are Propositions A.1 and A.2, taken from Elmachtoub and Grigas (2022), and the generalization-bound extension of El Balghiti et al. (2023). These are published, parameter-free results with stated assumptions that do not include the SPO-RC consistency claim being proved; under the review rule, such reuse counts as independent support rather than circularity. The SPO-RC+ loss is a constant shift of the cost+ metric with respect to the prediction, so its minimizer matching the cost+ minimizer is structural and not a fitted-input prediction. The paper's assertion that the SPO-RC and SPO-RC+ minimizers coincide with the cost/cost+ minimizers is under-justified in Appendix A.1 as written, since the indicator 1{a in U(x)} in SPO-RC can change the objective unless one also uses truncation or assumes conditional independence (Assumption 3.2). That is a correctness or assumption gap, not an equivalence-by-construction or self-citation circularity. The experiments are evaluated on out-of-sample test sets, and the KMM importance weights are estimated from separate target data rather than fitted to the reported test metric. No load-bearing step reduces to its own inputs, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central theoretical claims rest on standard convex-analysis propositions from prior work plus two domain assumptions: singleton optimal solutions and central symmetry (Assumption 2.3), and conditional independence of cost and constraint parameters given context (Assumption 3.2). The conformal coverage result relies on exchangeability. The use of KMM assumes the importance weight is bounded, which is stated in Lemma B.3 but not guaranteed by the estimation procedure. No physically invented entities are introduced; SPO-RC and SPO-RC+ are new loss functions with explicit definitions.

free parameters (5)
  • conformal miscoverage level alpha = 0.2 in experiments
    Set by the user to size the uncertainty set; controls the feasibility and performance trade-off. Not derived from first principles.
  • KMM bound B = 1000
    Set in experiments to constrain importance weights; chosen large because ratios rarely exceed 20.
  • KMM tolerance epsilon = (sqrt(m)-1)/sqrt(m)
    Chosen as a function of truncated sample size m; controls distribution matching tolerance.
  • Gaussian kernel bandwidth = implicit (k=exp(-||xi-xj||^2/2))
    Fixed in experiments; not tuned.
  • cost complexity degree deg_c = 2, 4, 6, 8 (varied); deg_a=4
    Synthetic data generation parameters that define instance difficulty; not method parameters but affect reported results.
assumptions (5)
  • domain assumption W*(E[c|x], Uhat(x)) is a singleton almost surely, c|x is centrally symmetric and continuous, and the interior of S(x) is non-empty (Assumption 2.3).
    Assumed in Theorem 2.4 to guarantee unique minimizers for the cost and cost+ risks; may fail for LPs with ties or asymmetric noise.
  • domain assumption c and a are conditionally independent given x (Assumption 3.2).
    Used in Lemma 3.3 to show truncation preserves the conditional distribution of c given x, reducing the bias to covariate shift on x.
  • standard math Calibration and test samples are exchangeable (i.i.d.) for split conformal prediction (Proposition 3.1).
    Standard conformal coverage guarantee; stated in Proposition 3.1.
  • ad hoc to paper P(a in U(x)|x) > 0 for all x (strict positivity of the truncation probability).
    Implied but not stated; the argmin of the truncated SPO-RC loss equals the argmin of the full loss only when this holds pointwise.
  • domain assumption The robust counterpart is tractable and admits a feasible solution (Section 2).
    Required for the method to compute w* and subgradients; holds for the LP and SOCP reformulations used.
invented entities (2)
  • SPO-RC loss independent evidence
    purpose: Measures decision error under robust constraints and penalizes uncertainty sets that either over-cover or fail to contain the true parameter.
    Explicitly defined in Section 2.3 and evaluated numerically; it is a loss function, not an unobserved physical entity.
  • SPO-RC+ loss independent evidence
    purpose: Convex surrogate for SPO-RC used for gradient-based training.
    Defined as cost+ minus the hindsight optimal cost; computed in experiments and in the consistency analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Smart Surrogate Losses for Contextual Stochastic Linear Optimization with Robust Constraints." pith.science (2026). https://pith.science/paper/J4HEXTVQ

@misc{pith2026250522881,
  author       = {Pith},
  title        = {Pith review of: Smart Surrogate Losses for Contextual Stochastic Linear Optimization with Robust Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4HEXTVQ}},
  note         = {Machine review of arXiv:2505.22881}
}
read the original abstract

We study an extension of contextual stochastic linear optimization (CSLO) that, in contrast to most of the existing literature, involves inequality constraints that depend on uncertain parameters predicted by a machine learning model. To handle the constraint uncertainty, we use contextual uncertainty sets constructed via methods like conformal prediction. Given a contextual uncertainty set method, we introduce the "Smart Predict-then-Optimize with Robust Constraints" (SPO-RC) loss, a feasibility-sensitive adaptation of the SPO loss that measures decision error of predicted objective parameters. We also introduce a convex surrogate, SPO-RC+, and prove Fisher consistency with SPO-RC. To enhance performance, we train on truncated datasets where true constraint parameters lie within the uncertainty sets, and we correct the induced sample selection bias using importance reweighting techniques. Through experiments on fractional knapsack and alloy production problem instances, we demonstrate that SPO-RC+ effectively handles uncertainty in constraints and that combining truncation with importance reweighting can further improve performance.

Figures

Figures reproduced from arXiv: 2505.22881 by the authors.

Figure 1
Figure 1. Visualization of the importance reweighting toy example [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Out-of-sample test set (size 3000) NormSPORCTest values of linear models with MSE and [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison across different tasks and evaluation criteria: (a)–(b) report [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of the truncation toy example [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 44 canonical work pages

  1. [1]

    Agrawal, A., Amos, B., Barratt, S., Boyd, S., Diamond, S., and Kolter, J. Z. (2019). Differentiable convex optimization layers. Advances in neural information processing systems , 32

    work page 2019

  2. [2]

    and Kolter, J

    Amos, B. and Kolter, J. Z. (2017). Optnet: Differentiable optimization as a layer in neural networks. In International conference on machine learning , pages 136--145. PMLR

    work page 2017

  3. [3]

    Angelopoulos, A. N. and Bates, S. (2021). A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511

    arXiv 2021

  4. [4]

    and Rudin, C

    Ban, G.-Y. and Rudin, C. (2019). The big data newsvendor: Practical insights from machine learning. Operations Research , 67(1):90--108

    work page 2019

  5. [5]

    Bartlett, P. L. and Mendelson, S. (2002). Rademacher and gaussian complexities: Risk bounds and structural results. Journal of Machine Learning Research , 3(Nov):463--482

    work page 2002

  6. [6]

    and Delage, E

    Bazier-Matte, T. and Delage, E. (2020). Generalization bounds for regularized portfolio selection with market side information. INFOR: Information Systems and Operational Research , 58(2):374--401

    work page 2020

  7. [7]

    and Kallus, N

    Bertsimas, D. and Kallus, N. (2020). From predictive to prescriptive analytics. Management Science , 66(3):1025--1044

    work page 2020

  8. [8]

    and Koduri, N

    Bertsimas, D. and Koduri, N. (2022). Data-driven optimization: A reproducing kernel hilbert space approach. Operations Research , 70(1):454--471

    work page 2022

Show all 53 references
  1. [9]

    and Van Parys, B

    Bertsimas, D. and Van Parys, B. (2022). Bootstrap robust prescriptive analytics. Mathematical Programming , 195(1):39--78

    2022

  2. [10]

    Bickel, S., Br \"u ckner, M., and Scheffer, T. (2009). Discriminative learning under covariate shift. Journal of Machine Learning Research , 10(75):2137--2155

    2009

  3. [11]

    Blondel, M., Berthet, Q., Cuturi, M., Frostig, R., Hoyer, S., Llinares-L \'o pez, F., Pedregosa, F., and Vert, J.-P. (2022). Efficient and modular implicit differentiation. Advances in neural information processing systems , 35:5230--5242

    2022

  4. [12]

    and Kwon, R

    Butler, A. and Kwon, R. H. (2023). Efficient differentiable quadratic programming layers: an admm approach. Computational Optimization and Applications , 84(2):449--476

    2023

  5. [13]

    Chen, W., Tanneau, M., and Van Hentenryck, P. (2023). End-to-end feasible optimization proxies for large-scale economic dispatch. IEEE Transactions on Power Systems

    2023

  6. [14]

    R., Bandi, N., and Delage, E

    Chenreddy, A. R., Bandi, N., and Delage, E. (2022). Data-driven conditional robust optimization. Advances in Neural Information Processing Systems , 35:9525--9537

    2022

  7. [15]

    Chenreddy, A. R. and Delage, E. (2024). End-to-end conditional robust optimization. In Proceedings of the Fortieth Conference on Uncertainty in Artificial Intelligence , pages 736--748

    2024

  8. [16]

    and Sen, S

    Deng, Y. and Sen, S. (2022). Predictive stochastic programming. Computational Management Science , 19(1):65--98

    2022

  9. [17]

    N., Grigas, P., and Tewari, A

    El Balghiti, O., Elmachtoub, A. N., Grigas, P., and Tewari, A. (2023). Generalization bounds in the predict-then-optimize framework. Mathematics of Operations Research , 48(4):2043--2065

    2023

  10. [18]

    predict, then optimize

    Elmachtoub, A. N. and Grigas, P. (2022). Smart “predict, then optimize”. Management Science , 68(1):9--26

    2022

  11. [19]

    N., Liang, J

    Elmachtoub, A. N., Liang, J. C. N., and McNellis, R. (2020). Decision trees for decision-making under the predict-then-optimize framework. In International conference on machine learning , pages 2858--2867. PMLR

    2020

  12. [20]

    and Morales, J

    Esteban-P \'e rez, A. and Morales, J. M. (2022). Distributionally robust stochastic programs with side information based on trimmings. Mathematical Programming , 195(1):1069--1105

    2022

  13. [21]

    Estes, A. S. and Richard, J.-P. P. (2023). Smart predict-then-optimize for two-stage linear programs with side information. INFORMS Journal on Optimization , 5(3):295--320

    2023

  14. [22]

    Fontana, M., Zeni, G., and Vantini, S. (2023). Conformal prediction: a unified review of theory and new challenges. Bernoulli , 29(1):1--23

    2023

  15. [23]

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., and Lempitsky, V. (2016). Domain-adversarial training of neural networks. Journal of machine learning research , 17(59):1--35

    2016

  16. [24]

    C., and Lee, J

    Hu, X., Lee, J. C., and Lee, J. H. (2023). Predict+ optimize for packing and covering lps with unknown parameters in constraints. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 37, pages 3987--3995

    2023

  17. [25]

    Huang, J., Gretton, A., Borgwardt, K., Sch \"o lkopf, B., and Smola, A. (2006). Correcting sample selection bias by unlabeled data. Advances in neural information processing systems , 19

    2006

  18. [26]

    Huber, J., M \"u ller, S., Fleischmann, M., and Stuckenschmidt, H. (2019). A data-driven newsvendor problem: From data to decision. European Journal of Operational Research , 278(3):904--915

    2019

  19. [27]

    and Cox, B

    Johnstone, C. and Cox, B. (2021). Conformal uncertainty sets for robust optimization. In Conformal and Probabilistic Prediction and Applications , pages 72--90. PMLR

    2021

  20. [28]

    and Mao, X

    Kallus, N. and Mao, X. (2023). Stochastic optimization forests. Management Science , 69(4):1975--1994

    2023

  21. [29]

    Kannan, R., Bayraksan, G., and Luedtke, J. (2021). Heteroscedasticity-aware residuals-based contextual stochastic optimization. arXiv preprint arXiv:2101.03139

    2021 arXiv

  22. [30]

    Kannan, R., Bayraksan, G., and Luedtke, J. R. (2024). Residuals-based distributionally robust optimization with covariate information. Mathematical Programming , 207(1):369--425

    2024

  23. [31]

    Kannan, R., Bayraksan, G., and Luedtke, J. R. (2025). Data-driven sample average approximation with covariate information. Operations Research

    2025

  24. [32]

    Lin, S., Chen, Y., Li, Y., and Shen, Z.-J. M. (2022). Data-driven newsvendor problems regularized by a profit risk constraint. Production and Operations Management , 31(4):1630--1644

    2022

  25. [33]

    Liyanage, L. H. and Shanthikumar, J. G. (2005). A practical inventory control policy using operational statistics. Operations research letters , 33(4):341--348

    2005

  26. [34]

    G., Tang, Q., and Xiao, Y

    Loke, G. G., Tang, Q., and Xiao, Y. (2022). Decision-driven regularization: A blended model for predict-then-optimize. Available at SSRN 3623006

    2022

  27. [35]

    Mandi, J., Kotary, J., Berden, S., Mulamba, M., Bucarey, V., Guns, T., and Fioretto, F. (2024). Decision-focused learning: Foundations, state of the art, benchmark and future opportunities. Journal of Artificial Intelligence Research , 80:1623--1701

    2024

  28. [36]

    W., and Heaton, H

    McKenzie, D., Fung, S. W., and Heaton, H. (2023). Faster predict-and-optimize with three-operator splitting. arXiv preprint arXiv:2301.13395

    2023 arXiv

  29. [37]

    Mulamba, M., Mandi, J., Diligenti, M., Lombardi, M., Bucarey, V., Guns, T., et al. (2021). Contrastive losses and solution caching for predict-and-optimize. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence , pages 2833--2840. ijcai. org

    2021

  30. [38]

    A., Zhang, F., Wang, S., Blanchet, J., Delage, E., and Ye, Y

    Nguyen, V. A., Zhang, F., Wang, S., Blanchet, J., Delage, E., and Ye, Y. (2024). Robustifying conditional portfolio decisions via optimal transport. Operations Research

    2024

  31. [39]

    Notz, P. M. and Pibernik, R. (2022). Prescriptive analytics for flexible capacity management. Management Science , 68(3):1756--1775

    2022

  32. [40]

    Ohmori, S. (2021). A predictive prescription using minimum volume k-nearest neighbor enclosing ellipsoid and robust optimization. Mathematics , 9(2):119

    2021

  33. [41]

    V., and Tak \'a c , M

    Oroojlooyjadid, A., Snyder, L. V., and Tak \'a c , M. (2020). Applying deep learning to the newsvendor problem. IISE Transactions , 52(4):444--463

    2020

  34. [42]

    Patel, Y., Rayan, S., and Tewari, A. (2024a). Conformal robust control of linear systems. arXiv preprint arXiv:2405.16250

    2024

  35. [43]

    P., Rayan, S., and Tewari, A

    Patel, Y. P., Rayan, S., and Tewari, A. (2024b). Conformal contextual robust optimization. In International Conference on Artificial Intelligence and Statistics , pages 2485--2493. PMLR

    2024

  36. [44]

    and Pagnoncelli, B

    Rahimian, H. and Pagnoncelli, B. (2023). Data-driven approximation of contextual chance-constrained stochastic programs. SIAM Journal on Optimization , 33(3):2248--2274

    2023

  37. [45]

    Romano, Y., Patterson, E., and Candes, E. (2019). Conformalized quantile regression. Advances in neural information processing systems , 32

    2019

  38. [46]

    Rychener, Y., Kuhn, D., and Sutter, T. (2023). End-to-end learning for stochastic optimization: A bayesian perspective. In International Conference on Machine Learning , pages 29455--29472. PMLR

    2023

  39. [47]

    Sadana, U., Chenreddy, A., Delage, E., Forel, A., Frejinger, E., and Vidal, T. (2024). A survey of contextual optimization methods for decision-making under uncertainty. European Journal of Operational Research

    2024

  40. [48]

    Stanton, S., Maddox, W., and Wilson, A. G. (2023). Bayesian optimization with conformal prediction sets. In International Conference on Artificial Intelligence and Statistics , pages 959--986. PMLR

    2023

  41. [49]

    Sun, C., Liu, L., and Li, X. (2023). Predict-then-calibrate: A new perspective of robust contextual lp. Advances in Neural Information Processing Systems , 36:17713--17741

    2023

  42. [50]

    and Khalil, E

    Tang, B. and Khalil, E. B. (2024). Pyepo: A pytorch-based end-to-end predict-then-optimize library for linear and integer programming. Mathematical Programming Computation , pages 1--39

    2024

  43. [51]

    Wang, I., Becker, C., Van Parys, B., and Stellato, B. (2023). Learning decision-focused uncertainty sets in robust optimization. arXiv preprint arXiv:2305.19225

    2023

  44. [52]

    Yeh, C., Christianson, N., Wu, A., Wierman, A., and Yue, Y. (2024). End-to-end conformal calibration for optimization under uncertainty. arXiv preprint arXiv:2409.20534

    2024

  45. [53]

    Zhang, Y., Liu, J., and Zhao, X. (2023). Data-driven piecewise affine decision rules for stochastic programming with covariate information. arXiv preprint arXiv:2304.13646

    2023

Pith tools

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