acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Scenario Generation for KL-DRO with a Learned Density Model

Abstract

Distributionally robust optimization (DRO) hedges data-driven two-stage stochastic programs against error in the distribution estimated from a finite record. Standard ambiguity sets treat realizations absent from the record in one of two ways. A divergence ball around the empirical distribution contains only reweightings of the record. A transport ball also contains distributions supported off the record, including on realizations the data-generating process cannot produce, against which the decision is then hedged. We propose a Kullback–Leibler (KL) ball centered at a density model fitted to the record, the learned nominal, so that the adversary can up-weight costly unobserved realizations in proportion to the probability the model assigns them. Because an independent sample of the few scenarios the problem can accommodate is likely to omit the rare realizations on which the worst case concentrates, we build the scenario set adaptively. We alternate between solving the finite problem and drawing new realizations from the learned nominal, steered toward the worst case at the current decision, with importance weights that keep the ball centered at the model. On a multi-item newsvendor problem and a capacity expansion problem on realistic grid data, centering the ball at the learned nominal rather than at the record lowers cost under distribution shift, at comparable in-distribution cost, when the record is scarce and the model class captures structure that lets it extrapolate beyond the record. The advantage fades as the record grows or as the shift departs from the structure the model has learned.

Primary area:
optimization
Keywords:
distributionally robust optimization, stochastic programming, generative models, importance sampling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WGmiOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors consider optimization problems where a decision xx must be made, a random variable ξ\xi is realized (drawn from a distribution PP in an ambiguity set), and then an action yy can be taken in response to xixi. The authors consider the ambiguity set to be those that are close in KL divergence to a nominal distribution QθQ_\theta. Connections/interpretations are drawn to when the QθQ_\theta is learned from a generative model, such as a diffusion model or normalizing flow. The authors motivate the use of the KL divergence by viewing it as requiring absolute continuity of PP relative to QQ (as PP's support must be a subset of QQ's), but allowing reweighting of elements in the support. They provide a sample, tilt, resample algorithm to deploy this idea, and demonstrate its effectiveness in a multi-item newsvendor problem and a power capacity expansion problem.

Critical strengths
  • The paper considers a very interesting problem where a 2-stage stochastic optimization takes place: one decision occurs before the realization of noise and the other decision takes place after. It is cool to consider how these ideas can connect with generative models providing the nominal distribution or center of ambiguity sets.
Critical weaknesses
  • The paper lacks a precise technical formulation of many of its claims. For example, a 'record', which intuitively is a subset of a larger dataset currently used for some type of learning, is never formally defined. Additionally, how the generative model is mathematically represented in the problem is never presented, and, as such, much of the claims of the paper are more 'intuitive' rather than formal.
  • Building on this, the paper has no theoretical contribution; the numerical results are of a caliber which would be good for validating a theoretical contribution, but not strong enough if the entire contribution of the paper is empirical, in my opinion.
  • The paper itself is a bit difficult to read in its organization. The discussion of generative models and scenarios feels very disconnected from Equation (1), and even the numerical experiments are primarily informal descriptions of the problem, rather than very explicit statements of the form "in this problem, Q(⋅)Q(\cdot) is this function". One unified problem formulation would be nice, as right now the notion of a KL divergence ambiguity set, the sequential aspect of the two-stage problem, the relationship to normalizing flows and generative models, the notion of scenario, are very disjoint and not connected, making the overall focus of this paper very unclear.
Questions for authors
  • Is there any theoretical insight to be gained from using the KL divergence for ambiguity sets? How does this part of the formulation differ from (Hu and Hong 2013)?
  • What mathematically changes about the problem when QθQ_\theta is generated from a diffusion process as opposed to the empirical distribution, or any other learning method?
  • Is there a way to formalize the notion of record and scenario more explicitly? How does the latent layer actually affect the optimization problem?
  • How does the sequential nature of the optimization problem (before and after the realization of uncertainty) relate to the discussion in the later parts of the paper? I had trouble seeing the connection.
  • Can you share more details regarding the computation time for the numerical results? Can this handle 'real world' scale problems?
4 comments

We thank the reviewer for their insights. The three identified weaknesses are closely related, and we plan to address them in the revision. We will prioritize more clearly defining the record, the learned nominal Qθ\mathbb Q_\theta, the KL ball and its dual, the finite master problem, and the latent problem, and we will state the decision xx, the uncertainty ξ\xi, and the recourse cost QQ explicitly at the start of each experiment. We agree that our contribution is primarily methodological and empirical, and we do not claim new theory for the KL-DRO problem itself, but the revision will state as formal results the equivalence of the latent and original problems and the condition, adopted from (Hu and Hong 2013), under which the problem centered at a learned nominal is well posed (https://openreview.net/forum?id=nNV05EIQf1&noteId=QFwiWI7cqX). We have also strengthened the empirical evidence with a held-out evaluation under pre-specified shifts (https://openreview.net/forum?id=nNV05EIQf1&noteId=eVFUivqQ8s) and a bounded-demand version of the newsvendor study (https://openreview.net/forum?id=nNV05EIQf1&noteId=mNHJQe1DTL). We address the reviewer's questions below.

Q1: Is there any theoretical insight to be gained from using the KL divergence for ambiguity sets? How does this part of the formulation differ from (Hu and Hong 2013)?

We do not claim new theoretical insight from the KL divergence itself, but we do rely on two of its standard properties. Every member of the KL ambiguity set assigns zero probability wherever its nominal does, so the nominal determines which realizations the adversary can reach, and the dual of (Hu and Hong 2013) turns the worst case over the ball into a single convex program on any finite scenario set (5). Our formulation otherwise follows (Hu and Hong 2013), except that we use a mean–CVaR objective instead of the expectation (which has many precedents and is not a contribution of ours) and center the ball at a learned nominal Qθ\mathbb Q_\theta rather than at the empirical distribution. This definition of the nominal distribution/center of the ambiguity set is what lets the adversary reach realizations beyond the record, and it is also what requires the finite scenario approximation that our sampler builds.

Q2: What mathematically changes about the problem when QθQ_\theta is generated from a diffusion process as opposed to the empirical distribution, or any other learning method?

Centered at the empirical distribution Q^N\hat{\mathbb Q}_N, the KL ball only reweights the NN recorded realizations, and the dual (4) of (Hu and Hong 2013) yields a finite convex program (5). Centered at a continuous Qθ\mathbb Q_\theta, the ball also contains distributions that place mass on realizations absent from the record, in proportion to the probability Qθ\mathbb Q_\theta assigns them, so the learned model determines what the adversary can reach. Since the expectation in (4) is then no longer a finite sum, problem (1) must in general be approximated with a finite scenario set, which is what our framework builds. The learning method matters only through Qθ\mathbb Q_\theta, so diffusion models with deterministic samplers, normalizing flows, and copulas are all compatible with the proposed framework.

Q3: Is there a way to formalize the notion of record and scenario more explicitly? How does the latent layer actually affect the optimization problem?

What we call the "record" is the finite sample of past realizations available to the decision-maker, ξ1,…,ξN\xi^1, \dots, \xi^N, with empirical distribution Q^N\hat{\mathbb Q}_N. The record is the only data used to fit Qθ\mathbb Q_\theta, and it is the nominal of the record-based baselines. Its realizations are not necessarily independent or identically distributed (they may exhibit seasonality and drift over time, as in the capacity expansion dataset), and none of our methods or statements rely on independence. A "scenario" is an element of the finite set on which the master (5) is solved, a realization ξi\xi_i with weight wiw_i, drawn from the record or from Qθ\mathbb Q_\theta depending on the method. We will make both definitions explicit in the revised manuscript.

The latent layer does not change the optimization problem we solve. For any decoder, the latent problem (7) and the original problem (1) have the same value and minimizers, by the argument in Sec. 4.2, which we will state as a formal theoretical result in the revision. The latent layer only changes how scenarios are drawn and weighted, since importance weights computed in the latent space are exact and in closed form.

Q4: How does the sequential nature of the optimization problem (before and after the realization of uncertainty) relate to the discussion in the later parts of the paper? I had trouble seeing the connection.

The two-stage structure is fundamentally what makes the number of scenarios the computational bottleneck. Because the first-stage decision xx is fixed before ξ\xi is realized, each scenario enters the master as its own copy of the second-stage problem, so the cost of solving the master grows quickly with the number of scenarios and only a few can be held in practice. Which scenarios are included largely determines xx, which is why we generate them adaptively at the current decision through the tilt. The same structure defines the evaluation under distribution shifts, where the plan xx is fixed from the record and the second-stage problem is re-solved on each day of the shifted distribution, so robustness is needed for the first-stage decision which cannot adapt to the realized ξ\xi.

Q5: Can you share more details regarding the computation time for the numerical results? Can this handle 'real world' scale problems?

Regarding computation time, a full run on the paper's capacity expansion model (20 zones, 1,440-dimensional daily scenarios) takes about 10–13 minutes on 8 CPU threads, of which 75–90% is spent re-solving the master in each of the 12 rounds as its scenario set grows to 48, while generating scenarios adds negligible time (KL-Gen takes 9.7–10.9 minutes against 10.4–10.7 for KL-Rec). Solving a single master on every record day instead (i.e., SAA over the full record) takes 15.5 minutes with half a year of record, 51 minutes with one year, and 6.1 hours with five years. KL-Gen's plans cost less than this SAA on 2020 at every record length, by 2.7–17.8 bn$/yr, although the SAA is cheaper in distribution with one or five years of record.

Regarding scale, the newsvendor study is a deliberately simple synthetic problem, which lets us know the true distributions and run each configuration on 300 records. The capacity expansion model, however, uses hourly demand and wind and solar capacity factors simulated from historical weather at a zonal resolution comparable to planning studies, and as such we consider this problem to be close to "real world" scale. To assess whether the approach extends to an even larger problem, we also considered a 214-bus version of the same system with 72-hour scenarios (46,224 dimensions), whose results we did not include because the full study was not complete in time for the manuscript. On this model, a run takes 3.2–3.6 hours, of which 91–92% is again the master, so the method scales as the scenario-based baselines do.

I appreciate your reply. Thank you. My review remains unchanged: in the absence of a strong theoretical contribution, the empirical portion of the results should be very comprehensive.

Thank you for the reply. We agree that a paper without a strong theoretical contribution needs a comprehensive empirical evaluation, and we acknowledge that our paper belongs in that category. With that in mind, we would like to explain why we believe the empirical part meets the bar you describe, particularly after the additional work completed during the rebuttal phase.

The original manuscript included two studies.

  • A newsvendor study with a known distribution, four controlled shifts, 300 records at each record size and five learned nominals, compared against SAA, KL-Rec, W-DRO, KL-OT and CSSC.
  • A capacity expansion study of the Texas system built on twenty years of hourly demand, wind, and solar data, considering a range of record lengths, compared against the same baselines.

During the rebuttal phase, we added the following analyses.

  • All 6,900 capacity expansion planning solutions were fixed and re-evaluated on five held-out years and pre-specified shifts.
  • The whole newsvendor study was re-run with bounded demand (for well-posedness), which did not change any verdict.
  • We ran ablations on the number of rounds, the initialization, and the risk measure.
  • We ran two convergence studies, which are reported in our reply to Reviewer KDsN.

Regarding your question about real-world scale, we also ran the method on a 214-county model with 72-hour scenarios and unit commitment, where a run takes 3.2 to 3.6 hours and the main comparison continues to hold, with KL-Gen attaining a lower out-of-distribution cost than KL-Rec in all 10 records (one-sided p=.001p = .001).

We also view the capacity expansion study as a contribution in its own right, since methods of this kind are often evaluated only on small synthetic problems, and we believe that a test on a planning problem of realistic size with weather-driven data and a held-out evaluation provides useful evidence for readers who would consider using such a method. In the same spirit, we have tried to honestly report negative results, such as the narrower advantage over KL-OT in the held-out evaluation.

We understand that you may weigh this evidence differently, but if it changes your view of the empirical part, we would be grateful if you would consider reflecting that in your evaluation.

Reviewer_KDsNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies two-stage stochastic programs with a mean-CVaR objective when the historical record is short and the optimization model can hold only a limited number of scenarios. The authors fit a generative nominal distribution, place a forward-KL ambiguity set around it, and build a weighted scenario set over several rounds. Each round solves the current master problem and selects new scenarios using an exponential tilt based on recourse cost at the current decision. The paper considers both resampling and a Gaussian proposal fitted in the generator's latent space, with importance weights to correct for the proposal.

The experiments cover a five-product newsvendor problem with four controlled shifts and a 20-zone Texas capacity-expansion model with 1,440-dimensional daily scenarios. The main finding is that generated scenarios can help with short records when the nominal captures dependence patterns relevant to costly events. The advantage is smaller for longer records and for shifts the fitted model does not anticipate.

Critical strengths
  1. The approach separates a learned nominal from the finite set used by the solver. I like that the generator is kept fixed while the sampling proposal reacts to the current decision. For direct draws from the defensive latent-space proposal, the likelihood ratio is available without evaluating the density of the decoded scenario, and the proposal retains coverage of the Gaussian latent prior. This is a sensible design for a two-stage model where each added scenario is expensive.

  2. The controlled newsvendor study is well chosen. It varies record size and the form of the distribution shift, uses 300 records per size, and selects settings without consulting the shifted evaluation law. The capacity-expansion study also includes useful checks on scenario severity and dependence structure, although its OOD scoring protocol raises a separate concern below.

Critical weaknesses

1. The continuous positive-radius KL formulation appears to be infinite for several central newsvendor nominals.

This follows directly from the recourse model. Write D=∑jξjD=\sum_j \xi_j, X=∑ixiX=\sum_i x_i, and let UU, VV, and SsubS_{sub} be total unmet demand, unused stock, and substitution flow. Summing the balance constraints in Equation (11) gives U−V=D−XU-V=D-X. Hence

Q(x,ξ)=100(D−X)+0.5Ssub+99.75V≥100(D−X).Q(x,\xi)=100(D-X)+0.5 S_{sub}+99.75 V \geq 100(D-X).

The base law is a Gaussian mixture in log demand, and several fitted nominals are log-normal or log-normal mixtures. Their demand marginals have infinite positive exponential moments. Since ψζ(u)\psi_\zeta(u) grows with a positive linear slope for large uu, it follows for every finite xx, ζ\zeta, and λ>0\lambda>0 that

EQθ[exp⁡(ψζ(Q(x,ξ))/λ)]=∞.\mathbb{E}_{Q_\theta}\left[\exp\left(\psi_\zeta(Q(x,\xi))/\lambda\right)\right]=\infty.

The issue is not merely failure of a convenient dual expression: the worst-case expectation over a positive-radius forward-KL ball is infinite for such right-heavy-tailed losses (Hu and Hong, 2013). Here the mean-CVaR risk is at least the mean, so the same problem carries over to the stated objective. It applies to the log-normal, log-normal mixture, true-law oracle, and the ordinary flow at N=16N=16, which the appendix says remains log-normal. I cannot establish the same claim for every trained flow at larger NN from the information given.

There is a related problem with Equation (8): for these nominals, its proposed population exponential tilt cannot be normalized, even if the finite master uses ρ=0\rho=0. The empirical tilt over a finite pool is well-defined, but it is not a sample from a finite-valued population KL adversary. Also, if the log-t nominal illustrated in Figure 4 is untruncated, its demand has infinite first moment, which would affect the nominal objective even at ρ=0\rho=0.

I do not dispute that the finite-support master can be solved. My concern is the connection to the continuous KL-DRO problem the paper claims to approximate. This requires a tail assumption, an explicit truncation, or a different formulation.

2. The capacity-expansion OOD year was also used for method development.

Appendix B.6 says that the KL radii, β\beta, marginal transforms and smoothing, and base-weight construction were chosen after comparisons on both ID and OOD days. The same 2020 days are then used to report OOD performance. The authors state that the flow architecture and its training procedure were fixed before scoring plans, which limits the scope of this concern. Still, the 2020 results are not an independent test of the completed method.

This matters because the capacity-expansion experiment is the main high-dimensional application. The reported advantage on the extreme year may be real, but the evaluation cannot tell us how much of it would remain on a year that played no role in development.

Questions for authors
  1. What are the exponential-integrability assumptions behind Equations (4) and (8), and which newsvendor nominals satisfy them? Are any demand distributions explicitly truncated in the implementation? Please also report the selected ρ\rho values for the main newsvendor results. How should the log-t configurations be interpreted if they have no finite nominal mean?

  2. Can the newsvendor results be reproduced with explicitly bounded demand support or an ambiguity set that gives a finite population objective for log-normal losses? Please also show how the decision and finite-master value change as the number of admitted scenarios grows. The answer would determine whether the headline improvements survive a well-defined population formulation.

  3. Can the capacity-expansion comparison be repeated with the full method frozen before any scores on the test period are examined? An untouched year or a pre-specified set of synthetic shifts would be useful. Does the KL-Gen advantage over KL-OT and W-DRO persist under that protocol?

7 comments
AuthorsOpenReview ↗

Q1: Well-posedness of the KL-DRO problem (Part 1)

We first thank the reviewer for raising this concern. This was an oversight on our part, and we appreciate the reviewer's help in catching it.

What are the exponential-integrability assumptions behind Equations (4) and (8)?

For ρ>0\rho > 0, the KL-DRO worst case is finite if and only if the cost has a finite exponential moment under the nominal, i.e., if and only if EQθ[eQ(x,ξ)/λ]<∞\mathbb{E}_{\mathbb Q_\theta} [e^{Q(x, \xi)/\lambda}] < \infty for some λ>0\lambda > 0. This corresponds to Assumption 1 in (Hu and Hong, 2013; https://optimization-online.org/wp-content/uploads/2012/11/3677.pdf) for expected value problems, which can be easily extended to the class of mean-CVaR problems we consider via the reformulation RP[Q(x,ξ)]=min⁡ζ∈R{ζ+EP[ψζ(Q(x,ξ))]}\mathcal R_{\mathbb P}[Q(x,\xi)] = \min_{\zeta\in\mathbb R} \lbrace \zeta + \mathbb E_{\mathbb P}[\psi_\zeta(Q(x,\xi))] \rbrace, where ψζ(u)=(1−β)(u−ζ)+β1−α(u−ζ)+\psi_\zeta(u) = (1-\beta)(u-\zeta) + \frac{\beta}{1-\alpha}(u-\zeta)_+ (equation 3 in our manuscript). Specifically, for any ζ\zeta, we have

sup⁡KL(P∣Qθ)≤ρRP[Q]≤ζ+sup⁡KL(P∣Qθ)≤ρEP[ψζ(Q)].\sup_{\mathrm{KL}(\mathbb P \\| \mathbb Q_\theta) \le \rho} \mathcal R_{\mathbb P}[Q] \le \zeta + \sup_{\mathrm{KL}(\mathbb P \\| \mathbb Q_\theta) \le \rho} \mathbb E_{\mathbb P}[\psi_\zeta(Q)].

The right side is the KL worst-case expected value of the loss ψζ(Q)\psi_\zeta(Q), which is finite if ψζ(Q)\psi_\zeta(Q) satisfies the exponential moment condition (Hu and Hong, 2013, Theorems 1 and 2); minimizing over ζ\zeta and dualizing gives equation 4 in the manuscript. Since ψζ(u)≤L(u−ζ)+\psi_\zeta(u) \leq L (u - \zeta)_+ with L:=1−β+β/(1−α)L := 1 - \beta + \beta/(1-\alpha), one can take λ′=Lλ\lambda' = L \lambda (with λ>0\lambda > 0 chosen to satisfy the exponential moment condition for the expected value case) and see that the "if" (i.e.,   ⟸  \impliedby) condition also holds for the mean-CVaR case:

EQθ[eψζ(Q)/λ′]≤EQθ[e(Q−ζ)+/λ]≤1+e−ζ/λ,EQθ[eQ/λ]<∞.\mathbb{E}_{\mathbb Q_\theta} [e^{\psi_\zeta(Q)/\lambda'}] \le \mathbb{E}_{\mathbb Q_\theta} [e^{(Q-\zeta)_+/\lambda}] \le 1 + e^{-\zeta/\lambda} \\, \mathbb{E}_{\mathbb Q_\theta} [e^{Q/\lambda}] < \infty.

Conversely, since ψζ(u)≥u−ζ\psi_\zeta(u) \ge u - \zeta, the mean-CVaR objective is always lower bounded by the expected value: RP[Q]≥EP[Q]\mathcal R_{\mathbb P}[Q] \ge \mathbb E_{\mathbb P}[Q], so whenever the condition fails, both the worst-case expected value and worst-case mean-CVaR are ∞\infty for any ρ>0\rho > 0, extending the "only if" (i.e.,   ⟹  \implies) condition from (Hu and Hong, 2013).

The tilt (equation 8) is also normalized under the same condition as its normalizing constant is ∫φ(z)eQ(x,Gθ(z))/λdz=EQθ[eQ(x,ξ)/λ]\int\varphi(z) e^{Q(x,G_\theta(z))/\lambda}dz=\mathbb E_{\mathbb Q_\theta}[e^{Q(x,\xi)/\lambda}].

Which newsvendor nominals satisfy them?

Beside the record-based baselines KL-Rec and KL-OT, whose nominals are finite, the copula nominal satisfies these conditions (by boundedness) as well as the Gaussian nominal, which has a finite exponential moment. These conditions are not satisfied by the log-normal and log-normal mixture nominals, the true-law oracle, or the flow variants, which model log-demand. This is the case for every trained flow, not only at N=16N = 16 where it remains log-normal. Specifically, the flow's coupling layers have bounded log-scales and shifts, so log-demand has at least Gaussian-type tails, and demand keeps a log-normal-type tail with no exponential moment. As the reviewer correctly identified, these nominals can be applied in our framework and evaluated in computational experiments, but they do not admit a well-posed problem at any ρ>0\rho > 0, which the revision will state explicitly.

AuthorsOpenReview ↗

Q1: Well-posedness of the KL-DRO problem (Part 2)

Are any demand distributions explicitly truncated in the implementation?

Demand was not truncated in the experiments for the original manuscript except for the heavy-tailed flow's decoder, whose draws were bounded at 100 times the record's per-product maximum (as a numerical guard). We have since re-run all newsvendor experiments in the paper with an explicit demand cap (which we will present in our response to Q2).

Please also report the selected ρ\rho values for the main newsvendor results.

The selected ρ\rho values for Table 1(a)'s subset of results are given below (note that ρ=0.8\rho=0.8 was the largest radius in the ρ\rho grid). Here, we also include the selected transport displacement ε\varepsilon where relevant, and entries split in the form "a / b" give the selected hyperparameters on the two halves of the records where they differ (see Appendix A.6). A †^\dagger symbol marks entries where no setting met the in-distribution budget, meaning the cheapest setting was used, following the specified selection rule (A.6). Note that the level, tail, and support shifts are merged as they share the same records (and hence selection) while the coincidence shift uses a different record (Appendix A.2).

Method (nominal)N=16N = 16, L/T/SN=16N = 16, CN=128N = 128, L/T/SN=128N = 128, C
KL-Rec0.80.80.80.80.80.80.80.8
W-DROε=10\varepsilon = 10ε=3/10\varepsilon = 3 / 10ε=0.1\varepsilon = 0.1ε=0.03\varepsilon = 0.03
KL-OT0.80.8 (ε=10\varepsilon = 10)0.80.8 (ε=3/10\varepsilon = 3 / 10)0.80.8 (ε=0.1/0.3\varepsilon = 0.1 / 0.3)0.80.8 (ε=0.03\varepsilon = 0.03)
KL-Gen-P (log-normal)0.80.80.80.80.80.80.015/0.020.015 / 0.02
KL-Gen-P (flow)0.80.80.80.80.03/0.80.03 / 0.80.01/0.0150.01 / 0.015
KL-Gen-P (Gaussian)0.80.80.80.80.05†0.05^\dagger0.05†0.05^\dagger
KL-Gen (flow)0.80.80.80.80.003†0.003^\dagger0.003†/0.007†0.003^\dagger / 0.007^\dagger

How should the log-t configurations be interpreted if they have no finite nominal mean?

The reviewer is correct that the log-t nominal has no finite mean, so the KL-DRO problem centered at it is not well-posed even at ρ=0\rho = 0. The data distributions, however, are log-normal mixtures with a finite mean-CVaR cost for any decision. Our computational study aims to answer whether solving the finite-sample KL-DRO master with the generator as the nominal yields good decisions under the true data distribution, before and after the shift. Since every decision is scored under the true distributions, the reported costs do not depend on whether the population problem behind the master is well-posed.

AuthorsOpenReview ↗

Q2: Newsvendor results with bounded support

Can the newsvendor results be reproduced with explicitly bounded demand support or an ambiguity set that gives a finite population objective for log-normal losses?

Following the reviewer's suggestion, we re-ran the newsvendor experiments with demand clipped at 2,000 (20 times the median demand of 100), both for the true distribution and the learned density models. Here, clipping refers to adding a point mass at 2,000 (as opposed to truncation, which would discard samples exceeding the demand bound). At a bound of 2,000, fewer than 0.00001% of scenarios are clipped under the base distribution as well as the level, support, and coincidence shifts, and 0.19% of scenarios are clipped under the tail shift. The following tables recreate Table 1 from the manuscript for the new experiments:

(a) Out-of-distribution cost at each method's selection rule (% above the optimum, median over 300 records)

MethodN=16N = 16, LTSCN=128N = 128, LTSC
SAA6431814342441291049839
KL-Rec64718243522714110511139
W-DRO61407512124533018
KL-OT67428111844674425
KL-Gen-P (log-normal)5261473135653515
KL-Gen-P (flow)5261433341683518
KL-Gen-P (Gaussian)2291221584925012617451
KL-Gen (flow)2291211524920611913846

(b) In-distribution cost needed to halve SAA's out-of-distribution cost (·: never reached; †^\dagger: lowest in under 80% of bootstrap resamples)

MethodN=16N = 16, LTSCN=128N = 128, LTSC
W-DRO5667569712421220
KL-OT57645797123012·
KL-Gen-P (log-normal)212322249†^\dagger·9†^\dagger15
KL-Gen-P (flow)18181823†^\dagger11·920
KL-Gen-P (Gaussian)40·4042····
KL-Gen (flow)36·3633····

These tables demonstrate that the headline results hold under the well-defined variant. In particular, costs do not move by much for the proposed methods (KL-Gen, KL-Gen-P) under the demand cap, except in the tail-shifted case (the only shift with non-negligible mass above the cap) where they rise. Interestingly, the demand cap does improve the baseline KL-OT and W-DRO results under the level, support, and coincidence shifts for N=16N=16 (possibly by reducing over-hedging), which narrows the margin by which KL-Gen-P improves over these transport baselines in some entries.

Please also show how the decision and finite-master value change as the number of admitted scenarios grows.

The following table presents the finite-master problem objective values relative to the population-level KL-DRO value for the capped nominal (20x cap, medians over 100 records, entries range over N∈{16,128}N \in \lbrace 16, 128 \rbrace and the two record families). Note that the manuscript uses n=K=100n = K = 100.

Methodρ\rhon=25n = 25n=50n = 50n=100n = 100n=200n = 200n=400n = 400
KL-Gen-P (log-normal)0.010.64–0.690.75–0.810.83–0.880.88–0.950.94–1.01
KL-Gen-P (log-normal)0.050.39–0.430.48–0.570.61–0.760.74–0.850.84–0.99
KL-Gen-P (log-normal)0.20.17–0.200.23–0.290.33–0.500.47–0.690.63–0.96
KL-Gen-P (flow)0.010.59–0.740.67–0.800.75–0.890.83–0.930.89–1.01
KL-Gen-P (flow)0.050.34–0.510.42–0.570.51–0.730.64–0.850.69–1.00
KL-Gen-P (flow)0.20.16–0.260.20–0.300.25–0.460.37–0.630.48–1.00
KL-Gen (log-normal, plain)0.010.49–0.540.50–0.550.54–0.560.57–0.600.61–0.64
KL-Gen (log-normal, plain)0.050.26–0.280.26–0.290.28–0.310.30–0.330.33–0.36
KL-Gen (log-normal, plain)0.20.12–0.130.11–0.120.12–0.130.13–0.140.14–0.16

With the proposal (KL-Gen-P), the master value approaches the population value as nn grows (here, values above 1 reflect Monte Carlo error in the population estimate), while with plain resampling (KL-Gen) it converges much more slowly. Importantly, the gap widens with ρ\rho, and at the manuscript's fixed n=100n = 100, decisions remain 0.23–0.51 (relative L2 distance) from those at n=400n = 400. At ρ=0.8\rho = 0.8, the radius the selection rule picks at N=16N = 16, two independent estimates of the population value from 10510^5 draws each differ substantially, but the master value at n=100n = 100 is at most about a quarter of either, so the reported decisions are best read as solutions of the finite-sample KL-DRO problem rather than approximations of the population KL-DRO solution at ρ=0.8\rho=0.8. Their quality under the true distribution is what Table 1 measures, and it does not depend on this distinction.

Reviewer_Ay1SOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for scenario generation in two-stage distributionally robust optimization (DRO). Specifically, the method samples from the nominal distribution, tilts it (re-weighting more costly scenarios), and fits a generative model that can then be resampled from, either iteratively or in one shot. Empirically, the authors compare in- and out-of-distribution solution quality metrics, and their approach consistently achieves a strong balance between the two.

Critical strengths
  • Clarity, Presentation, & Motivation: I do not work in the area, and I found the paper relatively easy to follow. The authors do a reasonable job of motivating each step and why. This is consequential; I can actually walk away with an understanding of the problem, motivation, and justification for the approach. I will caveat again that I am relatively unfamiliar with DRO, so I am mostly taking the authors' discussions/motivation/justifications at face value.
  • Empirical Results: The authors study two benchmarks and report relatively favorable results in terms of out-of-distribution and in-distribution performance. This is consequential because it empirically validates the method, which is especially important given its lack of the theoretical guarantees many existing DRO approaches have.
Critical weaknesses
  • Experimental breadth: Overall, the experiments are reasonable, but there are a few areas that could be strengthened. Performance over iterations. The authors propose an iterative algorithm and run it for 1 iteration in the first benchmark and 12 in the second, as I understand it (though the paper doesn't make this entirely clear). However, it would be useful to better understand performance as the number of iterations, TT, varies. Risk Measure. The computation is limited to a single risk measure with fixed parameters, so an ablation on this would be useful. Initialization. As I'll discuss below, the choice of initial decision may matter, and the authors use the risk-neutral problem to address this. It would be much more useful to better understand the impact of these decisions, e.g., the proposed method compared to random initializations, initializations coming from prior methods, etc. These are consequential, in my view, because the algorithm lacks strong theoretical guarantees, and more rigorous computation would help alleviate concerns about how well the approach works beyond the existing setting.

  • The decision(s) (xt)(x_t): The one algorithmic component I have a concern with is the decisions themselves. In Algorithm 1, the decisions are used to derive the tilt. However, the tilt itself is then a function of the decisions, and it may not be a reasonable tilt for the optimal decision, or any other decision for that matter. My concern is that the iterative approach may converge to poor local optima or fail to converge when second-stage costs vary across different decisions. This matters because it may be an algorithmic limitation the authors haven't fully considered and may limit the method's broader applicability to more complex decision-making problems.

Questions for authors
  • If the authors could elaborate on my last concern and if it is consequential (and specifically why it is not, either empirically or theoretically), that would be helpful.
6 comments
AuthorsOpenReview ↗

Performance over iterations

The reviewer's read is correct: we chose a single round (T=1T=1) in the newsvendor study to keep each run cheap enough to cover 300 records per configuration across all nominals, shifts, and radii, while the choice of T=12T=12 rounds for the capacity expansion study followed from a decision to impose a 48-scenario budget, with four scenarios admitted per round. We will make both settings explicit in the main text. In the following table, we report median values over the 50 experimental runs (10 records × 5 seeds) of the mean-CVaR planning cost for a representative subset of rounds tt (all results are for k=1/2k = 1/2, ρref=.2\rho_{\mathrm{ref}} = .2, ρ=.05\rho = .05), as well as the L1 distance in generation capacity (in GW) between the incumbent plan at round tt and the final plan at round TT.

Round tt124681012
In distributionKL-Gen303.972.036.231.331.030.930.9
KL-Rec225.694.445.435.434.333.132.7
KL-OT225.568.733.831.130.830.630.5
Out of distribution (2020)KL-Gen416.9146.960.839.234.633.533.4
KL-Rec334.3174.987.262.459.354.452.8
KL-OT334.3137.349.038.937.436.736.7
L1 distance to final planKL-Gen1491241037859430

We also re-ran KL-Gen in the same setting up to T=24T=24, which lowered in-distribution costs by 0.12±0.050.12 \pm 0.05 (median ±\pm standard error over records) and insignificantly lowered out-of-distribution costs by 0.30±0.440.30 \pm 0.44 relative to the T=12T=12 runs. In this case, the round-12 plan exhibited a median L1 distance of 59 GW from the final round-24 plan, while the round-23 plan exhibited a median distance of 30 GW. Altogether, these results demonstrate fast cost reductions for all methods in early iterations (particularly so for KL-Gen and KL-OT), with a plateau by round 6 in distribution and round 10 out of distribution for KL-Gen and KL-OT; KL-Rec appears to still be improving slowly at round 12. On the other hand, the plan itself continues to shift across iterations, which suggests the existence of many alternative plans that achieve similar costs.

AuthorsOpenReview ↗

Risk Measure

Below, we present results for a subset of the newsvendor experiments with β=0\beta=0 in the style of Table 1 in the manuscript. This changes the problem from having a worst-case mean-CVaR objective to a worst-case expected cost objective, which is commonly considered in DRO settings. As is the case in the manuscript's table, (a) describes the out-of-distribution cost at each method’s selection rule (see Appendix A.6) while (b) shows the in-distribution cost needed to halve SAA’s out-of-distribution cost. All entries denote costs above the optimum (%, median over 300 records), bold marks the lowest among the rows shown, and · means the target was never reached.

NNMethod(a) L(a) T(a) S(a) C(b) L(b) T(b) S(b) C
16SAA2423812864····
KL-Rec2493913163····
KL-OT6117604320322037
KL-Gen-P (log-normal)2810292610101012
64SAA65214315····
KL-Rec70214716····
KL-OT2312241391515·
KL-Gen-P (log-normal)141015661068

Two findings of the newsvendor study can be checked against these results and shown to hold in the β=0\beta=0 case, albeit with different margins. First, KL-Gen-P halves SAA's out-of-distribution cost at a lower in-distribution cost than KL-OT in every shift. For N=16N = 16 in particular, in-distribution costs need to be 10%-12% higher than the optimum versus 20%-37% for KL-OT to halve SAA's out-of-distribution cost under β=0\beta = 0, compared with 21%-24% versus 57%-97% under β=0.25\beta = 0.25. Second, under the coincidence shift, KL-Gen-P attains the lowest out-of-distribution cost: 26% versus 43% for KL-OT at N=16N = 16 and 6% versus 13% at N=64N = 64 under β=0\beta = 0, compared with 31% versus 159% and 21% versus 51% under β=0.25\beta = 0.25.

These results were obtained while producing the manuscript (i.e., they were already on-hand), and while the results partially demonstrate robustness of the approach to a different choice of β\beta, they do not constitute a full ablation as the reviewer's comment calls for, and in particular are missing results for high-performing baselines like W-DRO. Following the reviewer's suggestion, we have started re-running experiments needed to recreate Table 1 with β=0\beta=0 as a more thorough ablation, and will report the results when we have them, as well as include them in the revised version of the manuscript.

AuthorsOpenReview ↗

Impact of initialization

We agree with the reviewer's concern about the impact of initialization, and as such performed additional experiments for the capacity expansion case study to better understand robustness of decisions to initialization. Specifically, we ran experiments for KL-Gen with k=1/2k = 1/2, ρref=.2\rho_{\mathrm{ref}} = .2, ρ=.05\rho = .05 on 10 records x 3 seeds with four different initializations:

  1. The risk-neutral plan solved from one randomly drawn scenario (used in the manuscript's experiments, varies highly with the scenario)
  2. A plan obtained by SAA (which systematically under-builds peaking capacity)
  3. A plan obtained by KL-OT at δ=3\delta=3 (typically over-hedged)
  4. A plan obtained by solving KL-Rec

The following table presents the mean-CVaR cost of the final plan (median over 10 records, each the median over 3 seeds, ±\pm denoting standard error across records). The last two columns are paired differences from the default start on the same record and seed, and the last two rows give KL-Rec and KL-OT from the default initialization, for reference.

MethodInitial planIn-distribution costOut-of-distribution cost (2020)Change vs. default, in distributionChange vs. default, 2020
KL-Gen1. Risk-neutral, one scenario (default)31.0 ± 0.533.4 ± 3.3
KL-Gen2. SAA30.8 ± 0.133.1 ± 1.1−0.06 ± 0.06−0.52 ± 0.20
KL-Gen3. KL-OT (δ=3\delta = 3)30.8 ± 0.133.0 ± 0.4+0.00 ± 0.42−0.05 ± 2.28
KL-Gen4. KL-Rec31.0 ± 0.233.1 ± 1.0+0.03 ± 0.18−0.10 ± 0.29
KL-Recdefault32.5 ± 1.053.2 ± 5.2
KL-OT (δ=1\delta = 1)default30.4 ± 0.435.4 ± 3.1

These results show, in the capacity expansion case, that the final cost is nearly identical across initializations, both in distribution and out of distribution, and the paired differences from the default initialization are small relative to their standard errors with the only exception being the SAA initialization, which results in marginally lower out-of-distribution costs.

For the newsvendor study, where the single round is aimed at the SAA decision, we have a related result from experiments run while preparing the manuscript. In these experiments, we ran KL-Gen-P (with a log-normal nominal) twice in sequence, with the second run fitting its proposal with the first run's decision as the initialization rather than at the SAA decision and drawing a new set of scenarios that replaces the first, which we termed "re-aiming". The following table reports in-distribution and out-of-distribution costs above the optimum (%, median over records) at the selection rule used in Table 1.

NNShiftID, one roundID, re-aimedOOD, one roundOOD, re-aimed
16Level39275669
16Tail39275468
16Support39274750
16Coincidence44383231
64Level21143342
64Tail21145059
64Support21143134
64Coincidence31222117

These results show that re-aiming lowers the in-distribution cost in every setting while the out-of-distribution cost can increase or decrease depending on the nature of the distribution shift. Specifically, it rises under the level and tail shifts, most clearly at N=16N = 16, and falls under the coincidence shift at N=64N = 64. With T=1T=1, the decision at which the sampler is aimed plays a substantial role in governing the tradeoff between in-distribution and out-of-distribution performance. In the manuscript, we hold this decision fixed so that every method that requires an initialization uses the SAA decision, which depends only on the record, but do not claim that it is the best initialization method in general.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of responses and planned revisions

We thank the reviewers for their feedback, which guided new analysis and experiments for this work. We have responded to each reviewer's questions and summarize our findings below as well as planned revisions to the manuscript.

Theory. The revision will give the condition under which the KL-DRO problem centered at the learned nominal is well-posed, which follows from (Hu and Hong 2013) and is easily extended to the mean–CVaR objective. We are also working toward a convergence result for the adaptive algorithm, which we will include in the revision if it is completed in time.

Experiments. To address the reviewers' questions and give a more complete picture of the performance of our approach, we conducted additional experiments on the cost across iterations and the sensitivity to initialization, a reconstruction of Table 1 under the risk-neutral objective, the full newsvendor study with bounded demand (i.e., an extension of the previous newsvendor study to a well-posed setting, in which the conclusions of Table 1 hold), and an evaluation of the frozen capacity expansion plans on held-out years and pre-specified synthetic shifts. Importantly, this held-out evaluation narrowed two claims of the original submission in the capacity expansion setting, since the advantage over KL-OT holds only for the shortest records, and it did not appear under wind lulls alone, which disagrees with our proposed explanation of why KL-Gen outperforms under distribution shift, and we will revise the manuscript accordingly.

Writing. We are tightening the writing, in particular the definitions of the record, nominal, scenario, and recourse, and the connections between the two-stage formulation, the ambiguity set, and the sampler across sections.

We thank the reviewers for a discussion that led us to state the paper's setting and claims more precisely and to test them more broadly. We summarize the revision below by the concern each change answers.

The problem being solved and convergence (KDsN, Ay1S). The paper now states its setting directly. We target problems in which the scenario budget is tight, so the master is a finite-scenario approximation of the problem under the learned nominal. A new study (App. A.10) measures that approximation. Both tilted sampling schemes are closer to the population problem than independent sampling at every budget, and with 400 scenarios the fitted proposal is within about 3% of it at small radii. Under the true distribution, the learned nominal with a small radius gives the lowest cost in each of the four settings reported. The conclusions about protection also hold in the regime closest to the population problem (App. A.8). In capacity expansion, most of the decrease in cost occurs within the first four rounds, and running 24 rounds or starting from other initial plans changes the final cost little (App. B.8).

Well-posedness (KDsN). Demand in the newsvendor study is now bounded, so the KL problem is finite at every radius, with the condition cited to Hu and Hong (2013). The full study was repeated, and no finding changes with or without the bound (App. A.2).

Out-of-distribution evaluation (KDsN). Every final plan was evaluated without re-optimization on the held-out years 2015 to 2019 and under shifts specified in advance (App. B.9). With demand 10% higher, our method remains cheaper than KL-Rec at every record length and radius pair, at comparable cost without a shift, and its advantage over KL-OT is significant for the shortest records. The claims of Sec. 5.2 are stated with this scope.

Risk measure (Ay1S). Table 1 is repeated under the risk-neutral objective with all rows (App. A.9). The fitted proposal remains the cheapest protection for short records.

Formulation and positioning (WGmi). The record, the two stages, the recourse, the learned nominal, and a scenario are defined where the formulation is introduced, and each case study is stated in that notation (App. A.1, B.1). The paper positions its contribution relative to Iyengar et al. (2023) and Love and Bayraksan (2016), and the title now names the two-stage setting.

Scale and computation (WGmi). On a 214-county model with 72-hour scenarios, our method, with a Gaussian copula as the nominal at this scale, is cheaper than KL-Rec out of distribution at every record length, and we report the comparisons in which KL-OT does as well or better (App. B.10). A run takes about 10 minutes on the 20-zone model and 3 to 4 hours on the larger one, and generation adds no measurable time. Our method also remains cheaper out of distribution than an SAA on every record day (App. B.7).

Capacity expansion model. The capacity expansion study now also includes a linearized unit commitment of gas plants, corrected zone assignments, and a single capital budget in place of one per record, with all runs repeated. The conclusions of Sec. 5.2 are unchanged, and the out-of-distribution margins over KL-Rec are larger than in the reviewed version.

The additional experiments largely confirm the findings of the original submission, and they also showed us where some claims needed to be stated more carefully, which we have done in the revised text. We hope the revision resolves the concerns about soundness and scope, and we are grateful for the feedback that led to it.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.