acceptodds
Under review as a conference paper at ICLR 2027

Learning Context-Matched Uncertainty Representations for Robust Optimization

Abstract

Two-stage robust optimization is shaped by the adversarial scenarios admitted by its uncertainty set. When side information is observed before a decision, learning this set from paired context–scenario data requires both identifying relevant historical uncertainty patterns and representing their nearby variation. However, a learned uncertainty set may still admit scenarios poorly matched to the observed context. The adversarial problem can return these scenarios when they maximize recourse cost. Existing methods restrict such scenarios through context-matching constraints in the adversarial problem, which can complicate its solution and require problem-specific reformulations. To address this issue, we propose CUB-RO, a nonparametric uncertainty-set learner that incorporates context matching into set construction, avoiding additional context-matching constraints in the adversarial problem. Given a new context, CUB-RO retrieves the K nearest historical contexts using a prespecified dissimilarity measure and constructs the uncertainty set as a union of local regions around their paired complete scenarios. The matched centers and local radii jointly determine the uncertainty-set boundary. We establish sufficient conditions for the learned set to converge to the true conditional support as the sample size grows. Moreover, we derive finite-sample bounds on the excess worst-case cost of the resulting decisions relative to the true conditional optimum. Experiments on Feature-Based Newsvendor (FBNV), Robust Portfolio Optimization (RPO), and Security-Constrained Unit Commitment (SCUC) show improvements over the evaluated baselines on the respective primary metrics. Compared with robust optimization using only the retrieved historical scenarios as the uncertainty set, CUB-RO reduces value-at-risk and conditional value-at-risk on RPO by 10.63% and 12.92%, respectively, and lowers mean realized costs by 4.78% on FBNV and 0.58% on SCUC.

Primary area:
optimization
Keywords:
Two-stage robust optimization, Learned conditional uncertainty sets, Nonparametric learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Pvu5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CUB-RO, a nonparametric method for constructing contextual uncertainty sets for robust optimization. It selects historical scenarios with similar contexts and places a ball around each one to form an uncertainty set. The authors analyze set approximation and decision quality and evaluate the method on newsvendor, portfolio optimization, and unit commitment tasks.

Critical strengths

The proposed construction aims to reduce excessive conservatism and induce more effective downstream decisions, with encouraging results on the evaluated benchmarks.

Critical weaknesses

However, nearest-neighbor selection and unions of local neighborhoods are established ideas. The methodological novelty of combining these ingredients for decision problems appears limited.

Questions for authors
  1. In Eq. (5), each local radius is proportional to the distance to the nearest other selected scenario. A rare, isolated scenario can therefore receive a very large radius, potentially allowing many extreme scenarios into the prediction set and leading to conservative decisions. What justifies this radius rule, and how sensitive is performance to isolated centers?
  2. Section 5.2 should clarify what the target set C(s) should represent for different task objectives, such as minimizing mean loss or CVaR. How do the upper bounds on worst-case costs and decision suboptimality over C(s) relate directly to these objectives?
  3. The parameter-selection procedures need further justification. In the newsvendor experiment, mean loss is the primary metric and CUB-RO achieves the lowest test mean loss, yet K and \beta are selected by minimizing validation CVaR_{80} subject to a mean-loss constraint. Why use this criterion, and would tuning directly for mean loss change the conclusions? For SCUC, how were the fixed values K = 10 and \beta = 0.05 determined?
2 comments

Thank you for the questions about the contribution, radius rule, and evaluation objectives. We clarify these design choices and report additional radius and tuning controls.

W1. Methodological contribution. We will make clearer how the established ingredients serve the decision problem addressed here. A region fitted to context-relevant observations can still admit costly adverse directions weakly supported by those observations. CUB-RO uses retrieved complete trajectories as anchors and spacing-based local neighborhoods to control the adversarial domain before optimization, without adding context-matching constraints to the AP.

SRO-SI weights neighborhood-wise worst costs to form an average; CUB-RO uses context to select and shape the set for a single worst-case recourse objective. The analysis connects directional boundaries to worst-case costs and separates representation and search errors in the decision bounds. This contextual representation and its decision analysis form our contribution.

The same-center Portfolio comparison isolates expansion: VaR/CVaR decrease by 10.63%/12.92%, with paired CVaR difference -0.2249 percentage points (95% interval [-0.2774, -0.1707]). This provides controlled evidence for the construction on that task.

Q1. Radius rule and isolated centers. Your question identifies the trade-off captured by Proposition 3. Nearest-center spacing supplies a local scale: dense groups receive smaller neighborhoods and isolated centers larger ones, with a common multiplier controlling expansion. Larger radii can reduce missing uncertainty while introducing additional uncertainty. The consistency theorem requires the largest physical radius to vanish.

To examine sensitivity directly, we fixed K, the multiplier, and all centers in the 23 Newsvendor folds. For each query, we capped only the most isolated center's radius at the median radius or set it to zero, retaining the center. Neither intervention was tuned on test outcomes. On the same 92 decisions:

Radius treatmentMeanVaR95CVaR95
Original85.8413231.60430.67
Isolated radius capped at median84.3258231.60429.07
Isolated radius set to zero84.3627231.60428.91

Both interventions change 32/92 decisions; maximum changes are 115.84/118.88 demand units. The lower observed mean losses show that moderating an isolated radius can improve this benchmark. We will include this sensitivity and explain the spacing rule as a local-scale choice whose behavior should be checked for isolated centers. In saved SCUC runs, the median within-day maximum-to-median radius ratio is 1.62, with maximum 2.10; these describe radius heterogeneity rather than a controlled outlier intervention.

Q2. Target set and evaluation objectives. We will clarify the two roles. In Section 5.2, C(s) is the target scenario set for robust decision quality; the consistency theorem takes it to be the true conditional support. For a fixed decision, let HC(x,s)=sup⁡z∈C(s)Q(x,z)H_C(x,s)=\sup_{z\in C(s)}Q(x,z). If the conditional distribution is supported on C(s), then

\le H_C(x,s)\le H_U(x,s)+\varepsilon_{\rm miss}(s).$$ Adding first-stage cost gives total-cost bounds. The decision-gap guarantee compares with the robust optimum over C(s); mean loss and CVaR evaluate the resulting decisions empirically. A mean/CVaR regret guarantee would require comparison with the corresponding distributional optimum. We will make these roles explicit in Section 5.2. **Q3. Parameter selection.** Newsvendor's validation CVaR80 criterion balances tail performance with a mean-loss requirement of at most 5% above SAA. Each 52-week window uses 36 observations for fitting and 16 for validation, then refits the selected configuration on all 52. CVaR80 uses an upper-tail mass of 3.2 validation observations, versus 0.8 for CVaR95. If no candidate met the mean constraint, we selected the lowest-CVaR80 candidate from the full grid. This occurred in 1/23 windows; we will specify this fallback in Appendix B.2. Following your suggestion, we also selected directly by validation mean, keeping the split, grid, refitting, and held-out cases fixed, with ties broken by CVaR80 then VaR80. Mean/VaR95/CVaR95 become 86.04/231.60/430.33, versus 85.84/231.60/430.67 originally. The paired mean difference is 0.204 [-0.168, 0.803] (95% four-fold-block interval), indicating similar test performance. The separate same-neighborhood zero-radius control gives mean 85.05. We will report these controls to distinguish the tuning objective, neighborhood selection, and expansion. For SCUC, K=10 and beta=0.05 were fixed without tuning or a specific selection criterion. Figure 5(c) reports sensitivity on four preselected held-out days, including lower-cost settings.

Thank you for your detailed responses. I have reviewed the additional radius sensitivity and parameter-selection experiments. I will maintain my score.

Reviewer_DSX7OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes CUB-RO, a nonparametric method for constructing context-dependent uncertainty sets for robust optimization. Given an observed context, it retrieves nearby historical contexts and retains their paired complete uncertainty scenarios. It then constructs a union of ellipsoidal neighborhoods around those scenarios, with local radii determined by distances between the selected scenarios. Context matching is therefore incorporated into uncertainty-set construction rather than imposed through additional constraints in the adversarial optimization problem.

The analysis establishes convex-hull invariance for convex recourse costs, bounds decision suboptimality through omitted and introduced uncertainty, and incorporates error from approximate adversarial optimization. It also provides sufficient conditions for conditional-support and decision-value consistency. For LP recourse, the paper develops a critical-region-guided projected-gradient separator.

Experiments cover feature-based newsvendor, portfolio optimization, and security-constrained unit commitment. CUB-RO obtains the best reported primary point estimates across the three benchmarks, with the strongest statistical evidence coming from the paired portfolio CVaR comparisons. Additional studies examine representation distortion, solver behavior, parameter sensitivity, and variation across seeds and time blocks.

Critical strengths
  1. The uncertainty-set construction is simple and operationally useful. It retains complete historical trajectories, preserving their joint structure, while permitting controlled variation around them. The resulting representation supports exact adversarial queries for the newsvendor and portfolio problems and can also be integrated with mixed-integer first-stage decisions and LP recourse. This provides a practical alternative to learned generators and more complicated contextual adversarial formulations.

  2. The analysis gives a clear account of what the representation can and cannot accomplish. Convex-hull invariance prevents attributing the benefit merely to excluding interior combinations. The missing and spurious distortion terms explain the trade-off associated with local expansion, while the approximate-separation bound distinguishes representation error from optimization error. These are relatively direct geometric and stability arguments, but they are well matched to the proposed method.

  3. The portfolio evaluation provides the strongest evidence for the method. CUB-RO and zero-radius KNN-RO use the same context-matched centers, making their comparison informative about local expansion. The reported reductions are approximately 10.63% in VaR and 12.92% in CVaR. The paired time-block analysis supports the CVaR improvement over both KNN-RO and the closer cKNN-GB baseline. The year-stratified reporting also appropriately avoids treating asset-selection runs sharing market dates as independent repetitions.

  4. The experimental reporting is transparent about limitations. The paper reports newsvendor mean–tail trade-offs, intervals crossing zero for some comparisons, SCUC stopping failures, and a separation audit performed on decisions distinct from the main table. These disclosures make the evidence easier to assess and support confidence in the narrower conclusions established.

Critical weaknesses
  1. The evidence of improvement is uneven across benchmarks. On newsvendor, the paired mean-cost intervals against SAA and KNN-RO cross zero, and CUB-RO has higher tail losses than several more conservative alternatives. On SCUC, the paired mean-cost intervals against SAA and KNN-RO also cross zero. These findings do not invalidate the favorable point estimates, and the appendix acknowledges them, but they limit the breadth of the empirical advantage. The main newsvendor comparison also changes parameter selection as well as local expansion: CUB-RO selects its neighborhood size through validation, whereas KNN-RO retains a frozen neighborhood size.

  2. The main SCUC comparison does not fully separate uncertainty-set representation from adversarial solution quality. CUB-RO uses CRG-PGA, whereas the main AGRO and cAGRO results use standard PGA with different search budgets. Table 8 reports CRG-PGA runs on those generative representations, but not their corresponding held-out decision costs. Table 9 also shows substantially different stopping behavior across representations. The end-to-end pipeline comparisons remain useful, but attributing the cost differences specifically to uncertainty-set geometry is less secure. The separate separation-error audit does not resolve this for the main-table decisions.

Questions for authors
  1. What are the held-out mean costs and tail costs for the AGRO and cAGRO CRG-PGA runs already reported in Table 8? These results would help distinguish improvements associated with the uncertainty representation from differences in the adversarial solver.

  2. For SCUC runs that reach the outer iteration cap without satisfying the stopping criterion, which first-stage iterate is selected for settlement? Is the same selection rule used across methods?

  3. The newsvendor comparison with KNN-RO combines different neighborhood-size selection procedures with the presence or absence of local expansion. Are results already available for the zero-radius construction using CUB-RO’s selected neighborhood sizes? This would clarify how much of the reported difference is attributable specifically to expansion.

1 comment

Thank you for identifying the comparisons that would make the sources of improvement clearer. Following your suggestions, we evaluated the saved CRG-PGA decisions, increased search budgets, and performed the same-neighborhood Newsvendor comparison.

Q1 / W2. Held-out costs with CRG-PGA. The runs underlying Table 8 give the following costs, in hundreds of USD/day. Metrics are computed over 30 held-out days within each seed, then averaged over five seeds.

MethodMeanCVaR95CUB-RO minus baseline: mean difference [95% interval]
CUB-RO1935.742260.25—
AGRO, CRG-PGA1965.392330.73-29.65 [-44.46, -15.23]
cAGRO, CRG-PGA1971.302426.36-35.56 [-57.65, -14.69]

CUB-RO retains mean reductions of 1.51%/1.80%; both intervals are below zero under 20,000 paired circular five-day-block resamples with seeds fixed. We will add these costs alongside the solver results so readers can assess the comparison when the baselines also use CRG-PGA.

The manuscript's main AGRO/cAGRO rows use latent-space PGA. The Table 8 CRG-PGA rows share Appendix B.4's four starts, one additional independent trajectory, J=30, and 12-iteration outer cap. cAGRO uses the same standardized 24-hour context as CUB-RO. Both baselines hit the inner cap in 150/150 runs; outer stopping occurs in 2/150 AGRO and 6/150 cAGRO runs.

To examine the effect of additional computation, we kept each generator/set fixed on preselected day 420 and increased the outer cap from 12 to 48 with larger inner budgets:

MethodMean realized cost: original / enlargedOuter stopping: original / enlargedTime factor
AGRO2014.633 / 2015.1750/5 / 1/515.9x
cAGRO2006.084 / 2012.0440/5 / 4/58.6x

The original residual threshold remains 10−410^{-4}. Even among the subgroups meeting outer stopping, the mean same-seed cost increases are +0.149 for AGRO and +3.634 for cAGRO. At identical first-iteration decisions, enlarged-budget AP values are higher in all ten runs. Across five cAGRO seeds, extra first-stage cost (+6.112) exceeds the realized-recourse reduction (-0.153).

The higher realized costs are also observed in the stopped subgroups despite the substantial extra computation. This is consistent with Section 4: stronger AP search can expose costlier scenarios in the same set and induce extra hedging whose realized savings do not cover its cost. Outer stopping with a local AP is not a global certificate. Figures 1–2 provide complementary fixed-decision diagnostics; the separate separation audit remains distinct from the main-table decisions.

Q2. Decision selected at the iteration cap. Yes, the same selection rule is used across the iterative SCUC methods. We retain the visited first-stage decision with the smallest recorded first-stage cost plus recourse cost returned by its adversarial search. Held-out residuals enter only subsequent settlement. The score estimates the robust objective under approximate separation; it is not a certified upper bound. The stopping residual describes the terminal iteration, which can differ from the selected incumbent. We will state this common rule explicitly.

Q3 / W1. Newsvendor with the same selected neighborhood. Your proposed control isolates expansion from neighborhood selection. We held CUB-RO's selected K and centers fixed in all 23 folds and changed only the radius to zero:

ConstructionMeanVaR95CVaR95
Same neighborhood, zero radius85.05231.60430.09
CUB-RO selected radius85.84231.60430.67

Over 92 decisions, the paired mean difference is 0.785 [-0.259, 1.947], using 20,000 circular four-fold-block resamples. The controlled point estimate favors zero radius. We will include this ablation to distinguish expansion from the neighborhood selection that also changes in the main KNN-RO comparison.

W1. Scope of the empirical evidence. We will bring the appendix's paired-interval qualifications into the abstract and discussion. Portfolio provides the strongest controlled expansion evidence. Newsvendor's mean differences against SAA/KNN-RO are -1.20 [-16.97, 14.60]/-4.31 [-12.38, 3.02], and its CVaR95 is higher than KNN-RO's. SCUC's corresponding mean-cost intervals also cross zero. Its small percentage gains are contextualized by the limited gap to the perfect-information bound: CUB-RO closes 19.36%/22.29% of the SAA/KNN-RO gaps (Section 6.2, Appendix C.3). These complementary summaries clarify the magnitude and uncertainty of the observed gains.

Reviewer_L89LOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes a method for building context-dependent uncertainty sets for two-stage robust optimization. Given a new context, it retrieves the K nearest historical contexts and places a small ellipsoid around each paired scenario, with radius proportional to the distance to the nearest other center. The union of these ellipsoids is the uncertainty set. The authors argue this excludes scenarios poorly matched to the context. They bound decision suboptimality by the distance between learned and target sets, prove consistency as the sample grows, and give a projected-gradient separator (CRG-PGA) for LP recourse. Experiments cover a feature-based newsvendor, a portfolio problem, and a 144-dimensional SCUC instance on CAISO data.

Critical strengths

The premise is sound. In high-dimensional settings with little data, a region fitted to the data admits corner points that recombine marginally plausible coordinates into implausible joint scenarios. Anchoring the set to retrieved historical trajectories is a reasonable way to avoid this.

The IDCC diagnostic (Table 14) shows this failure: on every held-out day the returned adversarial scenario sits at a box bound in over 99% of coordinates while never leaving the marginal training range.

The support-function view is the right lens: for piecewise-affine recourse, only how far the set extends along the directions exposed by the recourse problem matters. The paper should build on it in the main text.

Critical weaknesses
  1. The SCUC result is not distinguishable from the simple baselines. In Table 7, the paired bootstrap intervals for CUB-RO against SAA and against KNN-RO both include zero. The abstract reports 0.58% against KNN-RO without noting this.
  2. Local expansion is the only component beyond KNN-RO, and it is supported on one benchmark. On newsvendor the selector picks zero radius in 10 of 23 folds. On SCUC the gap is within noise. Only portfolio, a single-stage problem with affine loss, shows a clear effect, but the paper does not say why.
  3. Sections 1 and 4 say CUB-RO excludes scenarios poorly matched to the context. But by Prop. 1 the union and its convex hull give the same worst-case value, and the hull contains those scenarios. What matching does is pick the K points that span the hull. Figure 1 is explained by Prop. 5, but Prop. 5 sits in the appendix and is not cited in Section 4.
  4. There are twelve numbered results; three are stated in the body and a fourth is summarized. Props. 6 and 8 and Cors. 9 to 12 are not cited in the main text. The paper should say which results are standard and which are new, and whether any of them guides the choice of K, the radius multiplier, or the scaling matrix.
  5. Two details weaken the SCUC comparison. The main configuration (K = 10, multiplier 0.05) is fixed, but how it was chosen is not reported, and Fig. 5(c) shows other settings doing better on the diagnostic days. Under CRG-PGA (Table 9), AGRO and cAGRO hit the inner search cap in all 150 runs and rarely meet the outer stopping criterion; the main-table PGA rows stop after three outer iterations. Neither indicates convergence, so margins against them are not reliable.
  6. The abstract leads with numbers the appendices do not support.
Questions for authors

Please see the weaknesses above. Additionally:

  1. How was the SCUC configuration chosen?
  2. Given Prop. 1, what does excluding poorly matched scenarios mean beyond selecting which K points span the convex hull?
  3. Can the authors report SCUC results with converged AGRO and cAGRO baselines?
  4. Is there a version of Figure 3(a) with the target defined by held-out realized residuals rather than by the retrieval rule?
  5. Why does local expansion help on portfolio and not on newsvendor or SCUC?
3 comments

Thank you for recognizing the motivation and highlighting the support-function perspective. Your questions help us connect the construction, analysis, and evidence more clearly. W1-W6 follow the original weakness order; this comment addresses W3-W4 and Q2/Q4, and Comment 2 addresses the remaining points.

Q2 / W3. Convex-hull invariance and boundary scenarios. The support-function perspective you highlight makes the intended mechanism precise. Proposition 1 establishes max⁡UQ(x,z)=max⁡conv⁡(U)Q(x,z)\max_U Q(x,z)=\max_{\operatorname{conv}(U)}Q(x,z) for our union U. A separately fitted region G can have different directional boundaries and hence a different worst-case value. We will make this comparison explicit.

To illustrate how boundary exclusion and convex-hull invariance fit together, suppose G is a convex region containing U. Then

\quad\Longrightarrow\quad v\notin\operatorname{conv}(U).$$ Otherwise, $v$ would be a convex combination of other points in $G$, contradicting its extremality. Thus, an extreme point admitted by the enclosing region but excluded from our union cannot be recovered by convexifying the union. If such points determine higher recourse costs, excluding them can change the worst-case value without contradicting Proposition 1. The centers, local radii, and scaling jointly determine our set's support: $$\sigma_U(b)=\max_i\{b^\top z_i+\widehat\beta d_i\|B_s^\top b\|_2\}.$$ Matching selects the centers; local radii and scaling determine the directional extent beyond them. For LP recourse, $H_G(x)=\max_p\{a_p(x)+\sigma_G(b_p)\}$. Proposition 5 gives a precise sufficient condition: if some actual recourse piece satisfies $a_p(x)+\sigma_G(b_p)>H_U(x)+\epsilon$, every $\epsilon$-accurate separator on G returns a value above $H_U(x)$. This requires no nesting; the enclosing-region example above is not an assumed relationship between every baseline and CUB-RO. We will connect Proposition 5 explicitly to Figure 1 and clarify the exclusion language in Sections 1 and 4. Table 14 illustrates the boundary issue empirically: on average, 99.63% of coordinates of the returned IDCC scenarios lie within 1% of the coordinate's box span from a bound, while none leave their marginal training ranges. This shows how joint boundary scenarios arise without marginal extrapolation; it is a fixed-decision diagnostic, separate from the baseline ranking. For computation, our implementation retains the local components and the decomposition $\max_{z\in U}Q(x,z)=\max_i\max_{z\in U_i}Q(x,z)$ in Eq. (19), including in 144-dimensional SCUC. This preserves separate component queries and support evaluations without explicitly reconstructing a high-dimensional convex-hull boundary. **W4. Standard results, analytical contributions, and parameter guidance.** Your suggestion will help readers distinguish the main analysis from its supporting tools. Our analytical contribution connects this contextual construction to decision quality through missing/introduced uncertainty (Theorem 2), the local-radius trade-off (Proposition 3), conditional-support and decision-value consistency (Theorem 4), and representation plus terminal separation error (Theorem 7). Propositions 1, 6, and 8 use standard convexity, exact C&CG bounds, and restart arguments. Proposition 5 applies standard support identities to the comparison above; Corollaries 9-12 specialize or combine the analyses. We will make these roles and their main-text connections explicit. Under Theorem 4's assumptions, $K_n\to\infty$, $K_n/n\to0$, and vanishing largest physical radius $R_n(s)=o_p(1)$ ensure support consistency. The radius term involves the multiplier and scaling matrix. This gives asymptotic design guidance and a radius trade-off; finite-sample parameter selection remains an empirical design choice. **Q4. Figure 3(a) with held-out realized residuals.** Following your suggestion, we repeated the diagnostic using each day's actual residual as the target, independently of retrieval, with the saved decision libraries (30 held-out days, 152 decisions). At K=10, median total recourse distortion is 285.23 for matched retrieval versus 771.17/781.77 for random/shuffled controls, averaging five ordering seeds per day. Paired five-day-block 95% intervals for the mean differences are [-617.65, -232.13]/[-614.97, -209.40]. The independent-reference comparison supports matching beyond the retrieval-defined target. With realized residuals as targets, K=40 no longer forces zero distortion. These are empirical recourse diagnostics; Theorem 4 separately addresses conditional-support consistency.

Thank you for identifying the comparisons needed to interpret the empirical gains. We address Q1/Q3/Q5 and W1-W2/W5-W6 below, with stopping status and controlled comparisons reported alongside the costs.

Q1 / W5. SCUC configuration. K=10 and beta=0.05 were fixed without tuning or a specific selection criterion. Figure 5(c) reports sensitivity on four preselected held-out days, including lower-cost settings. We will clarify that this setting is not claimed optimal.

Q3 / W5. Baseline convergence and search budgets. Following your suggestion, we increased search budgets and examined the runs meeting outer stopping. Their mean paired realized costs increased: +0.149 for AGRO and +3.634 for cAGRO. We also settled the CRG-PGA runs underlying Table 8 (hundreds of USD/day; metrics over 30 days per seed, then averaged over five seeds):

MethodMeanCVaR95
CUB-RO1935.742260.25
AGRO, CRG-PGA1965.392330.73
cAGRO, CRG-PGA1971.302426.36

CUB-RO's mean reductions remain 1.51%/1.80%; paired mean differences are -29.65 [-44.46, -15.23]/-35.56 [-57.65, -14.69] (95% intervals, 20,000 circular five-day-block resamples, seeds fixed). The favorable comparison persists when the baselines also use CRG-PGA. These rows share the Appendix B.4 configuration: four starts, one additional independent trajectory, J=30, and a 12-iteration outer cap. The main AGRO/cAGRO rows use latent-space PGA (Appendix B.2, Table 4). Table 8's PGA rows report mean/median values of three outer iterations, which is not convergence evidence. Both CRG-PGA baselines hit the inner cap in 150/150 runs; only 2/150 AGRO and 6/150 cAGRO runs meet outer stopping.

On preselected day 420, we kept each generator/set fixed and increased the outer cap from 12 to 48 with larger inner budgets. All five seeds per method originally missed outer stopping. Pairing the same seeds before/after gives:

MethodEnlarged-budget outer statusSeedsMean paired cost increaseIncrease
AGROMet stopping1+0.149+0.0074%
AGROMissed stopping4+0.640+0.0318%
cAGROMet stopping4+3.634+0.1810%
cAGROMissed stopping1+15.264+0.7635%

The threshold remains 10−410^{-4}; the remaining runs reach 48 iterations, and most inner searches exhaust their budgets. Mean solve time rises 15.9x/8.6x; one baseline seed has required about 12 hours. At identical first-iteration decisions, enlarged-budget AP values are higher in all ten runs. Across the five cAGRO seeds, extra first-stage cost (+6.112) exceeds the realized-recourse reduction (-0.153).

These results show that higher realized costs are also observed in the stopped subgroups. They are consistent with Section 4's mechanism: stronger AP search can induce extra hedging whose realized savings do not cover its cost. Outer stopping with a local AP is not a global optimality certificate.

Q5 / W2. Task-dependent expansion effects. The additional controls help explain this difference. On Portfolio, identical centers isolate expansion. Even for affine loss, expansion adds an allocation-dependent directional-support penalty. VaR/CVaR decrease by 10.63%/12.92%; paired CVaR differences against KNN-RO and cKNN-GB are -0.2249 [-0.2774, -0.1707] and -0.0745 [-0.1332, -0.0183] percentage points (95% intervals).

For scalar Newsvendor, extreme endpoints determine the robust objective. Holding CUB-RO's selected K and centers fixed, zero radius gives mean loss 85.05 versus 85.84 with expansion; the paired difference is 0.785 [-0.259, 1.947] over 92 decisions. This separates neighborhood selection from expansion in the original comparison: expansion alone gives no point-estimate gain here, consistent with zero radius being selected in 10 of 23 folds.

SCUC has limited remaining improvement space. As reported in Section 6.2 and Appendix C.3, the perfect-information lower bound is 1896.35, versus 1945.20 for SAA and 1947.04 for KNN-RO. Their gaps are only about 2.5%-2.6% of total cost. CUB-RO's 1935.74 closes 19.36%/22.29% of these gaps, despite total-cost reductions of only 0.49%/0.58%. This contextualizes the small percentage gains; the paired intervals against SAA and KNN-RO still include zero.

W1 / W6. Statistical claims and the abstract. We will bring the appendix's paired-interval qualification into the abstract: "With identical centers, CUB-RO reduces Portfolio VaR/CVaR by 10.63%/12.92%, with paired evidence for CVaR improvement. Newsvendor and SCUC show favorable mean-cost point estimates, but paired intervals against SAA and KNN-RO include zero." This aligns the headline claim with the evidence and separates expansion from end-to-end gains.

The rebuttal answers my questions with additional experiments, but the new controls show the method helps only on the single-stage portfolio problem. I am raising score by one point.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.