acceptodds
Under review as a conference paper at ICLR 2027

Generative Distributionally Robust Optimization

Abstract

Generative models increasingly supply scenarios for downstream optimization, but the resulting decisions are vulnerable to model misspecification. Standard transport-based distributionally robust optimization (DRO) accepts arbitrary nominal samplers without requiring worst-case laws to respect generator structure, while existing generative ambiguity-set methods enforce such structure but typically rely on model-specific access or training-data certificates. We introduce Generative Distributionally Robust Optimization (GDRO), a general framework that accepts any pretrained conditional generative model as its nominal model and optimizes against adversarial laws induced by a chosen conditional generator family. The key is to pair sampler representations of conditional laws with a Sinkhorn certificate that compares nominal and adversarial output laws from samples alone. This enables generator-constrained robust optimization across explicit and implicit nominal models, controlling distributional proximity at the decision context without requiring nominal likelihoods or gradients. We develop a stochastic primal–dual algorithm for an expected mini-batch formulation and establish bounds linking Sinkhorn proximity to changes in expected downstream loss and quantifying the Lagrangian relaxation gap. Experiments span convex contextual newsvendor problems with explicit generators on synthetic and M5 retail demand data, and nonconvex robot navigation with the implicit SocialGAN generator. Relative to nominal decisions, GDRO reduces rare-context inventory regret by 63% on synthetic data and navigation collisions by 66%, while maintaining a 99.7% arrival rate.

Primary area:
optimization
Keywords:
distributionally robust optimization, generative models, optimal transport, contextual stochastic optimization, Sinkhorn divergence

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_cS5vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Generative Distributionally Robust Optimization (GDRO) is proposed as a framework for robust decision-making built around a GAN-like adversarial game. A trainable generator maximizes expected decision loss, while a decision optimizer seeks to minimize it. A Sinkhorn constraint keeps the adversarial output distribution close to that of a frozen nominal generator. Primal-dual updates enforce this constraint and optimize the task-specific decision. The applications are promising, but the methodological novelty appears to lie mainly in integrating established generative adversaries, transport constraints, and alternating optimization.

Critical strengths

The framework can work with a frozen nominal generator using only its output samples, while restricting adversarial distributions to a chosen generator family. The experiments on inventory management and robot navigation show improvements over nominal optimization, including lower rare-context regret and fewer collisions while maintaining a high arrival rate.

Although the individual components are not particularly novel, the paper integrates them into a coherent framework.

Critical weaknesses

The methodological novelty is not sufficiently clear. The framework mainly combines generative adversaries, Sinkhorn constraints, and standard primal–dual optimization, and a sharper comparison with existing generative DRO methods is needed.

There is extensive theoretical work on GANs and related minimax optimization, but it is unclear how the analysis here relates to this literature or what new insight it provides for the proposed constrained adversarial setting. The optimization guarantees also assume exact inner maximization, which is not ensured by the finite gradient updates used in practice. The convexity requirement for the outer guarantee does not cover the nonconvex navigation experiment.

Questions for authors

Beyond putting familiar adversarial optimization components together for decision-making, what genuinely new formulation, capability, or theoretical insight does this paper provide?

2 comments
AuthorsOpenReview ↗

Response to Reviewer cS5v (1/2): Weakness 1 and your question

Response to Reviewer cS5v (1/2): Weakness 1 and your question

Thank you for the candid assessment. To our knowledge, GDRO is the first framework that lets a decision maker choose a pretrained generator to define which distribution shifts to guard against while prescribing a hard budget on how far its outputs may depart from a frozen reference distribution. We measure that budget directly from output samples, so it applies even to an implicit adversary such as SocialGAN, a case the GAS-DRO authors explicitly leave open.

Weakness 1. The methodological novelty is not sufficiently clear, and a sharper comparison with existing generative DRO methods is needed.

Answer: to our knowledge, GDRO is the only generative DRO method that keeps a user-prescribed budget as a constraint while accommodating any pretrained generator as the adversary (now in the main text: Introduction, "Relation to prior work", and Table 1; details in new Appendix C.1 with Table 14). The generator family is here part of the problem, not of the solver: it defines which laws can enter the worst case, so changing it changes the robust problem itself, even under exact optimization. Where a network approximates the worst case of a conventional ambiguity set, as the flow of FlowDRO does, the set carries no such restriction.

  1. Apart from the GAS-DRO line, the methods compared in Table 14 use a penalty in place of a budget. P-DRO, R-PDRO, DRAGEN and FlowDRO trade the adversarial loss against a divergence penalty, so how far the adversary departs from the reference is an outcome of the optimization.
  2. GAS-DRO, the closest method, also keeps a prescribed budget, but not for GANs. Its budget is the adversarial model's own reconstruction or denoising loss on nominal data, which requires a likelihood-based model. Its authors state that they "cannot currently provide a principled GAS-DRO design based on GANs" (authors' response).
  3. Why keeping the constraint matters for robust decision-making. GDRO measures the budget on output samples, where a transport discrepancy limits the change in a Lipschitz loss. Prescribing the budget therefore tells the user, before optimization, how far any admitted generator can move the expected loss of every candidate decision: by at most Lfρ+εlog⁡ML_f\sqrt{\rho+\varepsilon\log M} (Theorem 3.1). A penalty weight alone does not prescribe that budget. The algorithm must in turn meet it, and we audited how closely ours does: on fresh samples, the final adversaries' mean score is within the radius on 89.5% of navigation tasks and 51.5% of M5 contexts, and below 1.1 times the radius on 97.0% and 98.5% (Figures 4 and 8).

Our novelty is a bridge between DRO and the sampler perspective that GANs share, in which a distribution is handled through the samples of a generator. It serves a practical purpose: take the model a domain prefers as the reference, choose any pretrained generator as its adversary, set a budget, and obtain a robust decision. This bridge is what makes the bound above available. It even lets the reference and the adversary be different generators, since only output samples are compared: with a frozen CVAE–LSTM as the reference and a separately trained GAN as the adversary, GDRO keeps the cost reduction, at a larger radius and with more residual violations (Table 10).

Your question. Beyond putting familiar components together, what genuinely new formulation, capability, or theoretical insight does the paper provide?

Answer: building this bridge gave rise to new insights of each kind.

  1. A new capability. GDRO makes a pretrained implicit generator, here SocialGAN, available as the adversary under a prescribed budget, which the closest method does not offer (point 2 above). Figure 2, now in the main text, shows how that choice changes the adverse trajectories, and Table 4 shows what it buys the planner in both collisions and arrivals. Under the same Sinkhorn certificate, replacing the SocialGAN family by free support points raises collisions from 7.2 to 27.6 and lowers arrivals from 199.4 to 127.0 of 200.

  2. A new formulation. The ambiguity set consists of the generator laws within a prescribed budget of a frozen sampler, measured from output samples at the context where the decision is made (Eqs. 4 and 12). On rare inventory contexts, mean regret is 9.26 for GDRO, 16.79 for our conditional extension of GAS-DRO, and 25.09 for nominal planning (Table 2).

  3. A new theoretical insight. The algorithm solves the multiplier relaxation of the constrained problem. With a generator family as the adversary, the two can differ even under exact optimization, since the laws the family can produce need not form a convex set. We show that this gap is at most Lfρ+εlog⁡ML_f\sqrt{\rho+\varepsilon\log M}, the bound of Theorem 3.1, when a feasible generator reproduces the nominal law (Theorem 3.2; see part 2).

AuthorsOpenReview ↗

Response to Reviewer cS5v (2/2): Weakness 2

Response to Reviewer cS5v (2/2): Weakness 2

Weakness 2(a). Relation of the analysis to GAN and minimax theory, and its new insight for the constrained adversarial setting.

Answer: that literature asks whether and how well an adversarial game is solved. Theorems 3.1 and 3.2 answer two questions that arise only once the game is used to decide under a budget: what the budget guarantees for the loss of a decision, and how much the multiplier relaxation that the algorithm solves can overstate a decision's worst-case loss under that budget (now stated in the Introduction, "Relation to prior work"; details in new Appendix C.2).

  • Relation. GAN theory studies whether the game recovers a data distribution (Arora et al., 2017). Minimax theory studies how well a nonconvex game can be solved, for instance with an inner maximization oracle (Jin et al., 2020). Theorem 3.3 and its extensions belong to the second line and use standard tools.
  • Insight 1: what a budget estimated from samples guarantees (Theorem 3.1). In generative fitting, a discrepancy is a training signal. Here it is a certificate. With bounded outputs, squared transport cost and an LfL_f-Lipschitz loss, every generator within budget ρ\rho changes the expected loss of every candidate decision by at most C=Lfρ+εlog⁡MC=L_f\sqrt{\rho+\varepsilon\log M}.
  • Insight 2: the gap of the multiplier relaxation (Theorem 3.2). Once the adversary is restricted to a generator family, the laws it can realize need not form a convex set, so the multiplier problem is only a relaxation, even at a global solution. The square-root form of the first bound has a tangent at ρ\rho whose slope is one multiplier controlling all generators, including those over budget. When a feasible generator reproduces the nominal law, the relaxation therefore overstates each decision's worst-case loss by at most the same CC.

Weakness 2(b). The guarantees assume exact inner maximization, which finite gradient updates do not ensure.

Answer: you are right: Theorem 3.3 assumes a global inner maximizer. As is customary in analysis, it gives the assurance for the idealized case of abundant compute; the algorithm is what limited compute requires, and the revision connects the two by further analysis and by experiments (Point 3 of our comment to all reviewers; new paragraph "Theoretical analysis versus the practical algorithm" after Theorem 3.3).

  • Why the two differ. Solving the inner problem exactly presupposes, at the least, accurate evaluation of its expectations, which takes far more samples than a mini-batch. With limited compute, the adversary can only take a few stochastic gradient steps per update, so a gap between the two is unavoidable.
  • What the analysis still guarantees. First, the guarantees degrade linearly in the inner error: if the inner steps fall short of the Lagrangian maximum, every bound of Theorem 3.3 holds with the average expected shortfall δˉ\bar\delta added to Bμmax⁡/KB\mu_{\max}/\sqrt K (new Appendix A.7, Proposition A.9). We do not claim that a fixed number of gradient steps makes δˉ\bar\delta small for a nonconcave generator objective. Second, Theorems 3.1 and 3.2 involve no optimizer and hold at the batch size used: Eq. (21) bounds the loss change under whatever adversary the algorithm returns, in terms of that adversary's own expected score.
  • What the experiments show. First, the returned adversaries meet the budget closely: their mean score on fresh batches is within the radius on 179 of 200 navigation tasks and 103 of 200 M5 contexts, and below 1.1 times the radius on 194 and 197 (Figures 4 and 8). By Eq. (21), a 10% excess in the expected score, which this mean estimates, enlarges the loss bound by less than 5%. Second, longer inner updates alone need not improve the procedure: with ten times as many inner steps on 20 M5 contexts and 20 navigation tasks, M5 cost falls by 0.5%; in navigation, arrivals stay at 20/20 and the evaluation loss rises from 16.88 to 23.23 (Tables 9 and 11).

Weakness 2(c). The convexity requirement of the outer guarantee does not cover the nonconvex navigation experiment.

Answer: the revision extends the outer guarantee to smooth nonconvex losses, which cover the navigation loss once the pedestrian distance is smoothed at zero (new Theorem A.11 and Remark A.12, summarized in the main text after Theorem 3.3). Replace convexity by an ℓ\ell-Lipschitz gradient of the loss in the decision, and keep the other assumptions, the projected updates and the step sizes of Theorem 3.3. Then the relaxed objective is weakly convex, and the expected squared norm of the gradient of its Moreau envelope, averaged over the TT outer iterates, is O(T−1/2+K−1/2+δˉ)O(T^{-1/2}+K^{-1/2}+\bar\delta) (Eq. 71). Stationarity of the Moreau envelope is the standard criterion for weakly convex problems, and also the form of the outer guarantees of D-DRO and GAS-DRO.

Reviewer_1FLzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Generative Distributionally Robust Optimization (GDRO) for decisions made using pretrained conditional generators. The adversary generates alternative outcome distributions within a chosen generator family, while a Sinkhorn constraint limits their distance from the nominal distribution at the current context. The paper develops a stochastic primal–dual algorithm, analyzes risk and optimization bounds, and evaluates the approach on inventory management and pedestrian-aware robot navigation.

Critical strengths
  • Generative models can produce plausible scenarios while missing outcomes that matter for downstream decisions. I like the separation between access to the nominal sampler and optimization of the adversarial generator. Requiring only samples from the nominal model makes the framework potentially useful when likelihoods or gradients are unavailable.

  • The authors distinguish the population Sinkhorn constraint from the expected mini-batch constraint actually targeted by the algorithm. They also account for a possible Lagrangian relaxation gap rather than assuming strong duality for a nonconcave adversary. These distinctions make the analysis easier to interpret and are supported by detailed proofs.

Critical weaknesses
  • Theorem 3.3 assumes an exact maximizer of the expected inner Lagrangian at every step, as stated in Equation (26). In practice, the method takes a small number of stochastic gradient steps on a nonconcave generator objective. Consequently, the theorem gives limited assurance about the quality of the adversaries found in the experiments.

  • The methods use different radius mappings, sample counts, and optimization budgets. For example, GDRO uses adaptive budgets reaching 500 outer iterations in navigation, whereas W2 uses 180 outer iterations with a different sampling configuration. These differences do not establish an unfair comparison, but they make the results difficult to interpret without a common validation protocol and runtime measurements. The particularly poor arrival rate of W2 also warrants a broader examination of its radius–performance trade-off.

Questions for authors
  • How sensitive are the results to the number of adversarial updates and to initialization? Could you report constraint violations measured on fresh batches and the improvement obtained by running a substantially stronger inner optimizer?

  • How were the radii, risk-score mappings, and iteration budgets selected for each method? Do the main advantages persist when all methods are tuned on the same validation criterion and compared at similar wall-clock budgets?

  • When the nominal sampler and adversarial family have different architectures or latent interfaces, how should the radius be chosen to ensure feasibility? Could you demonstrate this setting and compare independent sampling with the shared-randomness setting?

  • Figure 3 provides an informative example of trajectory smoothness. Does this advantage hold across the full navigation task set, and does GDRO retain its advantage against a W2 adversary with a simple velocity or acceleration constraint?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1FLz (1/3): Weakness 1 and Q1

Response to Reviewer 1FLz (1/3): Weakness 1 and Q1

Thank you for the careful reading and the constructive questions. Each of them led to an experiment or a result that is now in the revised paper (changes in color). We answer in your order: Weakness 1 and Q1 here, Weakness 2 and Q2 in part 2, Q3 and Q4 in part 3.

Weakness 1. Theorem 3.3 assumes an exact inner maximizer (Eq. 26), so it gives limited assurance about the adversaries found in the experiments.

Answer: we agree about what Theorem 3.3 provides by itself. The revision adds the assurance that applies to the adversaries actually computed (Point 3 of our comment to all reviewers).

  • What Theorem 3.3 is (new paragraph after the theorem). Solving the inner problem exactly presupposes at least an accurate evaluation of its expectations. The theorem therefore states what the method attains when compute is abundant, the customary idealization: the original GAN analysis proves convergence for an optimal discriminator at every step, while its algorithm takes kk gradient steps (Goodfellow et al., 2014).

  • What holds for any computed adversary. Eq. (21) of Theorem 3.1 involves no optimizer: an adversary with expected mini-batch score ss changes the expected loss of every decision by at most Lfs+εlog⁡ML_f\sqrt{s+\varepsilon\log M}.

  • What an inexact inner step costs (new Appendix A.7, Proposition A.9). If the inner steps fall short of the Lagrangian maximum, every bound of Theorem 3.3 holds with Bμmax⁡/KB\mu_{\max}/\sqrt K replaced by Bμmax⁡/K+δˉB\mu_{\max}/\sqrt K+\bar\delta, with δˉ\bar\delta the expected shortfall averaged over its dual steps: the guarantees degrade linearly in the inner error.

Q1(a). Sensitivity to the number of adversarial updates.

Answer: more adversarial updates bring little on M5 and no improvement in navigation outcomes. We multiplied the number JJ of adversarial steps per dual update by 3, 5 and 10 (20 navigation tasks and 20 M5 contexts, seed 0, same initialization, all else unchanged; Tables 9 and 11).

JJNavigation: collisionsArrivalsEval. lossTasks with Sˉi>ρi\bar S_i > \rho_iM5: avg. costContexts with Sˉi>ρi\bar S_i > \rho_i
1×020/2016.882/201385.3310/20
3×020/2018.052/201382.1019/20
5×120/2022.196/201380.9118/20
10×120/2023.233/201378.3714/20

Arrivals stay at 20/20 with 0 or 1 collisions, the navigation evaluation loss rises with more updates, and M5 cost changes by 0.2% to 0.5%.

Q1(b). Sensitivity to initialization.

Answer: the results are stable across seeds, and also when the adversary starts from a different pretrained model.

  • GDRO's adversary is not initialized at random. In the main experiments it starts at the nominal generator, where the certificate is zero (Corollary A.4). What varies between runs is the sampling in the optimization: across five seeds, navigation gives 7.2 ± 1.6 collisions and 199.4 ± 0.9 arrivals of 200, and M5 an average cost of 1211.95 ± 1.34 (Tables 3 and 4).
  • In the experiment for your Q3, the adversary starts instead from a separately trained GAN aligned to nominal samples. Average M5 cost is 1213.18, against 1213.66 for the same-architecture run of Table 10 (seed 0; the aligned-GAN run also changes the architecture, the adversarial learning rate and the radius range, see part 3).

Q1(c). Constraint violations measured on fresh batches.

Answer: the estimated mean score is below 1.1ρi1.1\rho_i on 97.0% of navigation tasks and 98.5% of M5 contexts (Point 2). We froze each final adversary and evaluated it on 100 fresh batches (all 200 tasks or contexts, seed 0; shared latent noise within a batch, as in training; Figures 4 and 8).

Benchmarkmean within ρi\rho_iexcess below 5%5% to 10%10% or morelargest excess
SocialGAN, M=8M=8179 (89.5%)78628.1%
M5, M=128M=128103 (51.5%)6826319.4%

By Eq. (21), a 10% excess enlarges the risk bound by less than 5%. In navigation, 19 of the 21 tasks above the radius use the minimum radius (Figure 9).

Q1(d). Improvement from a substantially stronger inner optimizer.

Answer: small on M5 and none in navigation outcomes. With ten times more adversarial steps (last row of the table in Q1a), M5 cost falls by 0.5% at ten times the runtime (33.65 s to 350.41 s per context). In navigation the evaluation loss rises and arrivals are unchanged. The number of tasks or contexts above the radius does not fall in either benchmark. More steps per update therefore bring little here.

AuthorsOpenReview ↗

Response to Reviewer 1FLz (2/3): Weakness 2 and Q2

Response to Reviewer 1FLz (2/3): Weakness 2 and Q2

Weakness 2(i). Different radius mappings, sample counts and optimization budgets; the results are hard to interpret without a common validation protocol and runtime measurements.

Answer: the selection protocol is now stated in Appendix B.1, and the revision adds runtime measurements and a comparison at matched wall-clock time (Tables 7 and 8, Figure 6).

Weakness 2(ii). GDRO's budgets reach 500 outer iterations in navigation, W2 uses 180 with a different sampling configuration.

Answer: an outer iteration is not a common unit of budget, and neither difference explains the gap.

  • Iterations. One W2 outer iteration applies 15 gradient steps to all 512 support trajectories. One GDRO outer iteration applies 15 to 40 generator steps to mini-batches of 8 (Table 13).
  • Sampling configuration. Free-Sinkhorn, a new baseline, uses W2's configuration (512 support trajectories, same optimizer settings and radius rule) under GDRO's Sinkhorn certificate. It gives 27.6 collisions and 127.0 arrivals: no better than W2 (27.2 and 136.8) and far from GDRO (7.2 and 199.4; Table 4). GDRO's own result is stable for MM from 8 to 64 (Table 12).
  • More budget for W2. On M5, 17.5 times as many W2 iterations leave its result essentially unchanged

Weakness 2(iii). W2's poor arrival rate and its radius–performance trade-off.

Answer: W2's low arrival rate persists at the smallest radius.

  • Across the task set. At seed 0, the radius rule assigns its minimum, 0.5, to 163 of the 200 tasks; W2 fails to arrive on 41 of them, and on 18 of the other 37.
  • Radius sweep (Figure 2f). On the inspected task we varied W2's radius over two orders of magnitude. At the smallest value, 0.02, the 95th-percentile acceleration of its worst cases equals the nominal one, yet the maximum is still 19.7 times nominal: an empirical W2W_2 ball lets the budget concentrate on a few support points.
  • Other remedies. Neither kinematic penalties (W2-VA, 129.8 arrivals) nor the Sinkhorn certificate (Free-Sinkhorn, 127.0) restores arrivals.

Q2(a). How were the radii, risk-score mappings and iteration budgets selected for each method?

Answer: by one procedure for the adaptive methods, without using test outcomes (Appendix B.1).

  1. Risk score. On M5, the CVaR90 of simulated costs under the nominal decision; in navigation, the estimated collision probability of the nominal plan.
  2. Radii. Each adaptive method has two risk anchors, low and high. For each of KL, W2 and GDRO, one radius is selected at each anchor using validation contexts, and a context's radius interpolates between the two according to its normalized risk score, clipped at the anchors, by method-specific rules (Appendices B.3 and B.4; released code). The ranges are in Tables 6 and 13; W2-VA and Free-Sinkhorn use W2's range. GAS-DRO keeps a fixed radius; adapting it brought no improvement in preliminary experiments (Appendix B.3).
  3. Iteration budgets. For the primal–dual methods, we monitor that the certificate first rises and is then pushed back toward the prescribed boundary by the dual updates, and that the decision stops changing substantially. GDRO gives larger budgets to higher-risk contexts.
  4. The resulting configuration is run on the full validation set to assess downstream performance.

Q2(b). Do the main advantages persist when all methods are tuned on the same validation criterion?

Answer: yes, for the tuned methods: they share one validation protocol (Q2a), and Tables 3 and 4 report that comparison. For each of KL, W2 and GDRO, the radii were selected using validation contexts, and the budgets of the primal–dual methods by the same monitoring rule. W2-VA and Free-Sinkhorn inherit W2's range and settings, so their comparison with W2 is controlled.

Q2(c). And when compared at similar wall-clock budgets?

Answer: yes. Solver time per M5 context is 0.01 s for KL, 1.7 s for W2, 32.1 s for GDRO and 42.2 s for GAS-DRO (Table 8). We reran W2 at GDRO's wall-clock time, and evaluated GDRO at reduced outer caps from the saved trajectory of its original run (200 contexts, seed 0; Figure 6, Table 7).

RunCost of the runAvg. costCVaR90
W2, 1,500 iterations (paper)1.7 s per context1234.362720.38
W2, 26,300 iterations36.4 s per context (34.95 s on 20 reference contexts)1234.222721.52
GDRO (paper)33.65 s per context (same 20 contexts)1213.352566.55
GDRO, outer cap × 0.566% of the adversary updates1213.352619.12
GDRO, outer cap × 0.340% of the adversary updates1215.392671.26

W2 is essentially unchanged at matched time, and GDRO at 30% of its outer cap still has lower average cost and CVaR90. GAS-DRO already takes longer than GDRO.

AuthorsOpenReview ↗

Response to Reviewer 1FLz (3/3): Q3 and Q4

Response to Reviewer 1FLz (3/3): Q3 and Q4

Q3(a). With different architectures or latent interfaces, how should the radius be chosen to ensure feasibility?

Answer: align the adversary to the nominal sampler first, then set the radius above the score of the aligned model.

  • With shared random inputs: fine-tune the adversary at the active context to minimize the Sinkhorn score that defines the constraint (Eq. 75), then calibrate the radius range on validation contexts from fresh-batch scores.
  • Without them (different latent interfaces): draw the two batches independently and score the aligned model in the same way. The score of two independent batches from the nominal model itself, which needs nominal samples only, gives the scale: a radius below it excludes even an exact copy of the nominal model. Increasing MM lowers that scale.

Theorem 3.1 holds under either pairing: Eq. (21) turns the expected score of the final adversary, which the audit estimates, into a bound on risk, whether or not it is within the radius.

Q3(b). Could you demonstrate this setting?

Answer: yes, on M5, and GDRO keeps its cost reduction (Appendix B.3; Table 10, Figure 7). The frozen CVAE–LSTM only supplies samples; the adversary is a conditional LSTM-GAN trained separately on the same historical data.

  • Alignment. At validation context 21, the fine-tuned GAN's mean score with shared inputs is 28.82 at M=128M=128, below that context's radius of 31.76 (Figure 7).
  • Radius. With the original range [22, 40], a validation pilot gave violations on about half of the contexts, so we used [33, 60] with the same risk mapping.
  • Result (200 test contexts, seed 0; the first two rows are archived runs, as stated in Table 10):
AdversaryRadius rangeAvg. costQ95CVaR95
none (nominal)–1232.212680.693448.16
CVAE–LSTM, same architecture[22, 40]1213.662442.963192.54
aligned LSTM-GAN[33, 60]1213.182468.183250.15
  • Feasibility. The final adversaries' mean score is within the radius on 37 of the 200 contexts and above it on 163. On average it is 6.5% above. A different architecture thus needs a larger radius and tracks it less closely.

Q3(c). Independent sampling compared with shared randomness.

Answer: with shared inputs a small batch suffices. With independent draws the finite-batch score is dominated by sampling discrepancy, so the radius must be recalibrated or MM increased (validation context 21; Figure 7).

Mean Sinkhorn scoreShared inputsIndependent draws
GAN aligned with shared inputs, M=128M=12828.821,653
Same GAN, M=1,024M=1{,}024–1,149
Nominal against nominal, M=128M=128 / 1,0241{,}0240 (identical batches)1,668 / 1,157

A radius calibrated with shared inputs is far too small under independent sampling, even for the nominal model against itself. A larger MM lowers that level at a cost: one Sinkhorn evaluation takes 11.3 ms at M=128M=128 and 519.5 ms at M=1,024M=1{,}024. Both schemes are valid, and Eqs. (11)–(12) fix the pairing together with MM. We recommend shared inputs whenever the two samplers can share them.

Q4(a). Does the smoothness advantage of Figure 3 hold across the full navigation task set?

Answer: yes, for the decisions and for the worst-case trajectories (Figure 3 of the submission is Figure 2 of the revision). On all 200 tasks and five seeds (Table 4; evaluation loss in Appendix B.4, lower is better):

MethodCollisionsNear missesArrivals (of 200)Eval. loss
W227.2 ± 3.356.8 ± 6.1136.8 ± 5.053.9 ± 3.3
W2-VA16.0 ± 2.541.6 ± 7.1129.8 ± 3.849.0 ± 1.1
GDRO7.2 ± 1.642.8 ± 3.8199.4 ± 0.922.2 ± 1.0
  • Trajectories across the task set. In the archived run on the same 200 tasks (seed 0; GDRO with M=64M=64, the last row of Table 12), the mean task-level 95th-percentile acceleration of the worst-case trajectories is 0.037 for GDRO, against 0.317 for W2 and about 0.03 for nominal samples. GDRO's is the lower one on 182 of the 200 tasks (Appendix B.4).
  • W2 with penalties. Over all tasks and five seeds, penalizing speeds and accelerations beyond the nominal 99th percentile lowers W2's mean by 72%.

Q4(b). Does GDRO retain its advantage against a W2 adversary with a velocity or acceleration constraint?

Answer: yes (row W2-VA above; its speed and acceleration penalties are in Eq. 76). The penalties lower W2's collisions and bring near misses to GDRO's level, but arrivals stay near 130, against 199.4 for GDRO. On the inspected task, the 95th-percentile acceleration is 1.152 for W2, 0.120 for W2-VA and 0.053 for GDRO (nominal: 0.037); the W2-VA plan avoids the collision but does not reach the goal (Figure 2).

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

This paper proposes Generative Distributionally Robust Optimization (GDRO), a robust decision-making framework that uses pretrained conditional generative models. The method limits adversarial distributions to a chosen family of generators and uses sample-based Sinkhorn divergence to control their proximity to the nominal distribution at the decision context, without needing access to the nominal model's likelihoods or gradients. The authors introduce an expected mini-batch formulation and a stochastic primal–dual optimization algorithm. They also establish theoretical bounds on downstream risk deviation, the Lagrangian relaxation gap, and optimization convergence under specific assumptions. In experiments on a synthetic newsvendor problem, the M5 newsvendor benchmark, and robot navigation based on SocialGAN, improvements in inventory cost and collision avoidance were observed compared to the considered baselines.

Critical strengths
  1. A practical and flexible formulation for generative DRO. This framework addresses limitations of existing generative DRO methods by using sample-based Sinkhorn divergence to control the proximity between adversarial and nominal distributions. It also provides a method for optimizing the parameters of specified conditional adversarial generative models through generated samples, without requiring the nominal model's likelihoods or gradients. This approach leverages the structural inductive biases of the chosen generator family while keeping the adversarial distribution close to the nominal model.

  2. A theoretical framework for the proposed method. The paper provides theoretical analysis of the proposed formulation and algorithm, including downstream risk bounds, the application of Lagrangian relaxation, and optimization convergence under specific assumptions. The experiments also include a robot navigation task based on the pretrained SocialGAN model, which captures spatiotemporal patterns in pedestrian trajectories. The reported results show fewer collisions and a high arrival rate, suggesting that the proposed framework can be useful for decision-making under structured spatiotemporal uncertainty.

Critical weaknesses
  1. Insufficient ablation studies to isolate individual contributions. The performance comparisons with existing baselines involve multiple factors simultaneously—most notably differences in adversarial distribution representations, divergence constraints, and optimization procedures. Therefore, it is not clear whether the claimed benefits arise from the constraints imposed on the generator family or from other design choices. More controlled experiments—such as comparing generator-constrained and unconstrained adversaries under the same Sinkhorn constraint, or incorporating trajectory smoothness constraints into the original W2W_2 framework—would help clarify the contribution of each component.

  2. Incomplete empirical validation of constraint satisfaction and comparison fairness. Although the paper reports improvements in downstream performance, experimental evidence is insufficient to verify whether the optimized adversarial distributions satisfy the prescribed expected mini-batch Sinkhorn constraints in practice. Furthermore, the comparison involves different optimization budgets across baselines; in particular, GDRO uses a risk-adaptive budget in the M5 experiments. To strengthen confidence in the experimental conclusions, the authors should report constraint violation metrics evaluated using fresh, independently sampled batches and re-examine the performance comparisons under matched computational budgets or wall-clock time.

Questions for authors
  1. Effect of the generator-family constraint. Can the authors provide an ablation study that compares adversaries with and without the generator-family constraint under the same Sinkhorn constraint? It would also be interesting to see whether GDRO's advantage over W2W_2 remains when trajectory smoothness regularization is introduced into the W2W_2 baseline. Such comparisons would help clarify the contribution of each component.

  2. Constraint satisfaction. Can the authors report the expected mini-batch Sinkhorn divergence of the optimized adversarial distributions, estimated using fresh, independent batches? To be specific, how often and by how much does the estimated divergence exceed the prescribed radius ρ\rho?

  3. Computational budget and comparison fairness. In the M5 experiments, GDRO uses adaptive optimization budgets, whereas other methods use fixed budgets. Can the authors explain the computational costs of each method and compare their performance under matched wall-clock time or comparable computational budgets? Would the reported improvements remain under these settings?

  4. Practical implementation considerations. The theoretical analysis assumes that the inner Lagrangian function is globally maximized, whereas the proposed algorithm uses only a small number of stochastic gradient steps. How sensitive are the results to the number of inner optimization steps? Can the authors provide empirical convergence diagnostics to better justify this approximation?

  5. Connection to prior work. Can the authors explain the major methodological differences between GDRO and GAS-DRO, beyond just the choice of distributional constraint? To be clear, why is GAS-DRO classified in Table 1 as a method that does not support arbitrary frozen nominal samplers, given that its reconstruction-based constraint can also be estimated using samples?

3 comments
AuthorsOpenReview ↗

Response to Reviewer DyTE (1/3): Q1 and Q2 (Weaknesses 1 and 2)

Response to Reviewer DyTE (1/3): Q1 and Q2 (Weaknesses 1 and 2)

Thank you for the careful review. Each of your questions led to a new experiment or analysis, now in the revised paper (changes in color). Your two weaknesses are the subject of Q1 and of Q2–Q3, so we answer them under the questions, in your order: Q1–Q2 here, Q3–Q4 in part 2, Q5 in part 3.

Q1 and Weakness 1. Effect of the generator-family constraint; trajectory smoothness in the W2W_2 baseline.

Answer: the gain comes from the generator family. It disappears when the family is removed under the same Sinkhorn constraint, and GDRO keeps its advantage when W2W_2 is given smoothness penalties. We ran both controlled experiments you propose on the same 200 navigation tasks and five seeds (Table 4). This is Point 1 of our comment to all reviewers: a plausible worst case requires a generator as the adversary.

  • Free-Sinkhorn (generator-constrained against unconstrained, same Sinkhorn constraint). It keeps GDRO's Sinkhorn certificate and radius rule and removes the generator: the adversary moves trajectory support points directly. It is the W2W_2 baseline with its transport budget replaced by GDRO's certificate, with the same support size (512 trajectories) and optimizer settings (Table 13).
  • W2-VA (smoothness in W2W_2). It adds penalties on pedestrian speed and acceleration to the W2W_2 adversary, with thresholds at the nominal 99th percentiles (Eq. 76).
MethodAdversaryCollisionsNear missesArrivals (of 200)Eval. loss
Nominalnone21.2 ± 1.3106.2 ± 2.5200.0 ± 0.035.0 ± 0.6
W2free support, W2W_2 budget27.2 ± 3.356.8 ± 6.1136.8 ± 5.053.9 ± 3.3
W2-VAfree support, W2W_2 budget, smoothness penalties16.0 ± 2.541.6 ± 7.1129.8 ± 3.849.0 ± 1.1
Free-Sinkhornfree support, Sinkhorn budget27.6 ± 4.361.6 ± 5.4127.0 ± 7.057.6 ± 4.1
GDROgenerator family, Sinkhorn budget7.2 ± 1.642.8 ± 3.8199.4 ± 0.922.2 ± 1.0

The evaluation loss is the composite score of Appendix B.4 (lower is better).

  • Without the generator, same Sinkhorn constraint: collisions rise from 7.2 to 27.6, arrivals fall from 199.4 to 127.0, and the evaluation loss rises from 22.2 to 57.6. Changing only the divergence (W2 to Free-Sinkhorn, same support points and optimizer) brings no gain, so the Sinkhorn certificate alone does not produce GDRO's result. The gain appears when the generator family is added.
  • With smoothness in W2W_2: collisions fall to 16.0 and near misses reach GDRO's level, but arrivals stay near 130. GDRO keeps its advantage (7.2 against 16.0 collisions, 199.4 against 129.8 arrivals).
  • Why (Figure 2): on the inspected task the 95th-percentile acceleration of the worst-case trajectories is 0.053 for GDRO (nominal: 0.037), 1.152 for W2W_2 and 0.120 for W2-VA, whose plan avoids the collision but does not reach the goal. Over all 200 tasks (seed 0; GDRO at M=64M=64), the mean of this statistic is 0.037 for GDRO against 0.317 for W2W_2 (Appendix B.4).

Q2 and Weakness 2 (constraint satisfaction). The expected mini-batch Sinkhorn divergence on fresh, independent batches: how often and by how much is ρ\rho exceeded?

Answer: the estimated mean is within the radius on 89.5% of navigation tasks and 51.5% of M5 contexts, and below 1.1ρ1.1\rho on 97.0% and 98.5%; the largest excesses are 28.1% and 19.4% (Figures 4 and 8). This is Point 2: a budget the user sets requires a constraint that holds.

For every task and context (all 200, seed 0) we froze the final adversary and evaluated the Sinkhorn divergence on 100 fresh batches drawn independently of training. Within a batch the two generators share latent noise, which is the coupling that defines the constraint (Eq. 11). The table classifies tasks by the estimated mean Sˉi\bar S_i relative to the radius ρi\rho_i.

BenchmarkSˉi≤ρi\bar S_i\le\rho_iexcess below 5%5% to 10%10% or morelargest excess
SocialGAN, 200 tasks, M=8M=8179 (89.5%)78628.1%
M5, 200 contexts, M=128M=128103 (51.5%)6826319.4%
  • Individual batches: 25% of the fresh batches exceed the radius on a 20-task navigation subset and 43% on a 20-context M5 subset (Tables 11 and 9).
  • Where the excess occurs: in navigation, 19 of the 21 tasks above the radius use the minimum radius (Figure 9).
  • What an excess costs: by Eq. (21) of Theorem 3.1, an adversary with expected mini-batch score ss changes the expected loss of any decision by at most Lfs+εlog⁡ML_f\sqrt{s+\varepsilon\log M}, whatever optimizer produced it. A 10% excess therefore enlarges the risk bound by less than 5%.
AuthorsOpenReview ↗

Response to Reviewer DyTE (2/3): Q3 and Q4

Response to Reviewer DyTE (2/3): Q3 and Q4

Both questions concern the practical algorithm under limited compute, which is Point 3 of our comment to all reviewers.

Q3 and Weakness 2 (comparison fairness). Cost of each method; performance under matched wall-clock time; do the improvements remain?

Answer: yes, the improvements remain. We measured the cost of every method, reran W2W_2 at GDRO's wall-clock time, and evaluated GDRO at a fraction of its own budget. GDRO's advantage persists in both directions.

Cost of each method. Solver time per context on 50 M5 test contexts at the settings of the paper (seed 0; Table 8):

MethodMean (s)Median (s)90th percentile (s)
KL0.010.010.01
W21.71.81.8
GAS-DRO42.242.342.5
GDRO32.125.258.2

GDRO costs more than W2W_2 and less than GAS-DRO. Its time varies because its budget adapts to the risk score and it stops early on 28 of the 50 contexts. KL only reweights the nominal samples, and GAS-DRO already uses more time than GDRO, so W2W_2 is the baseline for which a matched budget matters.

Matched wall-clock. We raised W2W_2's outer iterations from 1,500 to 26,300, a number chosen from runtime alone, and reran all 200 contexts. We also evaluated GDRO at reduced outer caps; these values are read from the saved trajectory of the original run, whose updates do not depend on the cap (seed 0; Figure 6, Table 7).

Run (200 contexts)Cost of the runAvg. costCVaR90
W2, 1,500 iterations (paper)1.7 s per context1234.362720.38
W2, 26,300 iterations36.4 s per context (34.95 s on 20 reference contexts)1234.222721.52
GDRO (paper)33.65 s per context (same 20 contexts)1213.352566.55
GDRO, outer cap × 0.566% of the adversary updates1213.352619.12
GDRO, outer cap × 0.340% of the adversary updates1215.392671.26

W2W_2 is essentially unchanged with 17.5 times as many iterations, and GDRO at 30% of its outer cap still has lower average cost and CVaR90 than W2W_2 at matched time. The improvement is therefore not explained by GDRO's larger compute budget.

Q4. Sensitivity to the number of inner optimization steps; empirical convergence diagnostics.

Answer: more inner steps bring little on M5 and no improvement in navigation outcomes, and the final adversaries meet the budget closely on fresh batches. You are right that Theorem 3.3 assumes a global maximizer of the inner Lagrangian (Eq. 26). Solving the inner problem exactly presupposes at least an accurate evaluation of its expectations, so the theorem describes what the method attains when compute is abundant; with limited compute, the algorithm can only take a few stochastic gradient steps per update. The revision examines the gap between the two in three ways.

1. Sensitivity to inner steps. We multiplied the number JJ of adversarial gradient steps per dual update by 3, 5 and 10 on 20 navigation tasks and 20 M5 contexts (seed 0, same initialization), and audited each final adversary on 100 fresh batches (Tables 9 and 11).

JJNavigation: collisionsArrivalsEval. lossTasks with Sˉi>ρi\bar S_i > \rho_iM5: avg. costContexts with Sˉi>ρi\bar S_i > \rho_i
1×020/2016.882/201385.3310/20
3×020/2018.052/201382.1019/20
5×120/2022.196/201380.9118/20
10×120/2023.233/201378.3714/20

In navigation, arrivals stay at 20/20 with 0 or 1 collisions, and the evaluation loss rises with more steps. On M5, ten times more steps lower cost by 0.5% at ten times the runtime (33.65 s to 350.41 s per context). More inner steps thus buy little on M5 and nothing in navigation outcomes, and they do not consistently improve constraint satisfaction.

2. Convergence diagnostics. The iteration budgets were chosen by monitoring two signals (Appendix B.1): that the certificate first rises and is then pushed back toward the prescribed boundary by the dual updates, and that the decision stops changing substantially. For the solutions returned, the fresh-batch audit of Q2 is the check: the final adversaries' mean certificate is below 1.1ρ1.1\rho on 97.0% of navigation tasks and 98.5% of M5 contexts (Figures 4 and 8). Neither certifies inner optimality; see item 3.

3. Theory for an inexact inner step (new Appendix A.7, Proposition A.9). If the inner steps fall short of the Lagrangian maximum, every bound of Theorem 3.3 holds with Bμmax⁡/KB\mu_{\max}/\sqrt K replaced by Bμmax⁡/K+δˉB\mu_{\max}/\sqrt K+\bar\delta, with δˉ\bar\delta the expected shortfall averaged over its dual steps. The guarantees thus degrade linearly in the inner error. Theorems 3.1 and 3.2 do not involve the inner solver: Eq. (21) bounds the loss change under whatever adversary the algorithm returns in terms of its own expected score, which Q2 estimates.

AuthorsOpenReview ↗

Response to Reviewer DyTE (3/3): Q5

Response to Reviewer DyTE (3/3): Q5

Q5. The major methodological differences between GDRO and GAS-DRO, beyond the choice of constraint; why Table 1 classified GAS-DRO as not supporting arbitrary frozen nominal samplers.

Answer: on Table 1 you are right, and we have corrected it. The methodological differences are which models can serve as the adversary, what the budget certifies for the decision, and where the budget is evaluated. The comparison is now in the paper: revised Table 1 and introduction, and new Appendix C.1 with Table 14.

Table 1. A frozen nominal sampler can supply the data on which GAS-DRO's reconstruction constraint is estimated, as you say, so GAS-DRO should have a check mark in that column. The model-specific access is required on the adversarial side. The revised Table 1 separates the two: "Nominal used through samples" (GAS-DRO: yes) and "Implicit generator as adversary" (GAS-DRO: no; GDRO: yes).

Methodological differences.

GAS-DROGDRO
Nominal modelany frozen samplerany frozen sampler
Adversarydiffusion model or VAE decoderany generator, implicit ones included
The budget is placed onthe adversarial model's own training loss on nominal data (diffusion or VAE reconstruction)the Sinkhorn divergence between output samples of the two generators
What the budget impliesa bound on an inclusive KL divergence, up to model-dependent termsa bound on the change in expected loss of every decision
Where it is evaluatedover the data distribution, without a decision contextat the decision context
Adversary updatelikelihood-based, no gradient through the samplegradient through the generated samples

1. Which models can serve as the adversary (Point 1 of our comment to all reviewers). Your remark about sample estimation pinpoints the distinction. In its VAE construction, GAS-DRO passes a nominal observation through a fixed encoder and scores the candidate decoder by its reconstruction loss. Sampling supplies the observation; the score also needs the candidate's reconstruction mechanism. Two decoders with the same output law can therefore receive different scores. Take a standard Gaussian latent zz and two unit-noise Gaussian decoders with means +z+z and −z-z. Both generate N(0,2)N(0,2), yet with the encoder fixed to the first decoder's posterior their expected squared reconstruction errors are 1 and 5. Consistently, the GAS-DRO authors state that a principled design based on GANs is not currently available in their framework (GAS-DRO, Sections 4–5 and Appendix B; authors' response). GDRO's budget uses output samples only (Eq. 11, which also fixes how the two samplers' draws are paired within a batch). SocialGAN can therefore supply the nominal trajectories and generate the adverse ones.

2. What the budget certifies. GAS-DRO's Lemma 1 relates its budget to an inclusive KL divergence up to model-dependent terms, and its Corollary 1 bounds the error of its inner maximization under an expressivity assumption. In GDRO, with bounded outputs, squared transport cost and an LfL_f-Lipschitz loss, every generator admitted by the budget changes the expected loss of every candidate decision by at most Lfρ+εlog⁡ML_f\sqrt{\rho+\varepsilon\log M} (Theorem 3.1). When a feasible generator reproduces the nominal law, the same quantity bounds the gap between the constrained problem and its multiplier relaxation (Theorem 3.2).

3. Where the budget is evaluated. GAS-DRO is formulated without decision contexts. In our conditional extension of it, the reconstruction budget is an average over training pairs, so a large discrepancy at a rare context can contribute little to it. GDRO evaluates both the worst-case loss and the budget at the context where the decision is made. Table 2 compares the two implementations: on rare inventory contexts, mean regret is 9.26 for GDRO and 16.79 for the conditional GAS-DRO extension, against 25.09 for nominal planning.

Comments on the paper

2 comments
AuthorsOpenReview ↗

To all reviewers: what this paper offers

To all reviewers: what this paper offers

We thank the three reviewers. Your questions bear directly on what this paper offers, so we begin there.

What GDRO offers is practical: take the model your domain prefers as the reference, choose any pretrained generator as its adversary, set a budget, and obtain a decision that is robust to plausible errors of the reference. To our knowledge it is the first DRO framework that allows this. The original experiments, now augmented with more comprehensive and detailed ablation studies, show clearly that it works, and why.

In particular, we would like to direct your attention to Figure 2, now in the main text, which shows most clearly what the pretrained generator contributes: GDRO's worst cases remain plausible, whereas a transport-based adversary can still fail to produce plausible trajectories, even with penalties on speed and acceleration. Alongside it, Figures 4 and 8 (how closely the budget is met) and Table 4 (what the generator contributes to the decision) are, in our view, the most important additions in this revision. Changes are marked in color.

Each ingredient of GDRO is there because its practical use requires it, and each of your questions tests one of these requirements.

Point 1. A plausible worst case requires a generator as the adversary (Figure 2, Table 4). A free adversary is not bound to what the domain can produce, and the decision pays for it; a pretrained generator carries that knowledge. To admit any such generator, GDRO compares output samples only: it needs no likelihood or model-specific loss, so the generator can be implicit. The closest prior work leaves this case open: the GAS-DRO authors state that they "cannot currently provide a principled GAS-DRO design based on GANs" (authors' response).

  • With and without the generator (DyTE Q1, 1FLz Q4, cS5v). Under the same Sinkhorn certificate, free support points in place of the SocialGAN family raise collisions from 7.2 to 27.6 and lower arrivals from 199.4 to 127.0 of 200. In Figure 2, the 95th-percentile acceleration of the worst cases is 0.053 for GDRO (nominal: 0.037), against 1.152 for W2W_2 and 0.120 for W2W_2 with penalties. Over all 200 tasks (seed 0; GDRO at M=64M=64), the mean of this statistic is 0.037 for GDRO against 0.317 for W2W_2.
  • A generator of a different architecture (1FLz Q3). With a frozen CVAE–LSTM as the reference and a separately trained GAN as the adversary, GDRO keeps the cost reduction (1213.18 against 1232.21 for the nominal decision; Table 10), at a larger radius and with more residual violations.

Point 2. A budget the user sets requires a constraint that holds (Figures 4 and 8). P-DRO, DRAGEN and FlowDRO optimize a penalized relaxation of their divergence constraint. GDRO, like the GAS-DRO line, keeps the stated budget as a constraint and adapts a multiplier to meet it, and it measures that budget from the outputs of whatever generator was chosen (cS5v, DyTE Q5; new Appendix C). Setup, algorithm and analysis all address the constrained problem.

  • Does it hold on fresh samples? (DyTE Q2, 1FLz Q1) The final adversaries' mean score on fresh batches is below 1.1 times the radius on 97.0% of the 200 navigation tasks and 98.5% of the 200 M5 contexts, and within the radius on 89.5% and 51.5%.

Point 3. Theoretical analysis versus the practical algorithm (DyTE Q3–Q4, 1FLz Weakness 1 and Q1–Q2, cS5v). The analysis gives the assurance for abundant compute, the algorithm is what limited compute requires, and the experiments connect the two.

  • The analysis. Theorem 3.3 assumes that the inner problem is solved exactly. That presupposes, at the least, accurate evaluation of the expectations defining the worst-case loss and the budget, which takes very many samples. This is the standard idealization for assurance in analysis, and it states what the method attains when compute is not the limit. New Appendix A.7 extends it: with an inexact inner solution the bounds degrade only linearly in the inner error (Proposition A.9), and convexity can be replaced by smoothness (Theorem A.11).

  • The algorithm. With limited compute, we take a few mini-batch gradient steps. Theorem 3.1 bounds the shift in expected loss for any generator, even before optimization. Theorem 3.2 bounds the relaxation gap if a feasible generator matches the nominal law.

  • What the experiments show. The budget is closely met on fresh samples (Point 2), and three to ten times more adversarial steps change M5 cost by 0.2% to 0.5% and bring no improvement in navigation outcomes (Tables 9 and 11).

Dear Reviewers,

As today is the last day of the rebuttal period, please take this opportunity to review the authors’ responses and engage in further discussions with them if needed.

Thank you very much for your time and contributions!

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.