- 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
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 -Lipschitz loss, every generator within budget changes the expected loss of every candidate decision by at most .
- 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 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 .
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 added to (new Appendix A.7, Proposition A.9). We do not claim that a fixed number of gradient steps makes 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 -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 outer iterates, is (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.
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.
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.
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.
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).
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+εlogM, the bound of Theorem 3.1, when a feasible generator reproduces the nominal law (Theorem 3.2; see part 2).