acceptodds
Under review as a conference paper at ICLR 2027

Solving parametric variational inequalities via flow matching

Abstract

Variational inequalities (VIs) provide a unified framework for modeling optimization, equilibrium, and multi-agent decision problems. In many applications, the VI depends on exogenous parameters and must be solved repeatedly across varying instances, giving rise to parametric variational inequalities (PVIs). When solution sets are non-singleton, a single prediction can discard alternatives useful for downstream criteria such as welfare or robustness. To address this, we propose a conditional flow-matching approach that learns distributions over PVI solutions, enabling solution sampling and post-hoc selection for unseen instances. Theoretically, we establish VI-specific guarantees that translate flow-matching error into Wasserstein distributional, feasibility, and stationarity errors, together with finite-sample high-probability bounds that account for approximation, optimization, and generalization errors. Empirically, we evaluate the framework on three complementary benchmarks: Braess traffic equilibria, where it recovers diversity along a non-singleton equilibrium set; nonconvex AC optimal power flow, where it exhibits near-theory-consistent feasibility-error scaling as an out-of-assumption stress test and competitive time–accuracy tradeoffs; and non-monotone bimatrix games, where best-of- and welfare analyses evaluate flexible equilibrium selection.

Primary area:
optimization
Keywords:
variational inequalities, generative modeling, flow matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MLMVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a flow-matching framework for solving parametric variational inequalities. This is motivated by applications in which related VI instances must be solved repeatedly. The framework relies on learning a conditional distribution over solutions. The paper provides a theoretical analysis, showing that the proposed approach is theoretically grounded. In experiments across several benchmarks, the proposed method shows promising performance in terms of preserving solution diversity.

Critical strengths
  • The paper is well motivated and addresses an important problem. There are many applications in which the goal is to solve many related variational inequalities. Instead of solving each instance from scratch, the basic goal is to use some form of learning to speed up optimization when there's structure in between instances. This arises, for example, in various routing problems, where some of the parameters of the network may be changing slowly.

  • The approach put forward by the paper is very natural and promising. Flow matching has been a powerful paradigm for various problems. The way it's used in this paper is, to my knowledge, new. I like the fact that the approach is theoretically grounded, and the experiments provide a good sanity check. I'm not an expert on flow matching, but I learned much from reading this paper. Overall, the paper brings together ideas from two different lines of work, and can motivate a lot of interesting future research.

  • The paper is overall well written. The main ideas are clear and easy to follow. It also does a good job at motivating the problem. The paper is also well structured.

Critical weaknesses

My main criticism is that the experiments could go much further. First, the paper could do much more to show the scalability of the approach. Several of the benchmarks are really toy instances. For example, in Section 5.3, the largest bimatrix game is still quite small. The same criticism applies to Section 5.1. While these results are a good sanity check, I'm not convinced that the method would scale to much larger instances. The authors could address this limitation by conducting experiments to much larger instances. For example, in the context of Section 5.3, the authors could try to increase the dimensions in the thousands, so that the number of entries of each payoff matrix is in the millions.

The second issue in the experiments is the lack of good comparisons. For example, staying on Section 5.3, I don't think that comparing against algorithms like optimistic mirror descent or the extra-gradient method is meaningful; these are not methods designed to find Nash equilibria in general-sum games. On the other hand, there are many algorithms that are guaranteed to converge to exact Nash equilibria and are quite scalable. For example, see "Spatial branch-and-bound for computing multiplayer Nash equilibrium" and references therein. The paper doesn't compare against any of those methods, and in fact, in these small instances, these exact Nash solvers would converge instantly.

Overall, the experiments are not convincing enough. What I wanted to see is a class of realistic problems where using exact VI solvers fails to scale, but the proposed method is able to find good approximate solutions much faster.

Questions for authors
  • Using the term "optimistic gradient descent/ascent" makes sense in zero-sum games, but in bimatrix games. For example, if it's an identical interest game, both players are performing optimistic gradient ascent.
  • There's a closely related line of work that I strongly recommend should be discussed, which was started by Harris et al. ("Meta-Learning in Games") and Sychrovský et al. ("Learning Not to Regret"), and there have been some follow ups ever since. Even though the techniques are different, the motivation is very similar. The authors could also try to experiment in some of the game classes analyzed in those papers.
3 comments

We thank the reviewer for the encouraging assessment and the concrete suggestions, which shaped our new main experiment (Sec. 5.3). Below we respond point by point. Changes in the revised PDF are in blue, and all section, table, figure and lemma numbers refer to it.

W1 (scale). The benchmarks are toy-sized (Secs. 5.1 and 5.3); try games with thousands of strategies, i.e., payoff matrices with millions of entries.

W2 (baselines). Optimistic mirror descent and extragradient are not designed for general-sum Nash equilibria; compare with exact Nash solvers such as spatial branch-and-bound.

W3 (realistic regime). Show a realistic class where exact solvers fail to scale but the method finds good approximate solutions much faster.

Response: We agree on all three points and replaced the bimatrix games with a large electricity-market bidding game built from public data (new Sec. 5.3; the game and a table of results are also summarized in our response to Reviewer PZtA, and full details are in App. C.3).

Scale. With two sellers, each seller has up to n=1000n=1000 strategies, so each payoff matrix has up to 10610^6 entries as suggested. With three sellers, an instance has 1.1×1071.1\times10^7 (n=225n=225) or 10910^9 (n=1000n=1000) pure profiles, and payoffs are available only through a market-clearing simulator. The Braess network (Sec. 5.1) stays small on purpose, since its equilibrium set is known in closed form.

Baselines. We now compare with Nash solvers: Lemke–Howson, MIP-Nash and double oracle with two sellers; with three sellers, spatial branch-and-bound (Černý et al., "Spatial Branch-and-Bound for Computing Multiplayer Nash Equilibrium", 2026), Gambit's global Newton, logit-tracing and simplicial-subdivision solvers, and the nonconvex Nash model solved by Gurobi. We also compare with ADIDAS (Gemp et al., "Sample-based Approximation of Nash in Large Many-Player Games via Gradient Descent", 2022), best-response dynamics, random profiles and nearest-interval retrieval, and, since our method is a general VI method, with two general VI solvers (stochastic mirror-prox and projected extragradient) on the game's mixed extension (App. C.3). The reviewer's expectation holds when the matrices are available: with two sellers, Lemke–Howson is at least as fast as our sampler and finds more equilibria at every size (App. D.2). However, the picture reverses at scale. Details are in the next paragraph.

A realistic class where exact solvers fail. The three-seller game is such a class. At n=225n=225, complete solvers need the full tensor (1.1×1071.1\times10^7 clearings, i.e., market-clearing simulator calls), and within one hour only Gambit's global Newton method returns an equilibrium; spatial branch-and-bound returns none. With about 2% of those market-clearing simulator calls and 0.2 s per interval, our sampler with best-response repair certifies 24.1 equilibria per interval (i.e., game instance), against 16.5 for best-response dynamics, 16.7 for ADIDAS and 20.6 for the VI solvers, which use about twice our clearings. Trained toward the regulator's criterion (i.e., an equilibrium-selection criterion), it attains the least-consumer-surplus-loss equilibrium on 81% of intervals (best-response dynamics: 73%, paired gain +8.7+8.7 points, 95% CI [3.3,14.3][3.3,14.3]) and on 70% at n=1000n=1000, more often than every baseline at both sizes.

Q1. The term "optimistic gradient descent/ascent" fits zero-sum games but not general bimatrix games.

Response: We agree. That baseline belonged to the random bimatrix experiment, which we removed, so the term no longer appears in the paper.

Q2. Discuss the meta-learning-in-games line of work (Harris et al.; Sychrovský et al.) and possibly experiment on their game classes.

Response: Thank you for pointing us to this line of work. We now cite Harris et al. ("Meta-Learning in Games", 2023), Sychrovský et al. ("Learning Not to Regret", 2024) and their follow-up, Sychrovský et al. ("Meta-Learning in Self-Play Regret Minimization", 2025), in Sec. 2 and discuss them in App. E. Like these works, we exploit structure shared across related instances, but they learn game-specific equilibrium-finding dynamics (faster no-regret learning), whereas we learn, for general VIs, a conditional distribution over solutions, which yields many candidates per instance and allows selection, e.g., by consumer-surplus loss in our bidding game. Their experiments use two-player zero-sum games (matrix games and poker), where all equilibria share one value, so we focused our experiments on a general-sum game with many equilibria of different quality.

We hope these responses address the reviewer's concerns, and we would be happy to continue the discussion.

I thank the authors for the response. The new experiments are a strong addition to the paper. I have increased my score to a 3.

We thank the reviewer for the reply and for raising the score. The reviewer's suggestions shaped the new main experiment, and we are glad the reviewer finds it a strong addition. If any concerns remain, we would be happy to address them before the discussion period ends

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

This paper proposes to approximate parametric variational inequalities using transport theory and flow matching in particular. The key is to use a neural network to approximate the velocity field to drive an original random probability map into the desired one that represents the solution to a variational inequality. The paper connects ¯the fow-matching error to Wasserstein distributional accuracy, approximate feasibility, and VI gap-based stationarity. It provides finite-sample guarantees that account for model approximation, optimization, and generalization errors. Experimental results on game theory (Braess paradox and multi-player games) and AC-OPF highlight the potential of the approach.

Critical strengths

-- interesting topic -- both theoretical and experimental analysis

Critical weaknesses

-- experimental results are not demonstrating the benefits of the proposal at scale.

Questions for authors

N/A

2 comments

We thank the reviewer for the positive assessment of the topic and of our combination of theory and experiments. Below we respond to the reviewer's comment. Changes in the revised PDF are in blue, and all section, table, figure and lemma numbers refer to it.

Critical weakness. Experimental results are not demonstrating the benefits of the proposal at scale.

Response: We agree that the submitted experiments did not show the benefits at scale. We therefore replaced the random bimatrix games (old Sec. 5.3, whose largest case has only 50 strategies per player) with a large and realistic game (new Sec. 5.3). It is an electricity-market bidding game built from public data published by ERCOT, the Texas grid operator. Strategic bidding in such markets is a classic game in economics (Hortaçsu & Puller, "Understanding Strategic Bidding in Multi-Unit Auctions: A Case Study of the Texas Electricity Spot Market", 2008). Each of about 2100 five-minute market clearings is one game (VI) instance, and sellers choose price markups and capacity withholding, the classic market-power concern. Full details of the game (data, sellers (i.e., players) and strategies, market clearing, payoffs and the VI formulation) are in App. C.3. With three sellers an instance has 1.1×1071.1\times10^7 (n=225n=225 strategies per seller) or 10910^9 (n=1000n=1000 strategies per seller) pure profiles, and payoffs are available only through a market-clearing simulator, rather than given as a payoff tensor. Every method is charged for each simulator query (clearing) it makes.

The results show where our method helps. When the payoff matrices are small enough to write down (two sellers, up to 10610^6 entries each), exact solvers are the better tool: compared with our method, Lemke–Howson is faster and finds more equilibria (App. D.2, Table 10). We present part of the results in Table 10 here:

Two sellers, 16 candidates per methodn=100n=100n=400n=400n=1000n=1000
Ours (sampling + support polish)0.19 s, 4.5 NE0.55 s, 7.4 NE2.1 s, 8.4 NE
Lemke–Howson0.05 s, 7.4 NE0.24 s, 12.0 NE1.8 s, 30.0 NE

where each entry represents the online solving time (in seconds) and the number of distinct certified equilibria per problem instance.

At scale the picture reverses. With three sellers, Lemke–Howson does not apply. At n=225n=225, complete Nash solvers need the full payoff tensor, more than 40 times our clearings, and within one hour only Gambit's global Newton method returns an equilibrium (on 3 of 5 intervals); we did not run them at n=1000n=1000, whose tensor is 88 times larger. Our sampler with best-response repair certifies the most equilibria at n=225n=225 and, at n=1000n=1000, is second only to mirror-prox, which uses 1.5 times our clearings. Trained toward the regulator's consumer-surplus loss (CSL), it attains the least-CSL equilibrium most often at both sizes (at n=225n=225, 81% vs. 73% for best-response dynamics; paired gain +8.7+8.7 points, 95% CI [3.3,14.3][3.3,14.3]). Selected results (Table 1; 64 candidates per interval, mean over 3 seeds; clearings = simulator queries; NE = distinct certified equilibria; CSL opt. = share of intervals attaining the regulator's optimum):

Three sellersn=225n=225: clearingsNECSL opt.n=1000n=1000: clearingsNECSL opt.
Ours, uniform target2.6×1052.6\times10^524.170%1.7×1061.7\times10^618.949%
Ours, low-CSL target2.2×1052.2\times10^512.481%2.7×1062.7\times10^614.070%
BR dynamics2.7×1052.7\times10^516.573%1.6×1061.6\times10^615.553%
ADIDAS5.6×1055.6\times10^516.769%2.9×1062.9\times10^616.150%
Mirror-prox4.7×1054.7\times10^520.658%2.5×1062.5\times10^619.850%
Extragradient4.9×1054.9\times10^520.660%2.7×1062.7\times10^618.758%
Gambit, global Newton1.1×1071.1\times10^73/5 solved20%not run
Other complete solvers1.1×1071.1\times10^70/5 solved0%not run

Fig. 9 (App. D.2) shows this cost–quality trade-off for every method and budget.

Besides the new experiment, we (i) restated the theory in terms of a flow-matching error that can vanish with a perfect model, perfect training and a large sample size, with explicit rates in the sample size NN (Sec. 4), and (ii) now report the offline labeling and training costs with their break-even points (App. D.2).

We hope these revisions address the reviewer's concern, and we would be happy to continue the discussion.

Dear Reviewer PZtA, as the discussion period ends today, we would like to check whether our response has addressed your concern about our method's benefits at scale, which we addressed with a new large-scale experiment on an electricity-market bidding game with up to one billion pure profiles (new Sec. 5.3 in the revised pdf). We would be glad to clarify any remaining concerns. If our revisions have resolved your concern, we would appreciate it if you would take them into account in your final evaluation.

Reviewer_UK3EOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work considers parametric VI problems. First VI problems corresponding to each parameter is solved in an offline fashion and then flow matching technique is used to estimate the underlying parameter-dependent solution distribution of the parametric VI. Then at inference time, for a new parameter, a sampling from the underlying estimated distribution yields candidate VI solutions.

Critical strengths

The paper studies the class of PVI problems, where multiple candidate solutions can be valuable for welfare and robustness. Hence, the authors tried to estimate the underlying distribution via flow matching so as to sample the candidate VI solutions corresponding to a parameter.

Critical weaknesses
  1. We are solving PVI in an offline fashion. Then, we are solving a flow matching problem and the VI technique is used to find set of VI solutions corresponding to each of the parameters. How is the guaranteed that the offline VI solver provides the exact VI solution? I am also a little confused with the title.

  2. I do not understand the usefulness of the bounds stated in Theorems 4.1(a), 4.2(a) and 4.3(a). I believe these are some intermediate results that are useful to proceed further. Please clarify how these bounds are important on their own?

  3. I do not understand the sample complexity bounds in the theorems. The theorems consider the condition EN,δ/2≤1\mathcal{E}_{N,\delta/2} \leq 1 and the sample complexities depend on this term which is also linked to the quantity ηD\eta_D. There is no guarantee that this quantity will be small. Please clarify how any of this sample complexity bounds establish a convergence behavior?

  4. The PVI are solved in an offline fashion and then the flow-matching algorithm is performed. What is the guarantee that the offline computed PVI solutions are exact? I read the section on VI stationarity guarantee, but it is no where clearly explained. Also the complete method should be treated end to end starting from the offline VI computation till flow-matching and derive sample complexities for the whole process.

  5. While running the experiments, is the offline PVI compute time considered while comparing with other algorithms? Please clarify. I do not very well understand the simulation for equal-time.

  6. Is the Mean-Nash gap the best comparison index? How is the standard deviation for different sampling schemes? Is the performance improved substantially by following the flow matching approach?

  7. Please cross-check the value of the constant in the upper bound of MDM_D in line 1347 (page 25).

I would lean towards weak rejection since the sample complexity bound results are not strong.

Questions for authors

Please see the critical weaknesses section.

5 comments

We thank the reviewer for the careful reading, in particular for checking the constant in the appendix. Since the reviewer's main concern is the sample complexity results, we start with W2 and W3 and then respond to the remaining points. Changes in the revised PDF are in blue, and all section, table, figure and lemma numbers refer to it.

W2. Are the bounds in Theorems 4.1(a), 4.2(a) and 4.3(a) useful on their own, or only intermediate steps?

W3. Do the sample complexity bounds establish convergence, given the condition EN,δ/2≤1\mathcal E_{N,\delta/2}\le1 and the optimization error ηD\eta_{\mathcal D}?

Response: We agree that the submitted bounds did not show convergence, and we rewrote Theorems 4.1–4.3.

Submitted version. The bounds used the training (CFM) loss, which has an irreducible floor even with a perfect model and perfect training, and assumed EN,δ/2≤1\mathcal E_{N,\delta/2}\le1 to simplify the expressions. Part (b) held only with probability 1−δ1-\delta over ww.

Revised version. The bounds use the flow-matching error LFM\mathcal L_{\rm FM}, which, unlike the CFM loss, is zero for a model that matches the true velocity field, and the condition EN,δ/2≤1\mathcal E_{N,\delta/2}\le1 is no longer needed. Part (a) bounds the average error by LFM\mathcal L_{\rm FM} for every trained model, with no dependence on the parameter dimension pp. Since LFM\mathcal L_{\rm FM} is at most the CFM loss (Lemma B.1), the CFM loss makes (a) a computable certificate, which is why it matters on its own. Part (b) extends (a) to a random instance and, under the same assumptions as before, to every w∈Ww\in\mathcal W (Lemma B.9). Part (c) replaces LFM\mathcal L_{\rm FM} by εN,δ=AΘ+ηD+O~(N−1/2)\varepsilon_{N,\delta}=\mathcal A_\Theta+\eta_{\mathcal D}+\widetilde O(N^{-1/2}) (approximation, optimization and generalization errors). With a realizable class and training at the empirical minimum, the bounds vanish at these rates:

average over wwall ww
W22\mathbb W_2^2 (Thm. 4.1)N−1/2N^{-1/2}N−1/(2(p+1))N^{-1/(2(p+1))}
distance to Xw\mathcal X_w (Thm. 4.2)N−1/2N^{-1/2} (squared)N−1/(2(p+2))N^{-1/(2(p+2))}
VI gap (Thm. 4.3)N−1/4N^{-1/4}N−1/(2(p+2))N^{-1/(2(p+2))}

N−1/2N^{-1/2} (up to log factors) is the usual uniform-convergence rate of the generalization term (e.g., Wainwright, "High-Dimensional Statistics: A Non-Asymptotic Viewpoint", 2019). Uniformity over ww costs a dimension-dependent exponent. The optimization error ηD\eta_{\mathcal D} is not the focus of this work. Keeping it as an explicit term is common practice in learning theory, which splits the error into approximation, estimation and optimization parts and studies the last one in a separate line of work (Bottou & Bousquet, "The Tradeoffs of Large Scale Learning", 2007; Berner et al., "The Modern Mathematics of Deep Learning", 2022). With a suitable optimizer and assumptions, ηD\eta_{\mathcal D} can be made arbitrarily small by training longer: for instance, gradient descent on sufficiently overparameterized networks reaches a global minimum of the training loss (Du et al., "Gradient Descent Finds Global Minima of Deep Neural Networks", 2019; Allen-Zhu et al., "A Convergence Theory for Deep Learning via Over-Parameterization", 2019). The paragraph "Error decomposition and rates" at the end of Sec. 4 summarizes all this.

W1. How is it guaranteed that the offline VI solver returns exact solutions? The title is also confusing.

W4. Can the method be analyzed end to end, from the offline VI computation to flow matching, with sample complexities for the whole process?

Response: (i) Our analysis assumes that the offline solver returns exact solutions, i.e., the label distribution p(⋅∣w)p(\cdot\mid w) is supported on the solution set S(w)\mathcal S(w) (Sec. 3). Table 2 (App. C) in the revision records, for each experiment, how the labels are produced and whether they are exact. (ii) Our results extend to inexact solvers: if each label is within ε1\varepsilon_1 of Xw\mathcal X_w and has VI gap at most ε2\varepsilon_2, these errors add to the feasibility and stationarity bounds by the triangle inequality and Lemma B.6. End to end, the error is thus the offline solver's error plus the learning error bounded by Theorems 4.1–4.3. The number of offline solves is the sample size NN in Theorems 4.1–4.3, so the bounds also give the number of solves needed for a target accuracy. The offline cost is NN times the cost of one solve. That cost depends on the solver, and equilibrium computation is PPAD-hard in general, so we report it empirically (see W5). On the title: "solving" means amortized solving, i.e., a new instance is solved by sampling instead of running a solver from scratch.

W5. Is the offline compute time included in the comparisons, and what is the equal-time simulation?

Response: The equal-time comparison counted online solving time only, since labels and training are a one-time cost amortized over all future queries, as is standard in amortized optimization (Amos, "Tutorial on Amortized Optimization", 2023). That comparison belonged to the random bimatrix games, which we replaced by a large, realistic game built from market data, as the other two reviewers asked for larger instances (new Sec. 5.3; see also our responses to Reviewers PZtA and MLMV). There every method is charged for its market-clearing simulator queries, and our offline cost (labels and training) is reported separately (App. D.2, "Offline cost"). At n=225n=225, labels take 2.1 h and training 95 s, and the cost breaks even against exact enumeration after about 2100 new instances; at n=1000n=1000, after about 65.

W6. Is the mean Nash gap the right metric, what are the standard deviations and is the improvement substantial?

Response: We agree that the mean Nash gap alone is not a sufficient comparison index, since it hides the spread across instances and seeds. In the new large-scale game (Sec. 5.3), we report the mean Nash gap together with its std over seeds (Table 13). We find that our method's gains are substantial. At n=225n=225, the mean Nash gap of our samples is about 3.4 times lower than that of random profiles and ADIDAS, and the low-CSL model has the lowest mean gap of all methods (Table 13). Beyond the Nash gap, we now also report the number of distinct equilibria each method finds and the quality of the best one under two downstream criteria, joint profit and the regulator's consumer-surplus loss, so the comparison is more comprehensive than in the submission (Table 1; std over seeds in Tables 1, 11–13).

W7. Please check the constant in the upper bound on MDM_{\mathcal D} (line 1347).

Response: The reviewer is absolutely correct: the constant was wrong, and the correct one is C1=c (LCFMW)p/(p+1)(mWκ0κp)−1/(p+1)C_1=c\,(L^{\mathcal W}_{\rm CFM})^{p/(p+1)}(m_W\kappa_0\kappa_p)^{-1/(p+1)}. The rate was unaffected, and the revised proof of the uniform bounds no longer uses this bound (it now relies on Lemma B.9).

We hope these responses address the reviewer's concerns, in particular about the sample complexity results, and we would be happy to continue the discussion.

Dear Reviewer UK3E, as the discussion period ends today, we would like to check whether our response has addressed your concerns, in particular about the sample complexity results (rewritten Theorems 4.1–4.3, with explicit rates in Sec. 4 in the revised pdf). We would be happy to clarify any remaining concerns. If our revisions have resolved your concerns, we would appreciate it if you would take them into account in your final evaluation.

Comments on the paper

1 comment

We thank all reviewers for their feedback. We have uploaded a revised PDF with changes highlighted in blue. All section, table, figure and lemma numbers below refer to the revised PDF. We summarize the main changes and which reviewer(s) asked for them as follows:

  1. Theory (UK3E). Theorems 4.1–4.3 are restated in terms of the flow-matching error LFM\mathcal L_{\rm FM}, which, unlike the CFM loss, is zero for a model that matches the true velocity field. The condition EN,δ/2≤1\mathcal E_{N,\delta/2}\le1 is removed. Under the same assumptions, part (b) is now stronger: it holds for every instance ww rather than with probability 1−δ1-\delta over ww (Lemma B.9). The bounds vanish at explicit rates in the sample size NN, summarized in the new paragraph "Error decomposition and rates" (Sec. 4). We also corrected the constant pointed out by Reviewer UK3E.
  2. New main experiment at scale (PZtA, MLMV). The random bimatrix games (old Sec. 5.3) are removed from the paper and replaced by an electricity-market bidding game built from public data (new Sec. 5.3; details in App. C.3). With two sellers, each payoff matrix has up to 10610^6 entries; with three sellers, an instance has up to 10910^9 pure profiles, payoffs are available only through a market-clearing simulator, and every method is charged for its simulator calls.
  3. Baselines (MLMV). We now compare with exact Nash solvers (Lemke–Howson, MIP-Nash and double oracle with two sellers; spatial branch-and-bound, three Gambit solvers and a Gurobi model with three sellers), the large-game solver ADIDAS, best-response dynamics, and two general VI solvers (stochastic mirror-prox and projected extragradient). With two sellers, where the payoff matrices are small enough to write down, Lemke–Howson remains the better tool. With three sellers, the complete solvers fail within one hour, while our sampler certifies the most equilibria at n=225n=225 and attains the regulator's optimum most often at both sizes (Table 1).
  4. Metrics and statistics (UK3E). Standard deviations over seeds, the number of distinct certified equilibria, equilibrium selection by joint profit and by consumer-surplus loss with paired confidence intervals, and the Nash gap of the raw candidates (Tables 1 and 11–13).
  5. Offline cost (UK3E). Labeling and training costs, with their break-even points (App. D.2).
  6. Exactness of labels (UK3E). Table 2 (App. C) records how the labels of each experiment are produced and whether they are exact; our response to Reviewer UK3E explains how the errors of inexact labels add to the bounds.
  7. Related work (MLMV). Meta-learning in games, which the submission did not discuss, is now discussed in Sec. 2 and App. E.
  8. Further analyses (App. D.2). The role of the best-response repair, the target size of the low-CSL model, and the cost–quality frontier of every method (Fig. 9).
  9. Other. The term "optimistic gradient descent/ascent" left the paper with the bimatrix experiment (MLMV), and App. C has a new figure of the end-to-end pipeline (Fig. 6). We also added supporting lemmas in App. B, updated the introduction and related work, and corrected several citations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.