Advancing Differentiable Economics: A Neural Network Framework for Revenue-Maximizing Combinatorial Auction Mechanisms

Mai Pham; Peter Chin; Vikrant Vaze

REVIEW 3 major objections 5 minor 1 cited by

Advancing Differentiable Economics: A Neural Network Framework for Revenue-Maximizing Combinatorial Auction Mechanisms

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that gradient-trained neural networks, CANet and CAFormer, can design revenue-maximizing randomized combinatorial auctions that beat VCG and heuristic benchmarks on tested instances while keeping incentive-compatibility…

desk verdict The architectures and experiments are real, but the core feasibility lemma is unproven and likely false, so the paper's main claim about combinatorially feasible randomized CA mechanisms does not hold as written. read the letter →

arxiv 2501.19219 v1 pith:UTWTPC3B submitted 2025-01-31 cs.GT

Mai Pham , Vikrant Vaze , Peter Chin This is my paper · ORCID

classification cs.GT MSC 91B2668T07

keywords differentiableeconomicscombinatorialauctionsautomatedmechanismdesignrevenuemaximizationregretminimizationneuralnetworkrandomizedmechanismspermutation-equivariantnetworks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that deep learning, which has already produced near-optimal auctions for single items and simple multi-item settings, can be extended to combinatorial auctions where bidders value bundles of items. The authors propose two neural architectures, CANet and CAFormer, that output randomized (lottery) allocations while attempting to enforce the combinatorial constraint that no item is assigned to more than one bidder and each bidder receives at most one bundle. On test instances with two bidders and two, three, or five items, both networks report higher revenue than VCG and several heuristic mechanisms, with incentive-compatibility regret a few thousandths of revenue. The broader claim is that gradient-based training can substitute for analytically derived mechanisms in a setting where optimal designs are mostly unknown.

What carries the argument

The central object is the allocation matrix Z ∈ $R^{{n×k}}$ built from two components: an item-to-bundle probability matrix B_bundle and a bundle-to-agent probability matrix A_agent-bundle, with Z = B_bundle · A_agent-bundle. B_bundle is produced by applying a softmax over items for each bundle, masking non-positive entries, and taking the per-item minimum over bundles; A_agent-bundle is the elementwise minimum of a softmax over agents and a softmax over bundles. The construction is meant to turn the combinatorial constraints (4) and (5) into differentiable operations: the item softmax keeps each item from being over-allocated across bundles, and the agent/bundle softmaxes keep each bidder from receiving more than one bundle. The training loop then minimizes negative log-revenue plus a regret penalty computed by an adversarial inner maximization over misreports, with loss weights that are adaptively balanced and a regret budget that anneals toward zero.

What would settle it

Run either trained network on a batch of random valuation profiles, compute the output allocation matrix, and check whether for every item j the sum over bidders and over bundles containing j of z_{iS} is at most 1; the first violation refutes the feasibility lemma and invalidates the revenue comparisons.

Watch

Extended reading notes

Core claim

The paper's central claim is that CANet and CAFormer learn approximately dominant-strategy incentive-compatible, individually rational, combinatorially feasible randomized mechanisms whose revenue exceeds that of VCG, affine-maximizer auctions, and local-search baselines on every combinatorial instance tested (2×2, 2×3, 2×5, in both symmetric and asymmetric valuations). CAFormer, a transformer with permutation-equivariant exchangeable layers, consistently outperforms CANet, a fully connected network, in revenue while keeping regret near zero. The authors interpret this as evidence that differentiable economics can be moved into the combinatorial setting without restricting the allowable bundle structures or valuation distributions, and that the resulting randomized mechanisms capture revenue that deterministic heuristics miss.

Load-bearing premise

The load-bearing premise is that the softmax-min normalization used to build the allocation matrix really does guarantee that every item is allocated at most once in total; the proof of this lemma assumes a bound that does not follow when several bundles share the same item, so a counterexample would collapse the claimed feasibility of the output allocations.

Editorial extensions

If this is right

If the empirical results hold, revenue-maximizing combinatorial auctions can be designed by gradient descent without specifying allowable bundle structures or valuation distributions.
A single CAFormer model, being permutation-equivariant and size-tolerant, could be transferred to auctions with more bidders or items than it was trained on.
The two-step allocation decomposition provides a differentiable parameterization of combinatorial feasibility that could be reused in other constrained allocation problems.
The consistent revenue gap between CAFormer and CANet suggests attention-based architectures better model the complementarities and substitutions in bundle valuations.
Because the mechanisms are randomized, they can exploit lottery effects that deterministic winner-determination heuristics cannot, which may explain the revenue gains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The feasibility lemma's proof has a gap: it bounds the total probability allocated to bundles containing a given item by assuming the per-bundle item bounds add up, but they need not because multiple bundles share items; if the networks ever output a matrix violating constraint (4), the reported revenues are not achievable by any lottery.
A straightforward check is to decompose the output allocation matrix into a convex combination of deterministic feasible allocations; failure to find such a decomposition would show the network is not implementing a valid randomized mechanism.
The method's slight revenue underperformance in non-combinatorial settings hints that the extra bundle-level complexity can hurt convergence, so the practical advantage may depend on the presence of item complementarities.
If the constraint issue is repaired, the same softmax-min parameterization could be applied to other differentiable combinatorial optimization problems, but the current construction shows how easily such parameterizations can silently violate feasibility.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is worth knowing about because it extends RegretNet/RegretFormer to combinatorial auctions with a bundle-level allocation decomposition and attention-based architectures. The empirical work is substantial: they compare against VCG, AMA, local search, and RegretNet-family models on 2x2, 2x3, and 2x5 settings, and their reported revenues beat the baselines. The writing is clear, and the authors honestly note limitations like non-convergence and the challenge of large-scale lotteries.

But the central feasibility guarantee is not established. Lemma 1's proof has the exact gap your reader identified: per-bundle column constraints on A_agent-bundle do not imply per-item capacity when multiple bundles share an item, and the softmax normalization over items for each bundle does not control the sum over bundles for a fixed item. The final step is a non-sequitur—the number of bundles containing an item is 2^{m-1}, not 1, so (10) and (11) cannot yield constraint (4). The Birkhoff theorem invocation is also misplaced because the relaxed polytope is a set-packing polytope, not necessarily integral. The stress-test counterexample (z_{1,{1}}=1, z_{2,{1,2}}=0.5) is concrete and valid. So the networks may over-sell items, meaning the revenue figures in Tables 1 and 2 may not correspond to implementable lotteries.

I agree with the reader's verdict of REJECT. This is a load-bearing flaw, not a minor gap. The missing code and data, and the lack of error bars, add to the difficulty of assessing the empirical claims. The scalability argument is also strained: the bundle space is exponential, and the 2x5 case already uses 31 bundles.

That said, the paper is not without merit. It gives a plausible architecture family and a training recipe, and it honestly discusses the open theoretical question of whether the network can represent the true optimal mechanism. If the authors can fix the constraint enforcement—perhaps by using a proper doubly-stochastic relaxation that is actually lottery-decomposable—the empirical approach could be salvageable. But as it stands, the core assertion is unsupported.

For a serious editor: this deserves peer review because the problem is important and the architectures are interesting, even though the current form would need major revision. I would not cite it in its present state, and I would not bring it to the reading group except as a case study in how a plausible sounding feasibility proof can fail.

Referee Report

3 major / 5 minor

Summary. This paper proposes two neural architectures, CANet and CAFormer, for learning randomized revenue-maximizing mechanisms in combinatorial auctions (CAs). The allocation is constructed in Section 3.1 as a product of a per-bundle item-feasibility factor and a per-agent bundle-allocation factor, with softmax and min normalizations claimed to enforce the item-capacity, per-agent, and lottery-decomposability constraints. The networks are trained with a RegretNet-style objective plus an annealed regret budget, and the experiments compare revenue and regret against VCG, affine maximizer auctions, local-search baselines, and RegretNet/RegretFormer on 2x2, 2x3, and 2x5 instances.

Significance. If the feasibility guarantee and empirical comparisons were valid, the paper would make a useful step: it would provide a scalable, assumption-light, gradient-trained randomized-mechanism family for combinatorial auctions, and it would document improved revenue over several classical baselines. The effort to build differentiable constructions for combinatorial constraints and the inclusion of a permutation-equivariant transformer variant are creditable. However, the central lemma that certifies feasibility is not proven, and the Birkhoff-based justification for lottery implementability is misapplied; these gaps directly undermine the reported revenue and regret numbers. The paper should not be published in its present form.

major comments (3)

[§3.1, Lemma 1] The proof of item-capacity constraint (4) is invalid. After bounding each bundle column by (10) and each agent row by (11), the proof asserts that sum_i sum_{S containing j} A_{iS} <= 1. This is a non-sequitur: summing the per-bundle column bound over the 2^{m-1} bundles containing item j yields 2^{m-1}, not 1, and (11) is a per-agent row constraint. A concrete matrix satisfying the two bounds used in the proof is A_{1,{1}}=0.6, A_{1,{1,2}}=0.4, A_{2,{2}}=0.4, A_{2,{1,2}}=0.6, all other entries zero; each column and each row sums to 1, yet the total mass on bundles containing item 1 is 1.6 > 1. With B_bundle equal to 1 on these bundles, Z allocates item 1 to multiple bundles, violating (4). Thus CANet/CAFormer allocations can over-sell items, and the revenue and regret values in Tables 1 and 2 are not computed for known-feasible mechanisms.
[§2.1 and §3.1, Theorem 1] The invocation of Birkhoff's theorem to justify that every point satisfying (4), (5), and (7) is a lottery over deterministic feasible allocations is a category error. The constraints define a set-packing polytope over bundles, not the doubly stochastic matrices of a bipartite matching, and that set-packing polytope is not integral in general. For example, with n=2, m=3 and bundles {1,2}, {2,3}, {1,3}, the fractional allocation z_{1,{1,2}}=0.5, z_{1,{2,3}}=0.5, z_{2,{1,3}}=0.5 satisfies (4), (5), and (7), but it cannot be written as a convex combination of feasible integer allocations: bidder 1 receives a bundle with probability 1, and in every state where bidder 1 receives {1,2} or {2,3}, bidder 2 cannot simultaneously receive {1,3}. Therefore, even a point satisfying the relaxed inequalities need not be implementable as a lottery, and Theorem 1 does not supply the missing argument.
[§4, Tables 1 and 2] The main empirical claim that the proposed models 'consistently outperform' the baselines is not supported by statistical evidence. The tables report a single revenue and regret value per model with no standard errors, no number of random seeds, and no significance tests. Combined with the two feasibility issues above, the reported numbers cannot be used to evaluate the proposed mechanisms or to justify the 'new benchmarks' claim. At minimum, the authors would need to provide a valid feasibility guarantee and repeated-run statistics with confidence intervals.

minor comments (5)

[§3.1] The matrix product B · I in the definition of Badjusted is dimensionally inconsistent as written, since both B and I are stated to be in R^{m x k}; please clarify the intended multiplication or transposition.
[Algorithm 1] The bias-correction formula in line 7 of Algorithm 1 contains typographical errors in the exponents and parentheses; the update should be written cleanly.
[§4, training setup] The sentence describing exponentially annealing regret targets 'in 2/3 of the training iterationw' contains a typo, and the exact annealing schedule should be stated precisely or given as pseudocode.
[§5] The claim that the approach is 'inherently not limited to assumptions about the structures of bundles and valuation profiles' and can be applied when the distribution is unknown is overstated; training as described requires sampling from a known or estimable distribution to evaluate expected revenue and regret.
[§2.1] The citation of Birkhoff's theorem should be checked; the standard reference for the doubly stochastic matrix decomposition is Birkhoff (1946), and the date given in the text appears incorrect.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation is present; the feasibility lemma issue is a correctness problem, not a tautological reduction.

full rationale

The paper's empirical claims are validated against independent baselines (VCG, AMA, VVCA, BLAMA, ABAMA, BBBVVCA, RegretNet, RegretFormer), and reported revenue is the trained network's actual expected revenue on held-out profiles. No parameter is fitted to the target quantity and then renamed as a prediction; the regret budget is a user-specified annealing schedule, and revenue remains a genuine output of the optimization. There are no load-bearing self-citations: the authors do not rely on their own prior work, and the external results invoked (Birkhoff's theorem, RegretNet, Myerson) are independent, non-tautological inputs. The Section 3.1 construction is a parameterization of allocations; Lemma 1 attempts to prove feasibility, and even if its final step is invalid (per-bundle softmax does not imply per-item capacity), that is a mathematical error, not a definitional equivalence. The paper itself concedes non-convexity and lacking convergence guarantees in Section 5, further showing the claims are not derived by definition. Hence no circular step of any of the seven kinds is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Beyond standard quasi-linear utility assumptions, the paper's central construction leans on two unproven premises: the LP relaxation describes lottery-decomposable allocations, and the softmax-min normalization enforces item capacities. Both are load-bearing for the claim that the networks output feasible randomized mechanisms, and both are suspect given the proof gap in Lemma 1 and the non-integrality of set-packing relaxations.

free parameters (4)

softmax temperature theta = {10, 15, 25}
Controls concentration of allocation probabilities in A_agent-bundle; chosen by hand for training. Listed in Section 4 hyperparameters.
regret target budget rgt_bar = annealed from 0.05 to {0.0008, 0.001, 0.002, 0.003} in 2/3 of training iterations
Sets the allowed DSIC violation; reported revenue and regret values are conditional on this hand-chosen schedule.
revenue-regret factor alpha = {0.5, 1}
Adjusts target regret relative to revenue in Eq. (12); chosen by hand.
tanh weight scaler rho = 2
Scales weight normalization in Algorithm 1; chosen by hand.

assumptions (5)

domain assumption The relaxed polytope (4),(5),(7) equals the convex hull of deterministic feasible CA allocations, so any fractional allocation satisfying them is lottery-decomposable.
Invoked after Theorem 1 in Section 2.1. Birkhoff's theorem applies to doubly stochastic matrices; this set-packing relaxation is not generally integral, so the assumption is unsupported and false in general.
domain assumption The inner adversarial loop (Eq. 9) finds the best misreport, so estimated regret is accurate.
The paper concedes in Section 4 that inner-loop optimization might be inaccurate and that underestimated regret overstates revenue.
domain assumption Valuations are quasi-linear, drawn from known continuous distributions F_i, and the set of all 2^m - 1 bundles is available.
Standard CA model used throughout; experiments use uniform distributions. The paper claims generality beyond this but provides no evidence.
domain assumption The payment rule p_i = ptilde_i times sum_S z_iS b_iS with ptilde_i in [0,1] preserves individual rationality when the allocation is a realized lottery.
The payment is charged even if the realized allocation gives the bidder nothing, so ex-post IR can fail; only expected IR is ensured.
ad hoc to paper The matrices in Section 3.1 can be multiplied as written (B dot I) and reshaped to produce Z = B_bundle times A_agent-bundle.
B in R^(m x k) and I in R^(m x k) are not conformable for the stated product; the construction depends on an implicit re-definition of B.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Advancing Differentiable Economics: A Neural Network Framework for Revenue-Maximizing Combinatorial Auction Mechanisms." pith.science (2026). https://pith.science/paper/UTWTPC3B

@misc{pith2026250119219,
  author       = {Pith},
  title        = {Pith review of: Advancing Differentiable Economics: A Neural Network Framework for Revenue-Maximizing Combinatorial Auction Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UTWTPC3B}},
  note         = {Machine review of arXiv:2501.19219}
}

read the original abstract

Differentiable economics, which uses neural networks as function approximators and gradient-based optimization in automated mechanism design (AMD), marked a significant breakthrough with the introduction of RegretNet \citep{regretnet_paper}. It combines the flexibility of deep learning with a regret-based approach to relax incentive compatibility, allowing for approximations of revenue-maximizing auctions. However, applying these techniques to combinatorial auctions (CAs) - where bidders value bundles rather than individual items, capturing item interdependencies - remains a challenge, primarily due to the lack of methodologies that can effectively deal with combinatorial constraints. To tackle this, we propose two architectures: CANet, a fully connected neural network, and CAFormer, a transformer-based model designed to learn optimal randomized mechanisms. Unlike existing methods in traditional AMD, our approach is more scalable and free of assumptions about the structures of allowable bundles or bidder valuations. We demonstrate that our models match current methods in non-combinatorial settings and set new benchmarks for CAs. Specifically, our models consistently outperform benchmark mechanisms derived from heuristic approaches and provide empirical solutions where analytical results are unavailable. This work bridges the gap in applying differentiable economics to combinatorial auctions, offering a scalable and flexible framework for designing revenue-maximizing mechanisms.

Figures

Figures reproduced from arXiv: 2501.19219 by the authors.

**Figure 1.** CANet and CAFormer Architectures. Z ∈ R n×k into the product of two matrices: Z = B bundle · Aagent-bundle , where Aagent-bundle and Bbundle are defined below. This approach is motivated by a two-step allocation process: first, the allocation of items to the bundles, satisfying the itemwise constraint, and then the allocation of the bundles to the bidders, satisfying the other constraints. Each entry in the final… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Strategic Cyber Defense via Reinforcement Learning-Guided Combinatorial Auctions
cs.GT 2025-09 conditional novelty 5.0 of 10

RL Q-values are used as bids in a learned combinatorial auction that allocates defensive actions in the DARPA CAGE 2 simulation, giving revenue near an oracle and allocations loosely aligned with defender activity.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages · cited by 1 Pith paper

[1]

Optimal multi-object auctions

Mark Armstrong. Optimal multi-object auctions. The Review of Economic Studies , 67(3):455--481, 07 2000

work page 2000
[2]

Bundling and optimal auctions of multiple products

Christopher Avery and Terrence Hendershott. Bundling and optimal auctions of multiple products. The Review of Economic Studies , 67(3):483--497, 2000

work page 2000
[3]

The winner's curse, reserve prices, and endogenous entry: Empirical insights from ebay auctions

Patrick Bajari and Ali Hortaçsu. The winner's curse, reserve prices, and endogenous entry: Empirical insights from ebay auctions. The RAND Journal of Economics , 34(2):329--355, 2003

work page 2003
[4]

Combinatorial auctions via machine learning-based preference elicitation

Gianluca Brero, Benjamin Lubin, and Sven Seuken. Combinatorial auctions via machine learning-based preference elicitation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18 , pages 128--136. International Joint Conferences on Artificial Intelligence Organization, 7 2018

work page 2018
[5]

Machine learning-powered iterative combinatorial auctions, 2021

Gianluca Brero, Benjamin Lubin, and Sven Seuken. Machine learning-powered iterative combinatorial auctions, 2021

work page 2021
[6]

Edward H. Clarke. Multipart pricing of public goods. Public Choice , 11:17--33, 1971

work page 1971
[7]

Combinatorial Auctions

Peter Cramton, Yoav Shoham, and Richard Steinberg. Combinatorial Auctions . The MIT Press, December 2005

work page 2005
[8]

Differentiable economics for randomized affine maximizer auctions

Michael Curry, Tuomas Sandholm, and John Dickerson. Differentiable economics for randomized affine maximizer auctions. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence , IJCAI-2023, page 2633–2641. International Joint Conferences on Artificial Intelligence Organization, August 2023

work page 2023

Show all 35 references

[9]

Optimal auctions through deep learning

Paul Duetting, Zhe Feng, Harikrishna Narasimhan, David Parkes, and Sai Srivatsa Ravindranath. Optimal auctions through deep learning. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning , volume 97 of Pr...

2019
[10]

Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords

Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords. American Economic Review , 97(1):242–259, March 2007

2007
[11]

Zhe Feng, Harikrishna Narasimhan, and David C. Parkes. Deep learning for revenue-optimal auctions with budgets. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems , AAMAS '18, page 354–362, Richland, SC, 2018. International Foundati...

2018
[12]

Understanding the difficulty of training deep feedforward neural networks

Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Yee Whye Teh and Mike Titterington, editors, Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , volume 9 of Proceeding...

2010
[13]

Noah Golowich, Harikrishna Narasimhan, and David C. Parkes. Deep learning for multi-facility location mechanism design. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence , IJCAI-2018, page 261–267. International Joint Conferences on...

2018
[14]

Incentives in teams

Theodore Groves. Incentives in teams. Econometrica , 41(4):617--631, 1973

1973
[15]

Optimal-er auctions through attention, 2022

Dmitry Ivanov, Iskander Safiulin, Igor Filippov, and Ksenia Balabaeva. Optimal-er auctions through attention, 2022

2022
[16]

Differentially private machine learning-powered combinatorial auction design, 2024

Arash Jamshidi, Seyed Mohammad Hosseini, Seyed Mahdi Noormousavi, and Mahdi Jafari Siavoshani. Differentially private machine learning-powered combinatorial auction design, 2024

2024
[17]

Curry, Samuel Dooley, Ping yeh Chiang, Tom Goldstein, and John P

Kevin Kuo, Anthony Ostuni, Elizabeth Horishny, Michael J. Curry, Samuel Dooley, Ping yeh Chiang, Tom Goldstein, and John P. Dickerson. Proportionnet: Balancing fairness and revenue for auction design with deep learning. ArXiv , abs/2010.06398, 2020

2010 arXiv
[18]

Truthful and near-optimal mechanism design via linear programming

Ron Lavi and Chaitanya Swamy. Truthful and near-optimal mechanism design via linear programming. J. ACM , 58(6), December 2011

2011
[19]

Riley, and F

Eric Maskin, J. Riley, and F. Hahn. Optimal Multi-Unit Auctions , pages 312--335. Oxford University Press, 1989. Reprinted in P. Klemperer, The Economic Theory of Auctions, London: Edward Elgar, 2000

1989
[20]

Roger B. Myerson. Optimal auction design. Mathematics of Operations Research , 6(1):58--73, 1981

1981
[21]

Matthew Weinberg

Jad Rahme, Samy Jelassi, Joan Bruna, and S. Matthew Weinberg. A permutation-equivariant neural network architecture for auction design, 2021

2021
[22]

Matthew Weinberg

Jad Rahme, Samy Jelassi, and S. Matthew Weinberg. Auction learning as a two-player game, 2021

2021
[23]

Kominers, and David C

Sai Srivatsa Ravindranath, Zhe Feng, Shira Li, Jonathan Ma, Scott D. Kominers, and David C. Parkes. Deep learning for two-sided matching, 2023

2023
[24]

Sai Srivatsa Ravindranath, Zhe Feng, Di Wang, Manzil Zaheer, Aranyak Mehta, and David C. Parkes. Deep reinforcement learning for sequential combinatorial auctions, 2024

2024
[25]

A graph-based ant algorithm for the winner determination problem in combinatorial auctions

Abhishek Ray, Mario Ventresca, and Karthik Kannan. A graph-based ant algorithm for the winner determination problem in combinatorial auctions. Information Systems Research , 32(4):1099–1114, December 2021

2021
[26]

K. Roberts. The characterization of implementable choice rules. In J.-J. Laffont, editor, Aggregation and Revelation of Preferences , pages 321--348. North Holland Publishing, 1979

1979
[27]

Automated design of revenue-maximizing combinatorial auctions

Tuomas Sandholm and Anton Likhodedov. Automated design of revenue-maximizing combinatorial auctions. Operations Research , 63(5):1000–1025, October 2015

2015
[28]

Algorithm for optimal winner determination in combinatorial auctions

Tuomas Sandholm. Algorithm for optimal winner determination in combinatorial auctions. Artificial Intelligence , 135(1–2):1–54, February 2002

2002
[29]

Automated mechanism design: A new application area for search algorithms

Tuomas Sandholm. Automated mechanism design: A new application area for search algorithms. In Francesca Rossi, editor, Principles and Practice of Constraint Programming -- CP 2003 , pages 19--36, Berlin, Heidelberg, 2003. Springer Berlin Heidelberg

2003
[30]

Learning truthful, efficient, and welfare maximizing auction rules, 2022

Andrea Tacchetti, DJ Strouse, Marta Garnelo, Thore Graepel, and Yoram Bachrach. Learning truthful, efficient, and welfare maximizing auction rules, 2022

2022
[31]

Attention is all you need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural In...

2017
[32]

Counterspeculation, auctions, and competitive sealed tenders

William Vickrey. Counterspeculation, auctions, and competitive sealed tenders. The Journal of Finance , 16(1):8--37, 1961

1961
[33]

A clique-based exact method for optimal winner determination in combinatorial auctions

Qinghua Wu and Jin-Kao Hao. A clique-based exact method for optimal winner determination in combinatorial auctions. Information Sciences , 334–335:103–121, March 2016

2016
[34]

Deep sets, 2018

Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan Salakhutdinov, and Alexander Smola. Deep sets, 2018

2018
[35]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Optimal multi-object auctions

Mark Armstrong. Optimal multi-object auctions. The Review of Economic Studies , 67(3):455--481, 07 2000

work page 2000

[2] [2]

Bundling and optimal auctions of multiple products

Christopher Avery and Terrence Hendershott. Bundling and optimal auctions of multiple products. The Review of Economic Studies , 67(3):483--497, 2000

work page 2000

[3] [3]

The winner's curse, reserve prices, and endogenous entry: Empirical insights from ebay auctions

Patrick Bajari and Ali Hortaçsu. The winner's curse, reserve prices, and endogenous entry: Empirical insights from ebay auctions. The RAND Journal of Economics , 34(2):329--355, 2003

work page 2003

[4] [4]

Combinatorial auctions via machine learning-based preference elicitation

Gianluca Brero, Benjamin Lubin, and Sven Seuken. Combinatorial auctions via machine learning-based preference elicitation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18 , pages 128--136. International Joint Conferences on Artificial Intelligence Organization, 7 2018

work page 2018

[5] [5]

Machine learning-powered iterative combinatorial auctions, 2021

Gianluca Brero, Benjamin Lubin, and Sven Seuken. Machine learning-powered iterative combinatorial auctions, 2021

work page 2021

[6] [6]

Edward H. Clarke. Multipart pricing of public goods. Public Choice , 11:17--33, 1971

work page 1971

[7] [7]

Combinatorial Auctions

Peter Cramton, Yoav Shoham, and Richard Steinberg. Combinatorial Auctions . The MIT Press, December 2005

work page 2005

[8] [8]

Differentiable economics for randomized affine maximizer auctions

Michael Curry, Tuomas Sandholm, and John Dickerson. Differentiable economics for randomized affine maximizer auctions. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence , IJCAI-2023, page 2633–2641. International Joint Conferences on Artificial Intelligence Organization, August 2023

work page 2023