Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Bandit Learning in Matching Markets: Utilitarian and Rawlsian Perspectives

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper claims that both the utilitarian-optimal and the maximin stable matching can be learned when neither side knows its preferences, with regret that is nearly logarithmic in time.

desk verdict Fresh welfarist framing for bandit matching, but the Rawlsian half rests on a false lemma; the utilitarian half is plausible but under-proved. read the letter →

arxiv 2412.00301 v1 pith:M5R2YUQC submitted 2024-11-30 cs.LG cs.GT

Hadi Hosseini , Duohan Zhang This is my paper · ORCID
classification cs.LGcs.GT MSC 91B6868W2762L05
keywords two-sidedmatchingmarketsmulti-armedbanditsutilitarianwelfareRawlsianmaximinstableexplore-then-commitregretboundspreferencelearning
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

Two-sided matching markets — school choice, medical residency, ride-hailing — are usually analyzed as if preferences were known, and the standard deferred acceptance algorithm protects only the proposing side. This paper asks what can still be guaranteed when both sides must learn their preferences from noisy samples, and it answers with an explore-then-commit scheme. It claims that one version of the scheme converges to the stable matching maximizing total (utilitarian) welfare with regret $\tilde O(N^2 \log T)$, and another converges to the stable matching maximizing the utility of the worst-off participant (Rawlsian or maximin welfare) with regret $\tilde O(N \log T)$. Both guarantees require a minimum preference gap, meaning utilities that are almost tied make the problem harder; if the bounds are right, a planner can tolerate unknown preferences and still lose only polylogarithmic welfare over time.

What carries the argument

The load-bearing objects are the minimum preference gaps. The within-side gap $\Delta_a$ ($\Delta_b$) is the smallest utility difference between any two arms for one agent (any two agents for one arm); it controls whether sample averages preserve ordinal preferences. The cross-side gap $\Gamma$ is the smallest difference between any two utilities anywhere in the market; it controls whether the maximin objective survives estimation. Between them sits the rotation structure: starting from the agent-optimal stable matching obtained by deferred acceptance, repeatedly eliminating rotations traces all stable matchings to the arm-optimal one, and the utilitarian optimum is the minimum-weight closed set of rotations, computable by an $s$--$t$ min-cut. The algorithms are epoch explore-then-commit routines: in each epoch agents do round-robin exploration, form sample-average utilities, compute the estimated optimal matching, and commit to it for $2^l$ steps.

What would settle it

Enumerate all stable matchings of a small market (e.g., $N=3$) with randomly drawn utilities, compute the true maximin stable matching, and compare it to the output of Algorithm 3 run with the true utilities. If the algorithm's output has a lower minimum utility than some other stable matching, the side-restricted maximin routine is not computing the global maximin, and Lemma 5 and Theorem 2 would not target the stated objective.

Watch

Extended reading notes

Core claim

The paper's central claim is that the classical discrete structure of stable matchings — the lattice of rotations between agent-optimal and arm-optimal matchings — survives in the bandit setting and can be exploited without knowing either side's preferences. For the utilitarian objective, it recovers a cardinal-utility version of the rotation-based algorithm: starting from the two deferred-acceptance extremes, every stable matching is produced by eliminating a closed set of rotations, and the utilitarian optimum is the minimum-weight closed set. It proves that if every estimated mean is within $\min\{\delta/(4N), \Delta_a/2, \Delta_b/2\}$ of the truth, where $\delta$ is the welfare gap to the second-best stable matching, the recovered matching is exactly the utilitarian optimum; this yields regret $O(N^2 \log T \log\log T + N^3 \log T + 2^{l_1}N)$. For the Rawlsian objective, it adapts the side-restricted maximin procedure and shows that estimation error below $\Gamma/2$, where $\Gamma$ is the cross-side minimum preference gap, suffices to recover the maximin stable matching, with regret $O(N \log T \log\log T + 4N^2 \log T + 2^{l_1})$.

Load-bearing premise

The Rawlsian proof assumes that the procedure that maximizes the worst-off arm (and, symmetrically, the worst-off agent) produces the stable matching that maximizes the minimum utility across both sides; the paper does not demonstrate that equivalence.

Editorial extensions

If this is right

  • If the bounds are correct, a planner facing two-sided unknown preferences can guarantee that after an initial burn-in the market nearly always plays the targeted stable matching, with regret growing only polylogarithmically in the horizon.
  • The two gap notions give an instance-dependent measure of hardness: a small $\delta$ (two stable matchings with nearly equal utilitarian welfare) or a small $\Gamma$ (two utilities anywhere nearly tied) forces longer exploration.
  • The utilitarian bound scales as $N^2$ in the number of agents while the maximin bound scales as $N$ (ignoring gap and exponential terms), so the worst-off-focused objective is not harder in market size; the cross-side gap dominates instead.
  • For unequal numbers of agents and arms, the rural hospital theorem restricts attention to the fixed set of stable partners, so the same algorithms apply to the $\min\{N, K\}$ matched pairs.

Reading between the lines

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

  • The Rawlsian proof treats the arm-side maximin routine (and its agent-side mirror) as a way to compute the global two-sided maximin of Definition 4.1; this equivalence is assumed rather than shown, so a reader should check it before relying on Theorem 2's benchmark.
  • The additive $2^{l_1}$ term with $l_1$ exponential in $1/\Gamma^2$ indicates that near-tied utilities anywhere in the market force a burn-in exponential in the inverse gap; replacing the hard-gap analysis with gap-free or adaptive estimation is a natural extension.
  • The same rotation-lattice template should apply to other objectives defined on stable matchings, with the optimization oracle determining the learning rate rather than the exploration scheme.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper considers a two-sided matching market with N agents and N arms in which the cardinal utilities of both sides are unknown and are learned from stochastic rewards. The authors study two welfarist objectives among stable matchings: maximizing total utilitarian welfare and maximizing the minimum (Rawlsian) utility over all participants. They propose two epoch-based Explore-Then-Commit algorithms (Algorithm 2 for utilitarian, Algorithm 4 for maximin), relying on known-preference subroutines (Algorithm 1 for utilitarian-optimal and Algorithm 3 for maximin stable matchings). They claim regret bounds of O(N^2 log T log log T + N^3 log T + 2^{l1}N) and O(N log T log log T + 4N^2 log T + 2^{l1}), respectively, and PAC sample complexity bounds O(N^2/β^2 log(N^2/α)) and O(N^2/Γ^2 log(N^2/α)), supported by simulations on random preference profiles.

Significance. The welfarist framing is timely and, if the results were correct, the bounds would be the first regret guarantees for utilitarian and Rawlsian objectives in two-sided unknown-preference matching markets; the distinction between within-side and cross-side preference gaps is a reasonable analytical idea. The paper's utilitarian arm contains a credible epoch-ETC structure and a provable error-probability argument conditional on Algorithm 1. However, the Rawlsian half is not established: Lemma 5 is false as stated, Algorithm 3 is not shown to compute the global maximin stable matching, and the error-event analysis in Lemma 6 does not account for non-unique maximizers. Because Theorems 2 and 4 rely on these components, the advertised Rawlsian results cannot be accepted. The paper provides no machine-checked proofs or code release; the experiments are illustrative and lack baselines.

major comments (3)
  1. [Section 4.3, Lemma 5] Lemma 5 in Section 4.3 is false. The proof's key step is the assertion that if R(\hat m) < R(m*), then R(\hat m) + Γ ≤ R(m*) 'by construction of Γ'. But Γ = min{Γ_a, Γ_b} lower-bounds the distance between two utilities on the same side of the market, whereas R(m) is the minimum over the union of both sides; the gap between an agent-side utility and an arm-side utility is not controlled by Γ. A concrete instance with N = 2 shows the failure: set (μ_{a1,b1}, μ_{a1,b2}) = (100, 10), (μ_{a2,b1}, μ_{a2,b2}) = (10, 100), (μ_{b1,a1}, μ_{b1,a2}) = (1, 50), and (μ_{b2,a1}, μ_{b2,a2}) = (50, 1). The two stable matchings are M1 = {(a1,b1), (a2,b2)} with R(M1) = 1 and M2 = {(a1,b2), (a2,b1)} with R(M2) = 10, so Γ = min(90, 49) = 49 but R(M2) - R(M1) = 9 < Γ. If the estimates of μ_{a1,b2} and μ_{a2,b1} are both changed from 10 to 0, all estimate errors are below Γ/2, yet the estimated-profile maximin matching becomes M1, whose true Rawlsian value is 1 rather than 10. Thus the hypotheses of Lemma 5 can hold while its conclusion fails. Lemma 5 is used to prove Lemma 6, Theorem 2, and Theorem 4, so the Rawlsian regret and sample-complexity claims are not supported.
  2. [Section 4.1, Algorithm 3] Algorithm 3 is introduced as computing the arm-side maximin stable matching, but Definition 4.1 defines the maximin stable matching as maximizing the minimum over all agents and arms. No theorem or proof establishes that the arm-side maximizer has the global maximin value, or that combining arm-side and agent-side variants yields the global maximizer. Theorem 2 asserts that Algorithm 4 computes the maximin stable matching, and the analysis in Lemma 5 simply assumes that the matching output by Algorithm 3 maximizes \hat R over the stable set. This is a load-bearing gap: without such a proof, even a correct perturbation bound would not connect the algorithm's output to the stated Rawlsian objective.
  3. [Section 4.3, Lemma 6] Lemma 6 defines E_l as the event that the matching returned in epoch l differs from m*. Lemma 5, even if repaired to guarantee value optimality, does not imply uniqueness of the maximin stable matching. If several stable matchings attain the maximin value, the estimated-profile maximizer can differ from a fixed m* while still having optimal true value; the claimed probability bound P(E_l) ≤ 4N^2 exp(-c1 l log(l) Γ^2 / 8) then does not follow. For the regret analysis only the probability that the returned matching's true Rawlsian value is suboptimal is needed, and the lemma should be restated accordingly.
minor comments (5)
  1. [Throughout] There are numerous typos, including 'application' for 'applications' in Section 1, 'mimimum' in Section 4.1, 'optiaml' in Section 6, and 'decompostion' in the proof of Lemma 3 in Appendix A.1.
  2. [Section 4.3] The definition of the cross-side minimum preference gap is hard to parse: the displayed formula for Γ_a should explicitly state that the minimization is over ordered quadruples with non-zero utility difference and should show the absolute value unambiguously; the current notation does not make the index ranges and the absolute value clear.
  3. [Section 3.3, Theorem 3] In Theorem 3, the quantity β is used in the sample complexity bound but is not defined in the statement; it should be β = min{δ/(4N), Δ_a/2, Δ_b/2} as introduced in Lemma 4.
  4. [Appendix C, proof of Theorem 4] The proof of Theorem 4 twice refers to the target as 'the utilitarian-optimal stable matching'; it should refer to the maximin stable matching.
  5. [Algorithm 3] The pseudocode of Algorithm 3 is incomplete: the break-matching operation is not specified, ties among multiple minimum-reward arms are not addressed, and the stopping condition in line 3 ('no arm with reward R(m_{i+1})') is unclear.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation identified; self-citations are contextual and the proof gaps are correctness concerns, not circularity.

full rationale

The derivation chain for Theorem 1 is self-contained: Lemma 1 couples empirical to true ordinal preferences through the environment-defined within-side gaps Delta_a and Delta_b; Lemma 2 uses the per-pair error budget min{delta/(4N), Delta_a/2, Delta_b/2} to bound |R(m) - R_hat(m)| < delta/2 and then applies the definition delta = R(m*) - R(m2); Lemma 4 and Theorem 1 only union-bound deviations and sum epoch regrets. The Rawlsian chain is structurally the same, with Gamma defined from the input preference profile; no parameter is fitted to the target regret and no 'prediction' is a renamed fit. The self-citations to Hosseini et al. [2024] appear only in the related-work paragraph and in the experimental setup sentence ('we follow similar constructions with previous literature Liu et al. [2021], Hosseini et al. [2024]'), and neither supports the theorems; the combinatorial subroutines are credited to external work (Irving et al. 1987; Gusfield 1987). Thus no self-citation is load-bearing and no equation reduces to its own input. Two non-circular risks are flagged for the record: Section 4.1 asserts without proof that the arm-side maximin subroutine (Algorithm 3), possibly combined with its agent-side analogue, yields the global maximin of Definition 4.1; and Lemma 5's claim 'by construction of Gamma, R(m_hat)+Gamma <= R(m*)' is not justified since Gamma bounds within-side differences while R(m) is a minimum across both sides. These would threaten Theorem 2 and Theorem 4 if confirmed, but they are correctness gaps, not circularity.

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

The paper introduces no new physics or entities. Its free-parameter ledger is empty because the preference gaps are environmental parameters, not fitted numbers. The key axioms are the correctness of the matching subroutines and the existence of the gaps, with the maximin correctness being the most fragile.

assumptions (5)
  • standard math Standard subgaussian concentration bound (Lemma 7, from Lattimore and Szepesvari).
    Used in Lemma 4 and Lemma 6 to bound estimation error probabilities from the sample averages of subgaussian rewards.
  • domain assumption There exists a positive utilitarian welfare gap delta = R(m*) - R(m2) between the best and second-best stable matchings.
    Invoked in Lemma 2 and Theorem 1. The paper says 'we assume such matching exists without loss of generality', but this requires a unique optimal matching; ties in welfare would make delta = 0 and the error bound trivial.
  • domain assumption Algorithm 1 (rotation graph and min-cut) computes the utilitarian-optimal stable matching under bounded estimation error.
    The paper cites Irving et al. 1987, but the correctness of the cardinal-utility adaptation is not proved in this paper. Lemma 2 relies on this being true.
  • ad hoc to paper Algorithm 3 computes the global maximin stable matching as defined in Definition 4.1.
    The paper describes Algorithm 3 as finding the arm-side maximin stable matching and says the agent-side variant is symmetric, but it never proves that this yields the global maximin over both sides. Lemma 5 and Theorem 2 depend on this unstated assumption. This is the load-bearing gap.
  • domain assumption The cross-side minimum preference gap Gamma is positive.
    Gamma is the minimum nonzero difference between any two distinct utility values across all agents and arms. The analysis in Lemma 5 requires Gamma > 0. This holds with probability 1 for continuous utilities but is not guaranteed for arbitrary instances with ties or equal values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bandit Learning in Matching Markets: Utilitarian and Rawlsian Perspectives." pith.science (2026). https://pith.science/paper/M5R2YUQC

@misc{pith2026241200301,
  author       = {Pith},
  title        = {Pith review of: Bandit Learning in Matching Markets: Utilitarian and Rawlsian Perspectives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5R2YUQC}},
  note         = {Machine review of arXiv:2412.00301}
}
read the original abstract

Two-sided matching markets have demonstrated significant impact in many real-world applications, including school choice, medical residency placement, electric vehicle charging, ride sharing, and recommender systems. However, traditional models often assume that preferences are known, which is not always the case in modern markets, where preferences are unknown and must be learned. For example, a company may not know its preference over all job applicants a priori in online markets. Recent research has modeled matching markets as multi-armed bandit (MAB) problem and primarily focused on optimizing matching for one side of the market, while often resulting in a pessimal solution for the other side. In this paper, we adopt a welfarist approach for both sides of the market, focusing on two metrics: (1) Utilitarian welfare and (2) Rawlsian welfare, while maintaining market stability. For these metrics, we propose algorithms based on epoch Explore-Then-Commit (ETC) and analyze their regret bounds. Finally, we conduct simulated experiments to evaluate both welfare and market stability.

Figures

Figures reproduced from arXiv: 2412.00301 by the authors.

Figure 1
Figure 1. 95% confidence interval of utilitarian regrets, maximin regrets, and stability; 200 preference profiles generated uniformly at random. 5 Experimental Results In this section, we experimentally validate our theoretical results by examining the regrets and stability for two types of Epoch ETC algorithms. For this, we consider N = K = 5, 10, 15 and randomly generate preferences. In particular, we follow similar constru… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

    The New York City High School Match

    Atila Abdulkadiro g lu, Parag A Pathak, and Alvin E Roth. The New York City High School Match . American Economic Review, 95 0 (2): 0 364--367, 2005 a

    work page 2005

  2. [2]

    The Boston Public School Match

    Atila Abdulkadiro g lu, Parag A Pathak, Alvin E Roth, and Tayfun S \"o nmez. The Boston Public School Match . American Economic Review, 95 0 (2): 0 368--371, 2005 b

    work page 2005

  3. [3]

    Ride Sharing, pages 73--97

    Siddhartha Banerjee and Ramesh Johari. Ride Sharing, pages 73--97. Springer International Publishing, Cham, 2019. ISBN 978-3-030-01863-4. doi:10.1007/978-3-030-01863-4_5. URL https://doi.org/10.1007/978-3-030-01863-4_5

    work page doi:10.1007/978-3-030-01863-4_5 2019

  4. [4]

    Beyond log^2(t) regret for decentralized bandits in matching markets

    Soumya Basu, Karthik Abinav Sankararaman, and Abishek Sankararaman. Beyond log^2(t) regret for decentralized bandits in matching markets. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 705--715. PMLR, 18--24 Jul 2021

    work page 2021

  5. [5]

    Regret, stability & fairness in matching markets with bandit learners

    Sarah H Cen and Devavrat Shah. Regret, stability & fairness in matching markets with bandit learners. In International Conference on Artificial Intelligence and Statistics, pages 8938--8968. PMLR, 2022

    work page 2022

  6. [6]

    Two-sided bandits and the dating market

    Sanmay Das and Emir Kamenica. Two-sided bandits and the dating market. In IJCAI, volume 5, page 19. Citeseer, 2005

    work page 2005

  7. [7]

    Using stable matching to optimize the balance between accuracy and diversity in recommendation

    Farzad Eskandanian and Bamshad Mobasher. Using stable matching to optimize the balance between accuracy and diversity in recommendation. In Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization, pages 71--79, 2020

    work page 2020

  8. [8]

    College admissions and the stability of marriage

    David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The American Mathematical Monthly, 69 0 (1): 0 9--15, 1962

    1962

Show all 33 references
  1. [9]

    Gerding, Sebastian Stein, Valentin Robu, Dengji Zhao, and Nicholas R

    Enrico H. Gerding, Sebastian Stein, Valentin Robu, Dengji Zhao, and Nicholas R. Jennings. Two-sided online markets for electric vehicle charging. In Proceedings of the 2013 International Conference on Autonomous Agents and Multi-Agent Systems, AAMAS '13, page 989–996. Internat...

    2013

  2. [10]

    Three fast algorithms for four problems in stable marriage

    Dan Gusfield. Three fast algorithms for four problems in stable marriage. SIAM Journal on Computing, 16 0 (1): 0 111--128, 1987

    1987

  3. [11]

    Putting gale & shapley to work: Guaranteeing stability through learning

    Hadi Hosseini, Sanjukta Roy, and Duohan Zhang. Putting gale & shapley to work: Guaranteeing stability through learning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  4. [12]

    An efficient algorithm for the “optimal” stable marriage

    Robert W Irving, Paul Leather, and Dan Gusfield. An efficient algorithm for the “optimal” stable marriage. Journal of the ACM (JACM), 34 0 (3): 0 532--543, 1987

    1987

  5. [13]

    Learning equilibria in matching markets from bandit feedback

    Meena Jagadeesan, Alexander Wei, Yixin Wang, Michael Jordan, and Jacob Steinhardt. Learning equilibria in matching markets from bandit feedback. Advances in Neural Information Processing Systems, 34: 0 3323--3335, 2021

    2021

  6. [14]

    Complexity of the sex-equal stable marriage problem

    Akiko Kato. Complexity of the sex-equal stable marriage problem. Japan Journal of Industrial and Applied Mathematics, 10: 0 1--19, 1993

    1993

  7. [15]

    Mariages stables et leurs relations avec d'autres probl \`e mes combinatoires: introduction \`a l'analyse math \'e matique des algorithmes

    Donald Ervin Knuth. Mariages stables et leurs relations avec d'autres probl \`e mes combinatoires: introduction \`a l'analyse math \'e matique des algorithmes. (No Title), 1976

    1976

  8. [16]

    Matching with couples: Stability and incentives in large markets

    Fuhito Kojima, Parag A Pathak, and Alvin E Roth. Matching with couples: Stability and incentives in large markets. The Quarterly Journal of Economics, 128 0 (4): 0 1585--1632, 2013

    2013

  9. [17]

    Player-optimal stable regret for bandit learning in matching markets

    Fang Kong and Shuai Li. Player-optimal stable regret for bandit learning in matching markets. In Proceedings of the 2023 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1512--1522. SIAM, 2023

    2023

  10. [18]

    Improved bandits in many-to-one matching markets with incentive compatibility

    Fang Kong and Shuai Li. Improved bandits in many-to-one matching markets with incentive compatibility. Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (12): 0 13256--13264, Mar. 2024. doi:10.1609/aaai.v38i12.29226. URL https://ojs.aaai.org/index.php/AAAI/ar...

    2024 doi

  11. [19]

    Bandit algorithms

    Tor Lattimore and Csaba Szepesv \'a ri. Bandit algorithms. Cambridge University Press, 2020

    2020

  12. [20]

    Liu, Horia Mania, and Michael Jordan

    Lydia T. Liu, Horia Mania, and Michael Jordan. Competing bandits in matching markets. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, volume 108 of Proceedings of Machine Learning Research, pages 1618--1628. PMLR, 26--28 Aug 2020

    2020

  13. [21]

    Bandit learning in decentralized matching markets

    Lydia T Liu, Feng Ruan, Horia Mania, and Michael I Jordan. Bandit learning in decentralized matching markets. The Journal of Machine Learning Research, 22 0 (1): 0 9612--9645, 2021

    2021

  14. [22]

    Decentralized, communication- and coordination-free learning in structured matching markets

    Chinmay Maheshwari, Shankar Sastry, and Eric Mazumdar. Decentralized, communication- and coordination-free learning in structured matching markets. In Advances in Neural Information Processing Systems, volume 35, pages 15081--15092. Curran Associates, Inc., 2022

    2022

  15. [23]

    The stable marriage problem

    David G McVitie and Leslie B Wilson. The stable marriage problem. Communications of the ACM, 14 0 (7): 0 486--490, 1971

    1971

  16. [24]

    Deep learning for two-sided matching

    Sai Srivatsa Ravindranath, Zhe Feng, Shira Li, Jonathan Ma, Scott D Kominers, and David C Parkes. Deep learning for two-sided matching. arXiv preprint arXiv:2107.03427, 2021

    2021 arXiv

  17. [25]

    The Evolution of the Labor Market for Medical Interns and Residents: A Case Study in Game Theory

    Alvin E Roth. The Evolution of the Labor Market for Medical Interns and Residents: A Case Study in Game Theory . Journal of Political Economy, 92 0 (6): 0 991--1016, 1984

    1984

  18. [26]

    On the allocation of residents to rural hospitals: a general property of two-sided matching markets

    Alvin E Roth. On the allocation of residents to rural hospitals: a general property of two-sided matching markets. Econometrica: Journal of the Econometric Society, pages 425--427, 1986

    1986

  19. [27]

    The economist as engineer: Game theory, experimentation, and computation as tools for design economics

    Alvin E Roth. The economist as engineer: Game theory, experimentation, and computation as tools for design economics. Econometrica, 70 0 (4): 0 1341--1378, 2002

    2002

  20. [28]

    The Redesign of the Matching Market for American Physicians: Some Engineering Aspects of Economic Design

    Alvin E Roth and Elliott Peranson. The Redesign of the Matching Market for American Physicians: Some Engineering Aspects of Economic Design . American Economic Review, 89 0 (4): 0 748--780, 1999

    1999

  21. [29]

    Dominate or delete: Decentralized competing bandits in serial dictatorship

    Abishek Sankararaman, Soumya Basu, and Karthik Abinav Sankararaman. Dominate or delete: Decentralized competing bandits in serial dictatorship. In Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, volume 130 of Proceedings of Machine L...

    2021

  22. [30]

    Optimal analysis for bandit learning in matching markets with serial dictatorship

    Zilong Wang and Shuai Li. Optimal analysis for bandit learning in matching markets with serial dictatorship. Theoretical Computer Science, page 114703, 2024

    2024

  23. [31]

    Bandit learning in many-to-one matching markets

    Zilong Wang, Liya Guo, Junming Yin, and Shuai Li. Bandit learning in many-to-one matching markets. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 2088--2097, 2022

    2022

  24. [32]

    Decentralized two-sided bandit learning in matching market

    YiRui Zhang and Zhixuan Fang. Decentralized two-sided bandit learning in matching market. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024

    2024

  25. [33]

    Matching in multi-arm bandit with collision

    YiRui Zhang, Siwei Wang, and Zhixuan Fang. Matching in multi-arm bandit with collision. In Advances in Neural Information Processing Systems, volume 35, pages 9552--9563. Curran Associates, Inc., 2022

    2022

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.