Pith. sign in

REVIEW 3 major objections 4 minor 34 references

Learning to Incentivize in Repeated Principal-Agent Problems with Adversarial Agent Arrivals

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves tight minimax regret rates for a repeated principal-agent problem with adversarial agent arrivals: $\tilde{O}(\min\{\sqrt{KT\log N},K\sqrt{T}\})$ when each type's greedy best response is known, and…

desk verdict The adversarial-arrival model and the greedy-setting bounds are real contributions, but the smooth-model matching lower bound breaks on a step discontinuity that violates the paper's own Lipschitz assumption, so the flagship tight-rate claim does not hold as stated. read the letter →

arxiv 2505.23124 v2 pith:ZTRR2YFO submitted 2025-05-29 cs.GT cs.LG

Junyan Liu , Arnab Maiti , Artin Tajdini , Kevin Jamieson , Lillian J. Ratliff This is my paper
classification cs.GTcs.LG MSC 68W2791A26
keywords repeatedprincipal-agentproblemadversarialarrivalsregretminimizationgreedychoicemodelsmoothlinearbanditsincentivedesignLipschitz
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 asks how much regret a principal must pay when $K$ unknown agent types arrive in adversarial order and the principal can only steer their arm choices through incentives. Without any knowledge of how agents respond, the answer is linear regret: no algorithm can compete with the best fixed incentive in hindsight. If the principal knows each type's greedy best-response map, a reduction to adversarial linear bandits achieves $\tilde{O}(\min\{\sqrt{KT\log N}, K\sqrt{T}\})$ regret, and the paper constructs lower-bound instances showing this is tight up to a $\log K$ factor. If agents instead respond smoothly, meaning their arm-selection probabilities are $L$-Lipschitz in total variation with respect to the incentive vector, unknown types can still be handled at $\tilde{O}((LN)^{1/3}T^{2/3})$ regret, with a matching lower bound up to logarithmic factors. These rates delimit when sublinear learning is possible in a model that matches platforms facing arbitrary user arrivals.

What carries the argument

The central object is the one-hot linearization of the principal's payoff: for any incentive $\pi$, the per-type utilities $(U(\pi,1),\dots,U(\pi,K))$ form a vector $z_\pi\in\mathbb{R}^K$, and since the arriving type $j_t$ is encoded by the unit vector $y_t=e_{j_t}$, the round-$t$ payoff is $\langle z_\pi,y_t\rangle$. This carries the argument because it converts the principal-agent problem into an adversarial linear bandit over the finite set $\{z_\pi:\pi\in\Pi\}$, where $\Pi$ is a carefully chosen small incentive set. For known greedy agents, $\Pi$ collects threshold incentives $\pi_{i,j}$ with $\pi_i=\max_k\mu^k_j-\mu^i_j$ and then prunes them by the binary response pattern $h(\pi)\in\{0,1\}^K$, leaving at most $\min\{2KN,2^K\}+1$ representatives; for general incentives, the finite set comes from best-response polytopes $P_\sigma$ and $\varepsilon$-close approximations to their extreme points. For smooth agents, $\Pi$ is simply the $\varepsilon$-grid of single-arm incentives of size $N(\lceil 1/\varepsilon\rceil+1)$, and Assumption 1.1 bounds the per-round loss of replacing the best incentive by its nearest grid point by $2L\varepsilon+\varepsilon$. The lower-bound machinery is KL-divergence based: a Bernoulli KL bound and a chain rule control how few rounds are needed to distinguish two distributions over agent types, and the smooth lower bound reduces the constructed instance to a stochastic multi-armed bandit with Gaussian noise.

What would settle it

Compute the total-variation distance between the response distributions of Eq. (1) at two single-arm incentives that differ only in the $l$-th coordinate, one at $\pi_l=1/2-\delta$ and one at $1/2+\delta$, and take $\delta\to0$: the distance approaches a positive constant on the order of $1/N$, while $\|\pi-\pi'\|_\infty=2\delta\to0$, so Assumption 1.1 fails for any finite $L$. If instead the distribution is repaired by smoothing or renormalizing the jump, the lower-bound instance must be re-checked globally before the claimed $\Omega(L^{1/3}N^{1/3}T^{2/3})$ rate is established.

Watch

Extended reading notes

Core claim

The central discovery is a pair of tight rate boundaries for a repeated principal-agent model in which the sequence of agent types is chosen adversarially and the principal offers one or more arm-specific incentives per round. With greedy agents and known best responses, the per-type utility vector of any incentive defines a linear reward in the one-hot type indicator, so regret against the best incentive in hindsight equals adversarial linear-bandit regret over a finite set of $z_\pi\in\mathbb{R}^K$; the resulting bound is $O(\min\{\sqrt{KT\log N}, K\sqrt{T}\})$ up to log factors, and a matching $\Omega(\min\{\sqrt{KT\log N}/\sqrt{\log K}, K\sqrt{T}/\log K\})$ holds. If the greedy responses are unknown, the paper shows a $\Omega(T)$ wall even with $K=2$ and $N=3$. For smooth agents the paper proves an $\tilde{O}((2L+1)^{1/3}N^{1/3}T^{2/3})$ upper bound by grid-discretizing single-arm incentives and running an adversarial bandit algorithm, and constructs a lower-bound family of response distributions whose optimal incentive sits in a width-$\epsilon$ interval, yielding $\Omega(L^{1/3}N^{1/3}T^{2/3})$; the claimed match is up to log factors. Multi-arm incentives preserve these algorithmic results, with $O(K\sqrt{T\log(KT)})$ for greedy general incentives and $\tilde{O}(L^{N/(N+2)}T^{(N+1)/(N+2)})$ for smooth general incentives.

Load-bearing premise

The smooth lower bound requires the constructed response distribution of Eq. (1) to satisfy Assumption 1.1 for every pair of single-arm incentives; as written, the distribution jumps when a coordinate crosses $1/2$, so the proof's Lipschitz check, which covers only $[0,1/2]$, omits a boundary term that no finite $L$ can absorb.

Editorial extensions

If this is right

  • Known-greedy setting: a principal who knows the best-response map can use single-arm incentives and suffer at most $\tilde{O}(\min\{\sqrt{KT\log N}, K\sqrt{T}\})$ regret, so the per-round loss vanishes as $1/\sqrt{T}$ up to log factors.
  • Unknown greedy behavior: no algorithm can beat linear regret even with only two types and three arms, so sublinear incentive learning requires knowing or estimating responses before or during play.
  • Smooth behavior: with unknown types, the worst-case regret is $\tilde{O}((LN)^{1/3}T^{2/3})$, making the problem easier when responses are flatter (smaller $L$) and harder with more arms $N$.
  • General incentives: allowing multiple simultaneous arm incentives preserves sublinearity, with $O(K\sqrt{T\log(KT)})$ in the greedy case and $\tilde{O}(L^{N/(N+2)}T^{(N+1)/(N+2)})$ in the smooth case.
  • Instance-dependent improvement: when the near-optimal incentive set has adversarial zooming dimension $z$, the smooth bound improves to $O(T^{(z+1)/(z+2)}((2L+1)N)^{z/(z+2)}\log^5 T)$, for example $T^{3/5}$ for strictly concave induced rewards.

Reading between the lines

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

  • A natural middle ground the paper leaves implicit: partial knowledge of greedy responses, such as knowing each type's threshold order up to a perturbation, should yield regret rates between the linear wall and the $\sqrt{KT}$ rate, since the linear-bandit reduction needs only a covering set of plausible response maps.
  • The same one-hot linearization transfers to other per-type payoff structures, so the algorithmic template likely applies to contextual contract or persuasion problems whenever the principal's payoff is linear in the type indicator and the decision space admits an $\varepsilon$-cover.
  • A testable extension is to make $L$ self-estimated: the paper notes that fixed discretization with unknown $L$ costs linear regret, so an adaptive zooming rule that estimates the local Lipschitz constant from observed arm switches is the obvious next step and could realize the instance-dependent $T^{(z+1)/(z+2)}$ bound without knowing $L$ in advance.
  • Should the boundary issue in the smooth lower-bound construction be repaired, the same reduction to stochastic MAB with Gaussian arms suggests the $\Omega((LN)^{1/3}T^{2/3})$ rate is robust; if it is not repaired, the smooth model's true minimax rate may differ, and checking the full-domain Lipschitz condition is the fastest way to tell.
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.

Referee Report

3 major / 4 minor

Summary. The paper studies a repeated principal-agent problem with adversarial arrivals of K agent types, where the principal chooses an incentive vector each round without observing the arriving type. For greedy best-response agents with known response functions, it gives a reduction to adversarial linear bandits leading to regret O~(min{sqrt(KT log N), K sqrt(T)}) for single-arm incentives and O~(K sqrt(T)) for general incentives, together with lower bounds for the single-arm case. For agents whose choice probabilities are L-Lipschitz in the incentive (Assumption 1.1), it proposes a Tsallis-INF discretization achieving O~((LN)^{1/3}T^{2/3}) for single-arm incentives, claims a matching omega(L^{1/3}N^{1/3}T^{2/3}) lower bound, and adds an instance-dependent zooming bound and a general-incentive extension. It also proves linear regret when greedy behavior is completely unknown.

Significance. The paper introduces a natural new model, adversarial agent arrivals in repeated principal-agent problems, and the reduction-based upper bounds are a genuine strength: they use standard adversarial bandit algorithms (EXP3, Tsallis-INF) with explicit dependence on K, N, T, and no fitted constants. The greedy-setting lower bounds in Appendix C are detailed and appear internally coherent, and the general-incentive discretization idea is interesting. If the smooth lower bound were correctly proved, the paper would establish tight rates for both models. As submitted, however, the central matching lower bound in the smooth model is not supported because the constructed response distribution does not satisfy Assumption 1.1.

major comments (3)
  1. [§4.2 / Appendix E.1 (Eq. (1), Lemma E.1)] The lower-bound instance in Theorem 4.1 is not a valid instance of the smooth model. In Eq. (1), for l != i the arm-l selection probability is (1-pi_l)/(16N) when pi_l <= 1/2 but 1/(8N) when pi_l > 1/2. Taking two single-arm incentives pi = (1/2 - delta)e_l and pi' = (1/2 + delta)e_l, the total variation distance between the induced distributions tends to 3/(16N) as delta -> 0, while ||pi - pi'||_inf = 2 delta. Hence no finite L satisfies Assumption 1.1. Lemma E.1's mean-value-theorem calculation applies only on the [0,1/2] branch and silently omits this boundary jump, so its conclusion '<= L Delta' does not follow. Since the matching lower bound advertised in the abstract and in Section 1.2 rests on this construction, the claim is currently unsupported. The construction may be repairable by inserting a Lipschitz transition of width Theta(1/(LN)) at pi_l = 1/2, but as written the proof is invalid.
  2. [Appendix E.1, displayed expected-reward formula] The expected-reward calculation for the good interval is algebraically wrong under Eq. (1). For a single-arm incentive supported on arm i with pi_i in [j*epsilon, (j+1)*epsilon], Eq. (1) gives Pr[a = l] = 1/(16N) for l not in {i, N} and Pr[a = i] = (1 - pi_i)/(16N) + B(pi_i - j*epsilon); therefore E[r(pi)] = [(N-2) + (1 - pi_i)^2]/(16N) + (1 - pi_i)B(pi_i - j*epsilon), not (N-1)/(16N) + (1 - pi_i)B(pi_i - j*epsilon) as displayed. The missing term (2*pi_i - pi_i^2)/(16N) can be as large as 3/(64N) and reduces the claimed gap (L-1)*epsilon/16; with epsilon chosen as in the proof, the gap is not guaranteed to be positive for large T and pi_i near 1/2. The subsequent MAB gap and the final bound need to be re-derived with the correct reward expression.
  3. [Theorem 4.1 / Appendix E.1, quantifiers and final exponent] The theorem statement suppresses K: the construction uses a set J of size Theta(N/epsilon) = Theta(N^{4/3} L^{2/3} T^{1/3}) agent types, so the lower bound can only be asserted for K at least this large; the quantifier over K must be made explicit. Also, the final displayed lower bound in Appendix E.1 says Omega(sum Delta^{-1}_{(i,j)}) = Omega((L-1)^{1/3} N^{1/3} T^{1/3}), whereas Theorem 4.1 claims T^{2/3}; with epsilon and Delta as chosen, sum Delta^{-1} = Theta(N/((L-1) epsilon^2)) = Theta((L-1)^{1/3} N^{1/3} T^{2/3}), so the appendix exponent needs to be corrected and the argument reconciled.
minor comments (4)
  1. [Lemma E.1] In Lemma E.1, the proof sets l(pi) = N+1 when pi = 0, but the subsequent expressions use Pr[a(pi,(i,j)) = N+1], which is not defined; the zero-incentive case should be handled separately.
  2. [§2.1 / Appendix B] The proof of Theorem 2.1 describes the algorithm as acting 'based on the agent numbers j_1,...,j_t', although the principal observes only the chosen arm and not the agent type; the argument should be phrased over arm histories, which is the only information available to the algorithm.
  3. [§3.1 / Theorem 3.1] The abstract states a regret bound of O~(min{sqrt(KT log N), K sqrt(T)}), but the proof uses the arm set Z of size at most min{2KN, 2^K} + 1 and therefore obtains sqrt(KT log(KN)) in the first term; the log(KN) versus log(N) discrepancy should be stated consistently.
  4. [Appendix C.2] The proof assumes M and K_0/M are integers and K >= 1000 'for simplicity', but the theorem is stated for all K, N; a rounding argument is needed to remove these assumptions, or the theorem should state the extra conditions.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; results are derived from external bandit algorithms and standard KL arguments, with only minor technical self-citations.

full rationale

The central upper bounds are reductions to adversarial linear bandits (EXP3) or direct Tsallis-INF discretizations, with regret bounds following from standard external analyses (Lattimore & Szepesvari, Zimmert & Seldin, Podimata & Slivkins) rather than from fitted constants or from the target rates themselves. The greedy lower bounds use standard MAB lower-bound and KL-divergence arguments; the only self-citations (Maiti et al. 2023, 2025) supply a Bernoulli KL bound and a chain-rule decomposition, which are technical tools rather than the theorems being proved. The smooth-model lower bound's construction (Eq. 1) may violate Assumption 1.1 due to a jump at pi_l = 1/2, but that is a correctness/validity concern, not circularity: it does not reduce a predicted rate to an input fit or to a self-citation. No central claim is defined in terms of its own conclusion, so the paper is not circular in any load-bearing sense; the score reflects only the presence of minor, non-load-bearing self-citations.

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

No parameters are fitted to data; the regret bounds are minimax. The main load-bearing premise is the Lipschitz smoothness assumption, and the lower-bound proof violates it at the threshold.

assumptions (5)
  • domain assumption The principal knows the full best response function b(pi,j) for every agent type in the greedy setting.
    Section 1.1 and Section 3 require this knowledge to construct the finite incentive set Pi and to run the reduction to linear bandits; without it, Theorem 2.1 shows linear regret.
  • domain assumption Tie-breaking among equal-utility arms is fixed, consistent, and known, such as preferring the incentivized arm or using a hierarchical tie-break.
    Section 3.1 and Appendix D rely on predicting the selected arm at threshold incentives; the paper states the assumption can be relaxed, but the presented proofs use it.
  • domain assumption In the smooth model, the choice probability vector satisfies the L-Lipschitz condition in Assumption 1.1.
    This is the defining assumption of the smooth setting; all smooth results depend on it.
  • ad hoc to paper The constructed response distribution in the smooth lower bound satisfies Assumption 1.1 over the full single-arm domain D=[0,1]^N.
    This is asserted implicitly in Appendix E.1 but is false: Eq. (1) has a discontinuity at pi_l=1/2, so no finite L can satisfy the Lipschitz condition across that threshold.
  • standard math Standard lower-bound tools such as KL divergence bounds, Pinsker's inequality, and Yao's lemma are valid.
    These are used in the lower-bound proofs in Appendices A, B, and C; they are standard and not contested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Incentivize in Repeated Principal-Agent Problems with Adversarial Agent Arrivals." pith.science (2026). https://pith.science/paper/ZTRR2YFO

@misc{pith2026250523124,
  author       = {Pith},
  title        = {Pith review of: Learning to Incentivize in Repeated Principal-Agent Problems with Adversarial Agent Arrivals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZTRR2YFO}},
  note         = {Machine review of arXiv:2505.23124}
}
abstract

We initiate the study of a repeated principal-agent problem over a finite horizon $T$, where a principal sequentially interacts with $K\geq 2$ types of agents arriving in an adversarial order. At each round, the principal strategically chooses one of the $N$ arms to incentivize for an arriving agent of unknown type. The agent then chooses an arm based on its own utility and the provided incentive, and the principal receives a corresponding reward. The objective is to minimize regret against the best incentive in hindsight. Without prior knowledge of agent behavior, we show that the problem becomes intractable, leading to linear regret. We analyze two key settings where sublinear regret is achievable. In the first setting, the principal knows the arm each agent type would select greedily for any given incentive. Under this setting, we propose an algorithm that achieves a regret bound of $O(\min\{\sqrt{KT\log N},K\sqrt{T}\})$ and provide a matching lower bound up to a $\log K$ factor. In the second setting, an agent's response varies smoothly with the incentive and is governed by a Lipschitz constant $L\geq 1$. Under this setting, we show that there is an algorithm with a regret bound of $\tilde{O}((LN)^{1/3}T^{2/3})$ and establish a matching lower bound up to logarithmic factors. Finally, we extend our algorithmic results for both settings by allowing the principal to incentivize multiple arms simultaneously in each round.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Balcan, M.-F., Bernasconi, M., Castiglioni, M., Celli, A., Harris, K., and Wu, Z. S. Nearly-optimal bandit learning in stackelberg games with side information. arXiv preprint arXiv:2502.00204, 2025

    arXiv 2025

  3. [3]

    Principal-agent reward shaping in mdps

    Ben-Porat, O., Mansour, Y., Moshkovitz, M., and Taitler, B. Principal-agent reward shaping in mdps. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 9502--9510, 2024

    work page 2024

  4. [4]

    Optimal rates and efficient algorithms for online bayesian persuasion

    Bernasconi, M., Castiglioni, M., Celli, A., Marchesi, A., Trov \`o , F., and Gatti, N. Optimal rates and efficient algorithms for online bayesian persuasion. In International Conference on Machine Learning, pp.\ 2164--2183. PMLR, 2023

    work page 2023

  5. [5]

    and Dewatripont, M

    Bolton, P. and Dewatripont, M. Contract theory. MIT press, 2004

    work page 2004

  6. [6]

    Robustness and linear contracts

    Carroll, G. Robustness and linear contracts. American Economic Review, 105 0 (2): 0 536--563, 2015

    work page 2015

  7. [7]

    Learning to Price Homogeneous Data

    Chen, K., Huh, J. S., and Kandasamy, K. Learning to price homogeneous data. arXiv preprint arXiv:2407.05484, 2024

    work page Pith review arXiv 2024

  8. [8]

    M., and Aswani, A

    Dogan, I., Shen, Z.-J. M., and Aswani, A. Estimating and incentivizing imperfect-knowledge agents with hidden rewards. arXiv preprint arXiv:2308.06717, 2023 a

    arXiv 2023

Show all 34 references
  1. [9]

    M., and Aswani, A

    Dogan, I., Shen, Z.-J. M., and Aswani, A. Repeated principal-agent games with unobserved agent rewards and perfect-knowledge agents. arXiv preprint arXiv:2304.07407, 2023 b

    2023 arXiv

  2. [10]

    Multi-agent combinatorial contracts

    Duetting, P., Ezra, T., Feldman, M., and Kesselheim, T. Multi-agent combinatorial contracts. In Proceedings of the 2025 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pp.\ 1857--1891. SIAM, 2025

    2025

  3. [11]

    Simple versus optimal contracts

    D \"u tting, P., Roughgarden, T., and Talgam-Cohen, I. Simple versus optimal contracts. In Proceedings of the 2019 ACM Conference on Economics and Computation, pp.\ 369--387, 2019

    2019

  4. [12]

    Combinatorial contracts

    D \"u tting, P., Ezra, T., Feldman, M., and Kesselheim, T. Combinatorial contracts. In 2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS), pp.\ 815--826. IEEE, 2022

    2021

  5. [13]

    Multi-agent contracts

    D \"u tting, P., Ezra, T., Feldman, M., and Kesselheim, T. Multi-agent contracts. In Proceedings of the 55th Annual ACM Symposium on Theory of Computing, pp.\ 1311--1324, 2023

    2023

  6. [14]

    Fiez, T., Sekar, S., Zheng, L., and Ratliff, L. J. Combinatorial bandits for incentivizing agents with dynamic preferences. Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI), 2018

    2018

  7. [15]

    and Miller, R

    Gayle, G.-L. and Miller, R. A. Identifying and testing models of managerial compensation. The Review of Economic Studies, 82 0 (3): 0 1074--1118, 2015

    2015

  8. [16]

    Grossman, S. J. and Hart, O. D. An analysis of the principal-agent problem. In Foundations of Insurance Economics: Readings in Economics and Finance, pp.\ 302--340. Springer, 1992

    1992

  9. [17]

    Z., and Balcan, M.-F

    Harris, K., Wu, S. Z., and Balcan, M.-F. F. Regret minimization in stackelberg games with side information. Advances in Neural Information Processing Systems, 37: 0 12944--12976, 2024

    2024

  10. [18]

    Ho, C.-J., Slivkins, A., and Vaughan, J. W. Adaptive contract design for crowdsourcing markets: Bandit algorithms for repeated principal-agent problems. In Proceedings of the fifteenth ACM conference on Economics and computation, pp.\ 359--376, 2014

    2014

  11. [19]

    Moral hazard and observability

    Holmstr \"o m, B. Moral hazard and observability. The Bell journal of economics, pp.\ 74--91, 1979

    1979

  12. [20]

    J., Rand, D

    Horton, J. J., Rand, D. G., and Zeckhauser, R. J. The online laboratory: Conducting experiments in a real labor market. Experimental economics, 14: 0 399--425, 2011

    2011

  13. [21]

    and Siddiq, A

    Kaynar, N. and Siddiq, A. Estimating effects of incentive contracts in online labor platforms. Management Science, 69 0 (4): 0 2106--2126, 2023

    2023

  14. [22]

    and Szepesv \'a ri, C

    Lattimore, T. and Szepesv \'a ri, C. Bandit algorithms. Cambridge University Press, 2020

    2020

  15. [23]

    and Ratliff, L

    Liu, J. and Ratliff, L. J. Principal-agent bandit games with self-interested and exploratory learning agents. arXiv preprint arXiv:2412.16318, 2024

    2024 arXiv

  16. [24]

    Maiti, A., Jamieson, K., and Ratliff, L. J. Logarithmic regret for matrix games against an adversary with noisy bandit feedback. arXiv preprint arXiv:2306.13233, 2023

    2023 arXiv

  17. [25]

    J., and Farina, G

    Maiti, A., Fan, Z., Jamieson, K., Ratliff, L. J., and Farina, G. Efficient near-optimal algorithm for online shortest paths in directed acyclic graphs with bandit feedback against adaptive adversaries. arXiv preprint arXiv:2504.00461, 2025

    2025 arXiv

  18. [26]

    and Nair, H

    Misra, S. and Nair, H. S. A structural model of sales-force compensation dynamics: Estimation and field implementation. Quantitative Marketing and Economics, 9: 0 211--257, 2011

    2011

  19. [27]

    T., and Narasimhan, C

    Misra, S., Coughlan, A. T., and Narasimhan, C. Salesforce compensation: An analytical and empirical examination of the agency theoretic approach. Quantitative Marketing and Economics, 3: 0 5--39, 2005

    2005

  20. [28]

    and Slivkins, A

    Podimata, C. and Slivkins, A. Adaptive discretization for adversarial lipschitz bandits. In Proceedings of Thirty Fourth Conference on Learning Theory, volume 134, pp.\ 3788--3805, 2021

    2021

  21. [29]

    I., and Durmus, A

    Scheid, A., Capitaine, A., Boursier, E., Moulines, E., Jordan, M. I., and Durmus, A. Learning to mitigate externalities: the coase theorem with hindsight rationality. arXiv preprint arXiv:2406.19824, 2024 a

    2024 arXiv

  22. [30]

    Scheid, A., Tiapkin, D., Boursier, E., Capitaine, A., Mhamdi, E. M. E., Moulines, \'E ., Jordan, M. I., and Durmus, A. Incentivized learning in principal-agent bandit games. arXiv preprint arXiv:2403.03811, 2024 b

    2024 arXiv

  23. [31]

    Train, K. E. Discrete Choice Methods with Simulation. Cambridge University Press, 2 edition, 2009

    2009

  24. [32]

    Contractual reinforcement learning: Pulling arms with invisible hands

    Wu, J., Chen, S., Wang, M., Wang, H., and Xu, H. Contractual reinforcement learning: Pulling arms with invisible hands. arXiv preprint arXiv:2407.01458, 2024

    2024 arXiv

  25. [33]

    Zhu, B., Bates, S., Yang, Z., Wang, Y., Jiao, J., and Jordan, M. I. The sample complexity of online contract design. arXiv preprint arXiv:2211.05732, 2022

    2022 arXiv

  26. [34]

    and Seldin, Y

    Zimmert, J. and Seldin, Y. Tsallis-inf: An optimal algorithm for stochastic and adversarial bandits. Journal of Machine Learning Research, 22 0 (28): 0 1--49, 2021

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.