Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

Learning to Price with Resource Constraints: From Full Information to Machine-Learned Prices

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Re-solving the fluid model with boundary attraction (rounding tiny target demands to zero) gives $O(\log T)$ regret without non-degeneracy; forced exploration matches the optimal $O(\sqrt{T})$, and informed prices interpolate.

desk verdict The full-information boundary-attraction result is a genuine advance, but the no-information theorem is not proven for Algorithm 2 as written: the proof analyzes uncensored demands while the algorithm rejects some of them. read the letter →

arxiv 2501.14155 v1 pith:2WCZ5IPH submitted 2025-01-24 math.OC cs.LG

Ruicheng Ao , Jiashuo Jiang , David Simchi-Levi This is my paper · ORCID
classification math.OCcs.LG MSC 90B5091B2490C20
keywords onlinelearningdynamicpricingresourceconstraintsinformedpriceregretboundsfluidapproximationnon-degeneracylineardemandmodel
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

The paper studies a seller who posts prices for $n$ products over $T$ periods, where demand is linear in price, noisy, and each sale consumes limited shared resources. Its central claim is that one algorithmic template—solve the deterministic 'fluid' capacity problem, then round any product whose target demand is tiny down to zero—achieves logarithmic regret under full information even when the optimal fluid solution is degenerate, a case where earlier re-solving methods required a non-degeneracy assumption and could fail. When the demand model is unknown, the same template with deliberate price perturbations learns the parameters online and hits the worst-case $O(\sqrt{T})$ regret, which is optimal. When a pre-trained informed price-demand pair with known error bound $\varepsilon_0$ is available, the regret becomes $O(\min\{\rho\sqrt{T},\,(\varepsilon_0)^2 T\}+\log T)$, so accurate offline predictions recover the full-information logarithmic rate. The practical point is that a seller can select the right pricing policy based on how much it knows and how good its offline estimates are.

What carries the argument

The carrying object is the 'boundary attracted re-solve' step: at time $t$, solve the fluid program with remaining capacity $c_t$ and per-period budget $c_t/(T-t+1)$, then threshold the target demand at $\zeta(T-t+1)^{-1/2}$—coordinates below it are rounded to zero and the price is set to the modified expected demand. The threshold does double duty. It keeps every nonzero demand component large enough that zero-mean sub-Gaussian noise cannot make the realized demand flip the sign of $d^{\pi,t}_i - \epsilon_i^t/(T-t)$, and it deliberately parks the solution away from the degenerate face where the optimal basis is not unique. In the learning setting the same threshold is paired with an exploration perturbation $\sigma_0 t^{-1/4} e_{t-kn}$ around the periodic re-solve anchor, which is what keeps the design matrix in the least-squares problem well conditioned and makes the estimation error shrink. The regret analysis is carried by a hybrid-policy decomposition: compare the policy that follows the algorithm up to time $t$ and then uses noiseless fluid solutions, and bound each single-step gap.

What would settle it

Simulate Algorithm 2 on a degenerate two-product instance and record whether the demands zeroed by the rounding rule are stored for regression. If rejected observations are excluded, refit the parameters with and without them; an estimator bias large enough to push the empirical regret past the claimed $O(\sqrt{T})$ rate would show the proof relies on unstated uncensored-regression access. Separately, compute the minimum eigenvalue of the design matrix $P^t$ over the horizon and compare it with the claimed $\sigma_0^2\sqrt{kn}/(8n)$ lower bound; a persistent violation would invalidate inequality (35) and with it the estimation-error argument.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that degeneracy is not an obstacle if the re-solver pushes the fluid solution away from the boundary before noise acts. Algorithm 1 solves the fluid program at each remaining-horizon length, and any demand coordinate below $\zeta(T-t+1)^{-1/2}$ is set to zero before converting the demand target to a price. Because every surviving demand is large relative to the noise that could cancel it, the one-step loss of the policy relative to a hybrid noiseless benchmark is bounded by a term of order $1/(T-t)$, and these terms sum to $O(\zeta^2 n^2\|B^{-1}\|_2\log T)$. The no-information variant (Algorithm 2) re-estimates $\alpha,B$ by least squares every $n$ periods, prices as a momentum term plus a $\sigma_0 t^{-1/4}$ exploration kick, and rejects estimated low-demand types; its regret is $O((\zeta^2+\|B^{-1}\|_2^2)\sqrt{T})$, matching the established lower bound. Algorithm 3 uses an informed pair $(p_0,d_0)$ with known error $\varepsilon_0$ as an anchor, obtaining $O(\min\{\rho\sqrt{T},(\varepsilon_0)^2 T\}+C'\log T)$ and, when $\varepsilon_0=0$, recovering Algorithm 1's logarithmic bound.

Load-bearing premise

The no-information statements stand on the assumption that the exploration schedule makes the least-squares design matrix grow at the claimed rate, and on the assumption that the regression is fed uncensored realized demands; if either fails, the estimation-error control and the $O(\sqrt{T})$ regret bound no longer follow.

Editorial extensions

If this is right

  • A seller with exact knowledge of a linear demand model can price with only $O(\log T)$ regret even when the fluid optimal solution is degenerate, so the previously standard non-degeneracy condition is not needed.
  • With no prior information, $O(\sqrt{T})$ regret is the best any policy can achieve, so Algorithm 2 is worst-case optimal for the linear-demand knapsack pricing problem.
  • Knowing an upper bound $\varepsilon_0$ on the error of a machine-learned informed price pair changes the achievable rate: regret is $O((\varepsilon_0)^2 T+\log T)$ when the oracle is accurate and degrades gracefully to $O(\sqrt{T})$ when it is not.
  • Without a known bound on the oracle error, no policy can uniformly beat the no-information $\Omega(\sqrt{T})$ worst case, so the error bound is what converts offline data into a provable speed-up.
  • The estimate-then-select switch compares $(\varepsilon_0)^2 T$ with $\rho\sqrt{T}$ at the outset and selects between the learning policy and the informed-pair policy, giving a direct operational rule for when to trust offline estimates.

Reading between the lines

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

  • The boundary-attraction threshold behaves like a shrinking confidence interval, so the same rounding idea could transfer to non-linear demand models or to other continuous-action resource-constrained bandits where degenerate faces make estimation unstable.
  • The estimate-then-select comparison suggests a general meta-principle: trust an offline oracle only when its error cost $(\varepsilon_0)^2 T$ is smaller than the cost $\sqrt{T}$ of pure online learning, a rule that could apply to other predict-then-optimize settings.
  • A natural robustness test is to run Algorithm 2 with and without recording the demands that the algorithm rejects before updating inventory; if the stated $O(\sqrt{T})$ bound survives both variants the mechanism is stronger than the proof, and if it does not, a record-then-reject modification would repair the gap.
  • The exploration identity (35) can be checked directly by simulating the price sequence and measuring whether the minimum eigenvalue of the design matrix actually grows like $\sqrt{kn}$; a violation would locate exactly where the no-information argument needs extra proof.
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

4 major / 6 minor

Summary. The paper studies a T-period dynamic pricing problem with n products and m resource constraints under a linear demand model with sub-Gaussian noise. It proposes three policies: Algorithm 1, a full-information re-solve heuristic with a boundary-attraction threshold claimed to achieve O(ζ²n²‖B^{-1}‖² log T) regret without the non-degeneracy condition; Algorithm 2, a periodic-review re-solve heuristic with linear-regression parameter learning and price perturbations claimed to achieve O((ζ²+‖B^{-1}‖²)√T) regret with no prior information; and Algorithm 3, an estimate-then-select method using an informed price-demand pair with known error bound ε0 claimed to achieve O(min{ρ√T,(ε0)²T}+log T). The paper also reports numerical experiments and sketches proofs based on hybrid-policy regret decompositions, concentration inequalities, and stability of quadratic programs.

Significance. If the results were correct, the full-information logarithmic regret without non-degeneracy would be a meaningful extension of the re-solve heuristic literature, and the no-information √T bound would match the known lower bound. The informed-price interpolation is conceptually attractive and gives a falsifiable prediction about a phase transition in ε0. The paper makes a good-faith effort to connect with previous work and to validate the algorithms numerically. However, the proof of the headline no-information theorem does not analyze the algorithm as written, and the key exploration lower bound is derived by an invalid identity; these are load-bearing issues, so the significance is conditional on a substantial revision.

major comments (4)
  1. [Section 3, Algorithm 2, and Appendix B] Theorem 3.1 is not proven for Algorithm 2 as written. In the algorithm, lines 15–17 reject all types i∈I^t_r and update the inventory with the accepted demands, and the regression input for (6) is the observed demand history; in the proof, however, the realized demand is modeled as d_t=f(p_t)+ε_t for every t in the single-block regret decomposition (41), in the inventory recursion, and in the estimation error bound (36). The set I^t_r never appears in the proof. If the algorithm regresses on post-rejection (zeroed or 'accepted') demands, estimator (6) is biased; if it uses the uncensored realized demand, then the revenue and inventory updates in (41) do not match the algorithm's execution. In either case the O(√T) regret bound is not established for the stated algorithm. Algorithm 3 has the same rejection structure (lines 11–14), so Theorem 4.3 inherits the gap.
  2. [Appendix B, Eq. (35)] The lower bound J_k ≥ σ0²√(kn)/(8n) is derived from the identity J_k = n^{-1} Σ_{l=1}^k Σ_{i=1}^n (1−1/(n(l−1)+i)) σ0²(kn+i)^{-1/2}. This identity is false: the term p_s−p_{s−1}−(p̃_k−p_{kn}) equals the perturbation σ0 s^{-1/4} e_{s−kn} only for s in the block near kn; for earlier blocks the anchor is p̃_l−p_{ln}, not p̃_k−p_{kn}. The sum therefore does not reduce to the displayed perturbation sum, and the eigenvalue lower bound used in Lemma B.1 and Eq. (36) is unsupported. Since this bound controls the parameter estimation error, the proof of Theorem 3.1 depends on an invalid step.
  3. [Proposition 4.4 and Appendix D] Proposition 4.4 is misstated. It says 'without knowledge of the accuracy ε0' and claims regret O(min{ρ√T,(ε0)²T}) on instance (α,B) and Ω(max{ρ√T,(ε0)²T}) on (α′,B′), but this is not the lower-bound statement needed for Theorem 4.3, which assumes the upper bound ε0 is known. The proof in Appendix D merely refers to Proposition 4.1 and chooses ε of order T^{-(1−γ)/2}; it does not construct two instances with a fixed known ε0 and does not establish the claimed min/max trade-off. The proposition needs to be restated and proved.
  4. [Appendix A, Eq. (21) and Theorem 2.1] The proof of Theorem 2.1 has a condition mismatch and a tail-bound error. Theorem 2.1 assumes ζ ≥ 4σ², but the derivation of (21) requires ζ ≥ 4σ² log n and concludes P((E^t)^c) ≤ n exp(−2(T−t) log n) = exp(−2(T−t)); the intermediate inequality is dimensionally inconsistent because the sub-Gaussian tail of max_i ε_i^t depends on ζ²/(2σ²), not on ζ(T−t)/(2σ²). Thus the logarithmic regret bound for Algorithm 1 is not established as written.
minor comments (6)
  1. [Appendix A, cross-references] The concentration inequality is called Lemma A.2 but is cited as Lemma B.1 before Eq. (21); the cross-reference and the exponent in the sub-Gaussian tail should be corrected.
  2. [Appendix C, Eq. (55) vs Algorithm 3 line 9] The proof uses the perturbation σ0 t^{-1/2} B e_l for Δ^t_{II}, while Algorithm 3 line 9 defines the price perturbation as σ0 sgn(p̃_t−p0) t^{-1/4} e_l; these two expressions must be reconciled for the proof to describe the algorithm.
  3. [Assumption 1.2 and Lemma B.3] Assumption 1.2 calls B 'negative definite' but states λmin(B+B^⊤)>0, which is a positive-definiteness condition; the later use of λmin(B^{-1}+B^{-⊤}) in Lemma B.3 should be checked for sign consistency.
  4. [Section 5, Figures] Figure 4 has T on the x-axis while the caption describes different ε0 values, and Figure 2 is described as a log-log plot but the axes are not labeled as logarithmic; the figures should be re-labeled or captioned to match the text.
  5. [References and formatting] The Banerjee and Freund (2020) reference contains placeholder text 'sc.', and the reference list has several formatting errors (e.g., 'Wainwright 2019' vs the full entry); these should be cleaned up.
  6. [Section 1.3] There are typos such as 'CwK' for 'BwK' and 'ot' for 'to'; a careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the core regret bounds are derived against the external fluid benchmark, and the only self-citation is non-load-bearing provenance.

full rationale

The paper's three main regret bounds are derived against the fluid benchmark T r* (Proposition 1.3) and compared with external lower bounds (Lemma 3.3 from Keskin and Zeevi 2014), not against quantities fitted from the algorithm's own outputs. The regression estimators (6) and (9) are standard least-squares estimators; the proof then converts their estimation error into regret via stability lemmas (Lemma 3.2/B.2 from Bonnans and Shapiro 2013) and concentration bounds (Lemmas B.1/C.1 from Keskin and Zeevi 2014). No parameter is fitted to a subset of the data and then reported as a prediction of a closely related quantity; the regret bound is a theorem in terms of exogenous quantities such as zeta, sigma_0, B^{-1}, and epsilon_0. The self-citation to Jiang et al. (2022) for the single-step difference technique is present in the proof sketch of Theorem 2.1, but the technique is re-derived in Appendix A (Eqs. (12)-(29)), so the citation is not load-bearing. The serious gaps in Appendix B, namely Eq. (35) using the current block's anchor for all past blocks and the proof analyzing the uncensored process d_t = f(p_t) + epsilon_t while Algorithm 2 lines 15-17 reject types and update inventory with accepted demands, are correctness risks rather than circularity, because they do not make the theorem's output equivalent to its input by construction. Thus no circular reduction is exhibited; the appropriate score is low (2) to acknowledge the minor non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

Main assumptions are the linear demand model, negative definiteness of B, sub-Gaussian noise, and Assumption 4.2 on the informed price-demand pair. The no-information proof additionally relies on an orthogonality property of the exploration schedule (Eq. 35) that is not valid as written, and on an unstated censoring assumption for rejected demands. No invented entities are introduced.

free parameters (3)
  • zeta = zeta = 1 in experiments
    Rounding threshold in Algorithms 1-3; chosen by hand and constrained by lower bounds in Theorems 2.1 and 3.1; it appears in all regret bounds.
  • sigma_0 = sigma_0 = 1 in experiments
    Exploration perturbation scale in Algorithms 2 and 3; chosen by hand and appears in the estimation error and regret bounds.
  • rho = rho = 0.1 in experiments
    Tolerance parameter in Algorithm 3 that decides whether to trust the informed price or switch to Algorithm 2; appears in the regret bound.
assumptions (6)
  • domain assumption Linear demand model f(p) = alpha + B p, Assumption 1.1
    The entire fluid model, re-solve problems, and regressions assume demand is affine in price.
  • domain assumption B is negative definite, Assumption 1.2
    Used to guarantee invertibility, strong concavity of the fluid objective, and the continuity lemmas.
  • domain assumption Noise is zero-mean sub-Gaussian, Section 1.1
    All concentration bounds and regret estimates rely on the sub-Gaussian tail inequality.
  • domain assumption Informed price-demand pair satisfies d_0 > d*, Assumption 4.2
    The informed price is assumed to come from discounts or promotions, which is used to lower bound the variance J^t in the informed-price proof.
  • ad hoc to paper Exploration schedule yields orthogonal basis and Eq. (35)
    The proof of Theorem 3.1 requires lambda_min(P^t) to grow like sigma_0^2 sqrt(kn)/(8n); the equality in Eq. (35) is not valid for blocks before the current one.
  • ad hoc to paper Rejected low-demand observations do not bias the regression
    Algorithm 2 rejects low estimated demand types and updates accepted demands before using data in regression (6); the paper never states whether regression uses uncensored realized demands.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Price with Resource Constraints: From Full Information to Machine-Learned Prices." pith.science (2026). https://pith.science/paper/2WCZ5IPH

@misc{pith2026250114155,
  author       = {Pith},
  title        = {Pith review of: Learning to Price with Resource Constraints: From Full Information to Machine-Learned Prices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2WCZ5IPH}},
  note         = {Machine review of arXiv:2501.14155}
}
abstract

We study the dynamic pricing problem with knapsack, addressing the challenge of balancing exploration and exploitation under resource constraints. We introduce three algorithms tailored to different informational settings: a Boundary Attracted Re-solve Method for full information, an online learning algorithm for scenarios with no prior information, and an estimate-then-select re-solve algorithm that leverages machine-learned informed prices with known upper bound of estimation errors. The Boundary Attracted Re-solve Method achieves logarithmic regret without requiring the non-degeneracy condition, while the online learning algorithm attains an optimal $O(\sqrt{T})$ regret. Our estimate-then-select approach bridges the gap between these settings, providing improved regret bounds when reliable offline data is available. Numerical experiments validate the effectiveness and robustness of our algorithms across various scenarios. This work advances the understanding of online resource allocation and dynamic pricing, offering practical solutions adaptable to different informational structures.

Figures

Figures reproduced from arXiv: 2501.14155 by the authors.

Figure 1
Figure 1. Regret of our algorithms under different number of time horizon [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Regret of our algorithms under different number of time horizon [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Regret of our algorithms under different number of time horizon [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Regret of our algorithms under different number of time horizon [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Optimizing the Preconditioner: A Black-box Online-to-Nonconvex Conversion with Static Regret Minimization Oracles

    cs.LG 2026-07 conditional novelty 7.0 of 10

    An OCO algorithm with only O(√T) static regret, pluggable as a preconditioner selector, recovers the classical O(1/√T) stationarity rate on smooth stochastic nonconvex problems and the O(T^{-2/7}) rate on nonsmooth ones.

  2. Adaptive Resolving Methods for Reinforcement Learning with Function Approximations

    cs.LG 2025-05 conditional novelty 6.0 of 10

    An LP-based RL algorithm that identifies a small optimal basis and then resolves the corresponding linear equations adaptively, achieving an instance-dependent ~O(1/N) suboptimality gap for favorable instances.

  3. Constrained Online Decision-Making: A Unified Framework

    stat.ML 2025-05 reject novelty 5.0 of 10

    A general framework and algorithm for constrained contextual online decision-making with regret bounds expressed in terms of a generalized eluder dimension and an offline density estimation oracle.

Reference graph

Works this paper leans on

53 extracted references · 31 canonical work pages · cited by 3 Pith papers

  1. [1]

    Abbasi-Yadkori, Y., P \'a l, D., and Szepesv \'a ri, C. (2011). Improved algorithms for linear stochastic bandits. Advances in neural information processing systems , 24

    2011

  2. [2]

    and Devanur, N

    Agrawal, S. and Devanur, N. (2016). Linear contextual bandits with knapsacks. Advances in neural information processing systems , 29

    work page 2016

  3. [3]

    R., and Li, L

    Agrawal, S., Devanur, N. R., and Li, L. (2016). An efficient algorithm for contextual bandits with knapsacks, and an extension to concave objectives. In Conference on Learning Theory , pages 4--18. PMLR

    work page 2016

  4. [4]

    Amin, K., Rostamizadeh, A., and Syed, U. (2014). Repeated contextual auctions with strategic buyers. Advances in Neural Information Processing Systems , 27

    work page 2014

  5. [5]

    Ao, R., Chen, H., Simchi-Levi, D., and Zhu, F. (2024a). Online local false discovery rate control: A resource allocation approach. arXiv preprint arXiv:2402.11425

    arXiv 2024

  6. [6]

    Ao, R., Fu, H., and Simchi-Levi, D. (2024b). Two-stage online reusable resource allocation: Reservation, overbooking and confirmation call. arXiv preprint arXiv:2410.15245

    arXiv 2024

  7. [7]

    Auer, P. (2002). Using confidence bounds for exploitation-exploration trade-offs. Journal of Machine Learning Research , 3(Nov):397--422

    work page 2002

  8. [8]

    Badanidiyuru, A., Langford, J., and Slivkins, A. (2014). Resourceful contextual bandits. In Conference on Learning Theory , pages 1109--1134. PMLR

    work page 2014

Show all 53 references
  1. [9]

    Balseiro, S., Lu, H., and Mirrokni, V. (2020). Dual mirror descent for online allocation problems. In International Conference on Machine Learning , pages 613--628. PMLR

    2020

  2. [10]

    and Freund, D

    Banerjee, S. and Freund, D. (2020). sc. In Abstracts of the 2020 SIGMETRICS/Performance Joint International Conference on Measurement and Modeling of Computer Systems , pages 1--2

    2020

  3. [11]

    and Zeevi, A

    Besbes, O. and Zeevi, A. (2015). On the (surprising) sufficiency of linear models for dynamic pricing with demand learning. Management Science , 61(4):723--739

    2015

  4. [12]

    Bonnans, J. F. and Shapiro, A. (2013). Perturbation analysis of optimization problems . Springer Science & Business Media

    2013

  5. [13]

    and Vandenberghe, L

    Boyd, S. and Vandenberghe, L. (2004). Convex optimization . Cambridge university press

    2004

  6. [14]

    and Rusmevichientong, P

    Broder, J. and Rusmevichientong, P. (2012). Dynamic pricing under a general parametric choice model. Operations Research , 60(4):965--980

    2012

  7. [15]

    Bu, J., Simchi-Levi, D., and Wang, C. (2022). Context-based dynamic pricing with partially linear demand model. Advances in Neural Information Processing Systems , 35:23780--23791

    2022

  8. [16]

    Bu, J., Simchi-Levi, D., and Xu, Y. (2020). Online pricing with offline data: Phase transition and inverse square law. In International Conference on Machine Learning , pages 1202--1210. PMLR

    2020

  9. [17]

    and Wang, H

    Bumpensanti, P. and Wang, H. (2020). A re-solving heuristic with uniformly bounded loss for network revenue management. Management Science , 66(7):2993--3009

    2020

  10. [18]

    Chen, Q., Jasin, S., and Duenyas, I. (2021). Joint learning and optimization of multi-product pricing with finite resource capacity and unknown demand parameters. Operations Research , 69(2):560--573

    2021

  11. [19]

    G., Wang, H., and Wang, Z

    Chen, Q. G., Wang, H., and Wang, Z. (2022). Learning and Pricing with Inventory Constraints , volume 18, pages 103--135. Springer Nature

    2022

  12. [20]

    Chen, Z., Ai, R., Yang, M., Pan, Y., Wang, C., and Deng, X. (2024). Contextual decision-making with knapsacks beyond the worst case. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  13. [21]

    Cheung, W. C. and Lyu, L. (2024). Leveraging (biased) information: Multi-armed bandits with offline data. arXiv preprint arXiv:2405.02594

    2024

  14. [22]

    C., Lobel, I., and Paes Leme, R

    Cohen, M. C., Lobel, I., and Paes Leme, R. (2020). Feature-based dynamic pricing. Management Science , 66(11):4921--4943

    2020

  15. [23]

    Dudik, M., Hsu, D., Kale, S., Karampatziakis, N., Langford, J., Reyzin, L., and Zhang, T. (2011). Efficient optimal learning for contextual bandits. arXiv preprint arXiv:1106.2369

    2011 arXiv

  16. [24]

    predict, then optimize

    Elmachtoub, A. N. and Grigas, P. (2022). Smart “predict, then optimize”. Management Science , 68(1):9--26

    2022

  17. [25]

    J., Simchi-Levi, D., and Wang, H

    Ferreira, K. J., Simchi-Levi, D., and Wang, H. (2018). Online network revenue management using thompson sampling. Operations research , 66(6):1586--1602

    2018

  18. [26]

    and Van Ryzin, G

    Gallego, G. and Van Ryzin, G. (1994). Optimal dynamic pricing of inventories with stochastic demand over finite horizons. Management science , 40(8):999--1020

    1994

  19. [27]

    Grossman, S. J. and Stiglitz, J. E. (1976). Information and competitive price systems. The American economic review , 66(2):246--253

    1976

  20. [28]

    Jaillet, P., Podimata, C., Vakhutinsky, A., and Zhou, Z. (2024). When should you offer an upgrade: Online upgrading mechanisms for resource allocation. arXiv preprint arXiv:2402.08804

    2024 arXiv

  21. [29]

    Jasin, S. (2014). Reoptimization and self-adjusting price control for network revenue management. Operations Research , 62(5):1168--1178

    2014

  22. [30]

    and Kumar, S

    Jasin, S. and Kumar, S. (2012). A re-solving heuristic with bounded revenue loss for network revenue management with customer choice. Mathematics of Operations Research , 37(2):313--345

    2012

  23. [31]

    and Kumar, S

    Jasin, S. and Kumar, S. (2013). Analysis of deterministic lp-based booking limit and bid price controls for revenue management. Operations Research , 61(6):1312--1320

    2013

  24. [32]

    and Nazerzadeh, H

    Javanmard, A. and Nazerzadeh, H. (2019). Dynamic pricing in high-dimensions. Journal of Machine Learning Research , 20(9):1--49

    2019

  25. [33]

    Jiang, J., Ma, W., and Zhang, J. (2022). Degeneracy is ok: Logarithmic regret for network revenue management with indiscrete distributions. arXiv preprint arXiv:2210.07996

    2022 arXiv

  26. [34]

    Keskin, N. B. and Zeevi, A. (2014). Dynamic pricing with an unknown demand model: Asymptotically optimal semi-myopic policies. Operations research , 62(5):1142--1167

    2014

  27. [35]

    and Kleinberg, R

    Kumar, R. and Kleinberg, R. (2022). Non-monotonic resource utilization in the bandits with knapsacks problem. Advances in Neural Information Processing Systems , 35:19248--19259

    2022

  28. [36]

    and Szepesv \'a ri, C

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

    2020

  29. [37]

    and Ye, Y

    Li, X. and Ye, Y. (2022). Online linear programming: Dual convergence, new algorithms, and regret bounds. Operations Research , 70(5):2948--2966

    2022

  30. [38]

    Li, Y., Xie, H., Lin, Y., and Lui, J. C. (2021). Unifying offline causal inference and online bandit learning for data driven decision. In Proceedings of the Web Conference 2021 , pages 2291--2303

    2021

  31. [39]

    and Grigas, P

    Liu, H. and Grigas, P. (2022). Online contextual decision-making with a smart predict-then-optimize method. arXiv preprint arXiv:2206.07316

    2022 arXiv

  32. [40]

    Liu, S., Jiang, J., and Li, X. (2022). Non-stationary bandits with knapsacks. Advances in Neural Information Processing Systems , 35:16522--16532

    2022

  33. [41]

    Reiman, M. I. and Wang, Q. (2008). An asymptotically optimal policy for a quantity-based network revenue management problem. Mathematics of Operations Research , 33(2):257--282

    2008

  34. [42]

    Sankararaman, K. A. and Slivkins, A. (2021). Bandits with knapsacks beyond the worst case. Advances in Neural Information Processing Systems , 34:23191--23204

    2021

  35. [43]

    Shah, V., Johari, R., and Blanchet, J. (2019). Semi-parametric dynamic contextual pricing. Advances in Neural Information Processing Systems , 32

    2019

  36. [44]

    Sivakumar, V., Wu, S., and Banerjee, A. (2020). Structured linear contextual bandits: A sharp and geometric smoothed analysis. In International Conference on Machine Learning , pages 9026--9035. PMLR

    2020

  37. [45]

    Sivakumar, V., Zuo, S., and Banerjee, A. (2022). Smoothed adversarial linear contextual bandits with knapsacks. In International Conference on Machine Learning , pages 20253--20277. PMLR

    2022

  38. [46]

    and Banerjee, S

    Vera, A. and Banerjee, S. (2021). The bayesian prophet: A low-regret framework for online decision making. Management Science , 67(3):1368--1391

    2021

  39. [47]

    Wainwright, M. J. (2019). High-dimensional statistics: A non-asymptotic viewpoint , volume 48. Cambridge university press

    2019

  40. [48]

    and Wang, H

    Wang, Y. and Wang, H. (2022). Constant regret resolving heuristics for price-based revenue management. Operations Research , 70(6):3538--3557

    2022

  41. [49]

    and Zheng, Z

    Wang, Y. and Zheng, Z. (2023). Measuring policy performance in online pricing with offline data: Worst-case perspective and bayesian perspective. Journal of Systems Science and Systems Engineering , 32(3):352--371

    2023

  42. [50]

    Wang, Y., Zheng, Z., and Shen, Z.-J. M. (2024). Online pricing with polluted offline data. Available at SSRN 4320324

    2024

  43. [51]

    and Wang, Y.-X

    Xu, J. and Wang, Y.-X. (2021). Logarithmic regret in feature-based dynamic pricing. Advances in Neural Information Processing Systems , 34:13898--13910

    2021

  44. [52]

    and Wang, Y.-X

    Xu, J. and Wang, Y.-X. (2022). Towards agnostic feature-based dynamic pricing: Linear policies vs linear valuation with unknown noise. In International Conference on Artificial Intelligence and Statistics , pages 9643--9662. PMLR

    2022

  45. [53]

    and Cheung, W

    Zhang, X. and Cheung, W. C. (2024). Piecewise-stationary bandits with knapsacks. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.