acceptodds
Under review as a conference paper at ICLR 2027

Optimal Nonparametric Dynamic Pricing with Censored Demand and Adversarial Inventory

Abstract

We study online dynamic pricing with censored demand, where an arbitrary inventory level is revealed before pricing and may adapt to past observations, while demand follows an unknown, price-dependent distribution that is stationary over time. For a horizon of rounds, Xu et al., (2026) achieved regret under restrictive structural assumptions including linear demand, price-independent additive noise, and conditions relating inventory levels to the noise support. Our first contribution is to extend this framework to a substantially more general and statistically harder nonparametric setting, requiring only the natural assumption that expected sales are nonincreasing in price and allowing nonlinear demand curves and price-dependent noise. For this model, we first propose a simple baseline, Double-Grid UCB, which discretizes both price and inventory and achieves expected regret using separate revenue estimates for each price-inventory grid pair. Then, we develop Threshold-UCB, which improves the expected regret to . Unlike Double-Grid UCB, its key idea is to reuse sales observations across inventory levels through shared estimates of demand-tail probabilities, allowing the same data to support revenue upper bounds for multiple inventories rather than a single inventory bin. We also complement this upper bound with an lower bound via a reduction from stochastic posted pricing, establishing its minimax optimality. Finally, extensive experiments across inventory processes, demand functions, and noise models demonstrate consistently superior performance of Threshold-UCB over benchmark algorithms.

Primary area:
learning theory
Keywords:
online learning, dynamic pricing, adversarial inventory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bS84OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies dynamic pricing when inventory is revealed before pricing, may depend on past observations, and censors demand. Under monotone expected sales, Threshold-UCB shares demand-threshold observations across inventory levels and achieves an expected regret bound of order T^(2/3), up to logarithmic factors. A posted-pricing reduction gives a matching lower bound. Experiments cover 48 synthetic environments.

Critical strengths
  1. The broader demand model is useful. The paper removes the linear-additive structure and inventory-support conditions of Xu et al. [X26], while establishing a matching upper and lower regret rate for its own model class. It correctly explains why the two models have different minimax rates.
  2. The threshold estimator has a clear interpretation: an observation is used only when inventory makes the corresponding demand indicator observable. Appendix B explains both the adaptive-sampling argument and why pooling avoids a separate estimation cost for each inventory bin. I also appreciate the separate tuning paths, 30 held-out paths, and variability reporting.

[X26] Xu, J., Wang, Y., Chen, X., and Wang, Y.-X. (2026). Dynamic Pricing with Adversarially-Censored Demands. arXiv:2502.06168, revised January 23, 2026. https://arxiv.org/abs/2502.06168.

Critical weaknesses
  1. The experiments do not use Algorithm 1 exactly. Appendix D.3 omits the zero-threshold term and changes the confidence widths. The omitted population revenue term is pΔS_p (0), which depends on price; this is not simply removing a constant shared by all actions. The modifications are disclosed, but their effect on the comparisons is unclear. A small comparison with Algorithm 1 as stated would help connect the numerical results to the theorem.
  2. Section 4 omits Double-Grid-UCB, so the experiments do not directly measure the benefit of sharing observations across inventory levels. Moreover, Appendix D.2 generates every inventory path before interaction; piecewise changes do not test inventories that react to past observations. These limits do not invalidate the theoretical result, but the empirical conclusions should distinguish what was tested from what the theorem permits.
Questions for authors
  1. On a small representative subset, how do Algorithm 1 as stated, its experimental variant, and Double-Grid-UCB compare under matched grid choices and tuning budgets? I am not asking for a repeat of all 48 environments.
  2. Is any history-dependent inventory experiment available? If not, please make clear that the numerical evidence concerns pre-generated inventory paths; the adaptive-inventory guarantee remains a theoretical result.
  3. A short discussion of two related settings would help. Bu et al. [B23] learn from offline censored data, whereas this paper receives online feedback. Ao et al. [A25] study a shared capacity budget over the horizon, rather than exogenous inventory each period. Please clarify these differences in information and resource constraints; a direct comparison of regret rates is not needed.

[B23] Bu, J., Simchi-Levi, D., and Wang, L. (2023). Offline Pricing and Demand Learning with Censored Data. Management Science, 69(2), 885–903. https://doi.org/10.1287/mnsc.2022.4382. [A25] Ao, R., Jiang, J., and Simchi-Levi, D. (2025). Learning to Price with Resource Constraints: From Full Information to Machine-Learned Prices. Advances in Neural Information Processing Systems, 38, 91243–91281. https://proceedings.neurips.cc/paper_files/paper/2025/hash/83995c2acef585ae0f0d647154cfdd85-Abstract-Conference.html.

2 comments

We thank the reviewer for the time and constructive feedback. Below, we provide detailed responses to the reviewer’s comments. We have also submitted a revised PDF that includes a discussion and theoretical analysis of a new algorithm Shared Double-Grid-UCB (Algorithm 3), more related works, and additional experimental results for both Double-Grid-UCB and Shared Double-Grid-UCB. All changes are highlighted in red in the revised PDF.

Q1: How do Algorithm 1, the experimental implementation, and Double-Grid-UCB compare?

A1: Following your suggestion, we have included the results of the exact Algorithm 1, which keeps the zero threshold and uses the theoretical confidence radius log⁡(2(K+1)JT2)/(2N)\sqrt{\log(2(K+1)JT^2)/(2N)}, the experimental variant in Appendix D.3, which omits the zero threshold and uses the radius cconflog⁡(2(K+1)JT/δ)/(2N)c_{\rm conf}\sqrt{\log(2(K+1)JT/\delta)/(2N)} with δ=0.05\delta=0.05, and Double-Grid-UCB under the same price grid. Specifically, we fixed K=J=22K=J=22 and the confidence multiplier at one in the A, B, and C central-uniform additive settings. Each implementation uses its own stated confidence formula. Entries are mean final expected regret ± one sample standard deviation over the same 30 evaluation paths at T=10000T=10000. The results are shown in the following table. The two Threshold implementations perform similarly, and both have much lower regret than Double-Grid-UCB in all three settings.

EnvironmentAlgorithm 1Experimental variantDouble-Grid-UCB
A-1-U-Add136.64 ± 0.91137.73 ± 1.491187.24 ± 1.22
B-1-U-Add358.66 ± 3.02293.34 ± 3.322911.63 ± 2.72
C-1-U-Add117.94 ± 0.57135.28 ± 0.96929.87 ± 0.73

Q2: Is there an experiment with history-dependent inventory?

A2: All reported experiments use inventory paths generated before the interaction, so they evaluate exogenous inventory sequences. We clarified this in Appendix D.2. Inventory that adapts to past sales is covered by our theory: Theorem 3.2 holds for any inventory sequence whose conditional law depends on the observed past, because the confidence bounds only require that eligibility for each threshold is determined before the current demand is realized (Lemma B.2).

Following your suggestion, we also ran experiments in which each seller’s sales determine its next inventory. We use the paper’s linear Demand A, Yt(p)=1.4−0.8p+εtY_t(p)=1.4-0.8p+\varepsilon_t, with independent shocks εt∼Unif[−0.1,0.1]\varepsilon_t\sim\mathrm{Unif}[-0.1,0.1]. Each seller starts with low inventory γ1=L\gamma_1=L and observes γt\gamma_t before posting a price. If all available inventory sells in round tt, the seller receives HH units in round t+1t+1. Otherwise, it receives LL units. Thus, two algorithms facing the same demand shocks can generate different inventory paths through their pricing decisions.

For each inventory pair (L,H)(L,H), we compared Threshold-UCB, fixed-horizon C20CB, Double-Grid-UCB, and Shared Double-Grid-UCB, a variant of Double-Grid UCB suggested by Reviewer 4Csy which uses each sale to update the revenue estimates for all lower inventory levels. Similar to all the experiments in the paper, we evaluated 20 configurations per algorithm by Bayesian search on five tuning paths and selected the one with the lowest mean final regret. We then evaluated each selected configuration on 30 separate paths of T=10000T=10000 rounds, using common demand shocks across algorithms. Each round’s benchmark is the best expected revenue at that algorithm’s realized inventory. The table reports mean final cumulative expected regret ± one sample standard deviation across the 30 paths.

Inventory (L,H)(L,H)Threshold-UCBC20CBDouble-Grid-UCBShared Double-Grid-UCB
(0.50,1.40)(0.50,1.40)53.87 ± 1.00774.86 ± 297.38214.80 ± 1.21151.31 ± 1.20
(0.65,1.40)(0.65,1.40)32.62 ± 1.00907.15 ± 312.71143.89 ± 0.8689.00 ± 0.90
(0.80,1.40)(0.80,1.40)55.59 ± 2.56531.84 ± 159.05189.69 ± 0.28167.94 ± 0.31
(0.86,1.29)(0.86,1.29)42.20 ± 2.0746.26 ± 44.58201.62 ± 0.45159.02 ± 1.61

The result shows that Threshold-UCB has the lowest mean final cumulative expected regret among the four algorithms in all four inventory settings.

Q3: How does the setting differ from Bu et al. and Ao et al.?

A3: Thank you for pointing out these related works. As the reviewer noted, Bu et al. (2023) learn from an offline data set of historical prices, inventory levels, and censored sales, whereas our learner chooses prices online and the inventory arrives sequentially and possibly adversarially. Ao et al. (2025) study dynamic pricing with a resource budget shared over the whole horizon, whereas in our model inventory is supplied before each period and the comparator optimizes revenue at each realized inventory, so there is no budget to allocate across periods. We have added this discussion to the related work in Section 1.1.

Reviewer_bS84OpenReview ↗

I will keep my score

Thanks for updating the paper.

Reviewer_4CsyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies online nonparametric dynamic pricing under censored demand and adversarial inventory. The demand model allows nonlinear price responses and price-dependent noise, requiring only that expected sales are nonincreasing in price. The authors first introduce Double-Grid-UCB, which separately estimates revenues across price-inventory pairs and achieves O~(T3/4)\widetilde O(T^{3/4}) regret. They then develop Threshold-UCB, which exploits the integral representation of expected revenue to share demand-tail information across inventory levels, achieving a minimax-optimal O~(T2/3)\widetilde O(T^{2/3}) regret bound.

Critical strengths
  1. A broader nonparametric adversarial-inventory pricing setting.

The paper considers a broader nonparametric model than Xu et al. (2026), allowing nonlinear demand, price-dependent noise, and inventory sequences without restrictive support conditions. For this model class, the authors establish matching O~(T2/3)\widetilde O(T^{2/3}) upper and Ω(T2/3)\Omega(T^{2/3}) lower regret bounds, characterizing its minimax rate up to logarithmic factors.

  1. Effective information sharing with minimax-optimal regret.

The authors exploit the integral representation of expected revenue to construct Threshold-UCB, which pools observable demand-tail information across inventory levels. This information-sharing design effectively utilizes censored sales feedback and achieves O~(T2/3)\widetilde O(T^{2/3}) regret, matching the minimax lower bound up to logarithmic factors.

Critical weaknesses
  1. An overly simplistic Double-Grid-UCB baseline.

Double-Grid-UCB is essentially a naive discretization-based UCB algorithm that treats price-inventory pairs separately, updating only the current inventory bin despite the availability of information for lower inventory levels. It offers little methodological novelty and serves mainly as a pedagogical warm-up. Consequently, the highlighted improvement from O~(T3/4)\widetilde O(T^{3/4}) to O~(T2/3)\widetilde O(T^{2/3}) does not, by itself, clarify the incremental benefit of Threshold-UCB over stronger information-sharing baselines.

  1. Limited methodological novelty in the main algorithm and analysis.

Threshold-UCB largely builds on a classical tail-integral representation of expected revenue and the observability of demand-threshold events below the available inventory. The resulting algorithm combines these properties with relatively standard UCB estimation, concentration, and regret-counting arguments. While effective, the overall methodological contribution appears relatively modest.

Questions for authors

Q1. Comparison with a strengthened Double-Grid-UCB baseline.

A natural improvement to Double-Grid-UCB would be to update all eligible lower-inventory revenue estimates using each sales observation, rather than only the current inventory bin. Could the authors investigate the regret guarantees and empirical performance of this strengthened baseline? What additional advantages would Threshold-UCB offer over such an information-sharing variant?

Q2. Dependence across threshold estimates.

The threshold estimators share observations and are statistically dependent, although the current analysis handles them through individual confidence bounds. Could a joint confidence construction exploiting this dependence yield tighter revenue bounds or improved finite-sample performance?

4 comments

We thank the reviewer for the time and constructive feedback. Below, we provide detailed responses to the reviewer’s comments. We have also submitted a revised PDF that includes a discussion and theoretical analysis of a new algorithm Shared Double-Grid-UCB (Algorithm 3), more related works, and additional experimental results for both Double-Grid-UCB and Shared Double-Grid-UCB. All changes are highlighted in red in the revised PDF.

Q1: What does lower-inventory sharing achieve, and what advantage does Threshold-UCB provide?

A1: Following your suggestion, we analyzed and implemented Shared Double-Grid-UCB (Algorithm 3 in Appendix B.3): every inventory representative u≤γtu\le\gamma_t receives the sample ptmin⁡{Xt,u}p_t\min\{X_t,u\}, and price selection uses the current representative's revenue estimate together with the original PDPD confidence radius evaluated at the shared count. We added its pseudocode and analysis to the new Appendix B.3.

Guarantee of Shared Double-Grid-UCB. We argue that this modified algorithm does not improve upon Double-Grid UCB when the inventory is adversarial. Intuitively, since the inventory sequence is adversarial, the extra term in the count, from earlier rounds with the same price and a higher inventory bin, can be zero in every round. Consider a nondecreasing inventory sequence, which our model allows. Then a sale only updates bins at or below the current one, and none of these bins except the current one is used for pricing again. Consequently, the bin used for pricing in each round has received exactly the samples from its own earlier visits, and Shared Double-Grid-UCB posts exactly the same prices as Double-Grid-UCB in every round. Therefore, this kind of information sharing gives no advantage over Double-Grid-UCB. We added this explanation to Section 3.1 and Appendix B.3.

In fact, the worst-case regret for both algorithms remains Ω(T3/4)\Omega(T^{3/4}). Fix any K,J≥1K,J\ge 1. If J≤T1/4J\le T^{1/4}, rounding inventory down to the lower endpoint of its bin can already cost Ω(PDT/J)\Omega(PDT/J) regret at constant inventory. If J>T1/4J>T^{1/4}, consider Θ(T1/4)\Theta(T^{1/4}) successive blocks of increasing inventory, each in a new bin and lasting Θ(T3/4)\Theta(T^{3/4}) rounds. Both algorithms start each block with an empty revenue estimate at every price in the new bin. By choosing demand below every inventory level used in these blocks, we obtain the same posted-pricing problem in every block. The classical posted-pricing lower bound then gives one fixed demand distribution under which the common block policy incurs Ω(PD(T3/4)2/3)=Ω(PDT1/2)\Omega(PD(T^{3/4})^{2/3})=\Omega(PDT^{1/2}) expected regret per block. Summing over the blocks gives Ω(PDT3/4)\Omega(PDT^{3/4}) expected regret for both algorithms.

For completeness, in the revision, we also show in Proposition B.5 that for every K,J≥1K,J\ge1, Shared Double-Grid-UCB achieves

\le \frac{PDT}{K}+\frac{PDT}{J} +6PD\sqrt{(K+1)JT\log(2(K+1)JT^2)}+PD,$$ which is the same $\widetilde{O}(T^{3/4})$ bound as for Double-Grid-UCB (Theorem 3.1) after picking $K$ and $J$ optimally. **Advantage/novelty of Threshold-UCB.** The key novelty of Threshold-UCB lies in its revenue estimator and in how it shares information across decisions to achieve a better guarantee. Unlike Double-Grid-UCB, Threshold-UCB writes the revenue at inventory $\gamma$ as a sum over a threshold grid, $p\Delta\sum_{j:u_{j-1}<\gamma}\Pr(Y(p)>u_{j-1})$, and each threshold estimate pools all past sales at that price whose inventory exceeded the threshold. Hence, when inventory increases, all thresholds below the previous maximum inventory are already well estimated from earlier rounds, and only the newly reached thresholds between the previous maximum and the current inventory start with few samples. In addition, each of these new thresholds has weight only $\Delta=D/J$ in the index, so the new uncertainty is $\mathcal{O}(P\Delta)$ per threshold, whereas Double-Grid-UCB and its shared variant must estimate the full revenue at the new inventory bin from scratch, with uncertainty of order $PD$. Summed over thresholds, the weights $D/J$ cancel the factor $J$, which removes the extra $\sqrt J$ factor and yields $T^{2/3}$ after balancing with the grid errors.

Following the reviewer's suggestion, we also implemented both Double-Grid methods on the original evaluation paths. The revised manuscript includes both baseline algorithms directly in Figure 1 and all eight figures in the appendix, and Appendix B.3 describes and analyzes the Shared Double-Grid-UCB algorithm. The full appendix comparison covers all 48 environments. In the new results, Threshold-UCB has lower mean final regret than both grid baselines in every environment, which is consistent with the theory. The following table shows the mean final expected regret ± one sample standard deviation over 30 paths at T=10000T=10000 in the environments shown in Figure 1.

EnvironmentThreshold-UCB (Figure 1)Double-Grid-UCBShared Double-Grid-UCB
A-1-U-Add76.16 ± 1.77216.49 ± 0.54186.34 ± 0.83
A-2-U-Add57.31 ± 1.18478.67 ± 2.86336.45 ± 2.47
B-1-U-Add37.63 ± 2.11115.39 ± 0.6670.03 ± 0.36
C-1-U-Add68.11 ± 1.04128.71 ± 0.6189.16 ± 0.53
C-3-G-Add61.26 ± 0.75344.85 ± 1.85388.52 ± 2.27
A-1-U-Mult106.93 ± 11.09337.22 ± 7.44328.86 ± 10.22

Q2: Can joint confidence improve the result?

A2: We do not expect a joint confidence construction to improve the regret rate significantly, as the current construction already achieves the minimax T2/3T^{2/3} regret up to logarithmic factors, with the lower bound proven in Theorem C.1. Since the union bound enters only through the factor log⁡(KJT)\log(KJT), a joint construction could at most improve constants and logarithmic factors.

Dear Reviewer 4Csy,

Thank you again for your constructive feedback on our paper. In our response, we analyzed and empirically tested Shared Double-Grid-UCB, the variant you suggested. We explained why updating revenue estimates for lower inventory levels leaves its worst-case regret rate unchanged and how Threshold-UCB uses sales at lower inventory levels to inform pricing at higher ones. In all 48 environments, Threshold-UCB had lower mean final regret than both grid baselines. Feel free to also check our revised revision for details.

We would be grateful if you could acknowledge our rebuttal and let us know whether it addresses your concerns. Any further feedback would also be greatly appreciated.

I thank the authors for their detailed response and additional clarifications. I will keep my score unchanged.

Reviewer_j6kmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies dynamic pricing with censored feedback when inventory is given to the seller before each pricing decision and may depend on past observations. The demand distribution is fixed over time, and expected sales are assumed to decrease with price at every inventory level.

The authors first present Double-Grid-UCB, which learns separately for different price–inventory combinations and achieves O~(T3/4)\tilde O(T^{3/4}) regret. Their main algorithm, Threshold-UCB, uses each sales observation to estimate demand probabilities relevant to several inventory levels. This improves the regret bound to O~(T2/3)\tilde O(T^{2/3}). The paper also shows these rates are optimal and conducts numerical experiments.

Critical strengths
  • The proposed model allows nonlinear demand and price-dependent noise, while inventory can vary arbitrarily and respond to past observations. Both settings are among the most general of all closely related works. The main result shows that the seller can still attain the usual (T^{2/3}) worst-case regret rate for nonparametric pricing.

  • The algorithm combines this information across rounds to estimate expected revenue at different inventory levels. This avoids learning each price–inventory combination separately. The paper also does a good job explaining why this sharing of information improves the regret bound.

Critical weaknesses
  • The setting is largely theoretical, and the paper does not provide a very convincing motivation for why this setting is interesting in general. While the paper is very well written and makes considerable effort to highlight what is different in its algorithm design, almost all components are very similar to existing works in this line of research. No fundamentally new phenomenon is revealed or demonstrated. The (T^{2/3}) regret rate is also very classical and well established in the literature.

  • Numerical experiments are not very thorough. The paper omits Double-Grid-UCB from the experiments, although it is the most direct comparison for this purpose. Many of the remaining comparisons involve C20CB outside its assumptions, while KM-UCB and Naive-UCB are heuristic baselines.

Questions for authors

Please see weaknesses. In general, could the authors explain more clearly why this setting is interesting and challenging, and why existing methods may fail or be insufficient? I would be willing to consider raising my score if these concerns are adequately addressed.

4 comments

We thank the reviewer for the time and constructive feedback. Below, we provide detailed responses to the reviewer’s comments. We have also submitted a revised PDF that includes a discussion and theoretical analysis of a new algorithm Shared Double-Grid-UCB (Algorithm 3), more related works, and additional experimental results for both Double-Grid-UCB and Shared Double-Grid-UCB. All changes are highlighted in red in the revised PDF.

Q1. Why is this setting interesting and challenging? Why are existing methods insufficient here, and what is new in Threshold-UCB?

A1. We extend the model of Xu et al. (2026), which is motivated by many applications where the inventory is given to the seller before pricing and can vary from one round to the next. As mentioned in Xu et al. (2026), examples include touring performances, where the organizer receives the venue's seating capacity before setting the ticket price, and fruit retail, where the retailer sets the price after seeing how much ripe fruit the farm delivers that day. Compared with the classic dynamic pricing problem with unit demand, where a purchase or a rejection reveals the buyer's demand at the posted price, the main challenge of this problem is that a stockout hides how much demand exceeded the available inventory.

Xu et al. (2026) first considered this problem under a set of model assumptions, as discussed in Section 2.1. First, they assume that demand is linear in price with a common additive noise distribution, which allows observations to be transferred across prices. Second, they assume that when the learner posts the highest (or a nearly highest) price, demand is never censored. These assumptions allow their C20CB algorithm to first post high prices and estimate the model parameters from uncensored demand (see Appendix A), so it never needs to handle censored observations. This estimate then gives an unbiased revenue estimate at every price and inventory level, which makes the problem easier. Our model generalizes theirs in three respects, as also discussed in Section 2.1. First, we consider a general nonparametric demand model whose only assumption is that expected sales are nonincreasing in price. This class contains their model and many nonlinear demand curves, and their algorithm is no longer applicable under this extension because it relies on estimating the linear demand parameters. Second, we do not assume that there exists a price at which demand is uncensored. This is more challenging because the seller must learn from censored sales even at the highest price. Third, we do not assume a common additive noise. Instead, we allow the noise distribution, including its variance, to depend on price. Hence, existing algorithms cannot use a common noise distribution to transfer demand observations across prices.

The contribution of Threshold-UCB (Algorithm 1) lies in its revenue estimator and in how it shares information across decisions. At price pp and inventory γ\gamma, the expected revenue is p∫0γPr⁡(Y(p)>u)dup\int_0^\gamma \Pr(Y(p)>u)du, where Y(p)Y(p) denotes demand at price pp. Threshold-UCB discretizes this integral on a grid of thresholds and estimates each survival probability Pr⁡(Y(p)>u)\Pr(Y(p)>u) from every past sale at price pp whose inventory exceeded uu, because such a sale reveals whether demand exceeded uu even when it is censored. We construct an upper confidence bound for each survival probability and sum these bounds up to the current inventory. Every threshold in the selected revenue index receives an observation in that round, and its weight is proportional to the grid spacing. This removes the separate estimation cost for each inventory level. This way of sharing information is not trivial. Specifically, Reviewer 4Csy suggested a variant of Double-Grid-UCB (Algorithm 2) that uses each sale to update the revenue estimates for all lower inventory levels. We show this variant algorithm in Algorithm 3 in the revision. However, this sharing does not improve the guarantee: the variant still achieves only T3/4T^{3/4} regret, because the estimate used for pricing at a larger inventory does not benefit from observations at smaller inventories (see our response to Reviewer 4Csy and the new Appendix B.3). In contrast, Threshold-UCB achieves T2/3T^{2/3} regret up to logarithmic factors. Since the matching lower bound already holds at fixed inventory (Theorem C.1), adversarial inventory and censored sales preserve the minimax regret rate of nonparametric posted pricing.

Q2: Numerical Experiments

Following your suggestion, we added Double-Grid-UCB (Algorithm 2) to Figure 1 and all eight figures in the appendix. Following Reviewer 4Csy's suggestion, we also implemented Shared Double-Grid-UCB (Algorithm 3, Appendix B.3), which uses each sale to update the revenue estimates at the posted price for all lower inventory levels whose sales are observable. Unlike the heuristic KM-UCB and Naive-UCB, both grid baselines have provable guarantees in our model (Theorem 3.1 and Proposition B.5), so this comparison directly measures the benefit of sharing observations across inventory levels. In the appendix of the updated revision, we added the results of these two Double-Grid-UCB algorithms across all 48 environments, and Threshold-UCB has lower mean final regret than both of these two baselines in all 48 environments. The table below gives the six comparisons in Figure 1 on the original 30 held-out paths at T=10000T=10000, reported as mean cumulative expected regret ± one sample standard deviation. The experiment setup details are shown in Appendix D.

EnvironmentThreshold-UCB (Figure 1)Double-Grid-UCBShared Double-Grid-UCB
A-1-U-Add76.16 ± 1.77216.49 ± 0.54186.34 ± 0.83
A-2-U-Add57.31 ± 1.18478.67 ± 2.86336.45 ± 2.47
B-1-U-Add37.63 ± 2.11115.39 ± 0.6670.03 ± 0.36
C-1-U-Add68.11 ± 1.04128.71 ± 0.6189.16 ± 0.53
C-3-G-Add61.26 ± 0.75344.85 ± 1.85388.52 ± 2.27
A-1-U-Mult106.93 ± 11.09337.22 ± 7.44328.86 ± 10.22

Dear Reviewer j6km,

Thank you again for your constructive feedback on our paper. In our response, we explain in more detail what motivates this pricing problem, why learning is challenging compared with the previous work, and the novelty of Threshold-UCB, which lies in how the algorithm uses each sale to estimate revenue across different inventory levels. Is there any aspect of the motivation or the algorithm’s contribution that you would like us to clarify further?

Following your suggestion, we compared Threshold-UCB directly with Double-Grid-UCB across all 48 environments. We also included the information-sharing variant proposed by Reviewer 4Csy in these comparisons. Do the new results address your concern about the experimental baselines?

We would be happy to discuss any remaining concerns or suggestions.

I thank the authors for the response. After reading the responses as well as other review comments and responses, I decided to maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.