Bypassing the Simulator: Near-Optimal Adversarial Linear Contextual Bandits

We consider the adversarial linear contextual bandit problem, where the loss vectors are selected fully adversarially and the per-round action set (i.e. the context) is drawn from a fixed distribution. Existing methods for this problem either require access to a simulator to generate free i.i.d. contexts, achieve a sub-optimal regret no better than $\widetilde{O}(T^{\frac{5}{6}})$, or are computationally inefficient. We greatly improve these results by achieving a regret of $\widetilde{O}(\sqrt{T})$ without a simulator, while maintaining computational efficiency when the action set in each round is small. In the special case of sleeping bandits with adversarial loss and stochastic arm availability, our result answers affirmatively the open question by Saha et al. [2020] on whether there exists a polynomial-time algorithm with $poly(d)\sqrt{T}$ regret. Our approach naturally handles the case where the loss is linear up to an additive misspecification error, and our regret shows near-optimal dependence on the magnitude of the error.

Paper

Similar papers

Peer review

Reviewer 6s4U7/10 · confidence 3/52023-07-04

Summary

The paper addresses the problem of adversarial linear contextual bandits, where loss vectors are selected adversarially and the context for each round is drawn from a fixed distribution; traditional approaches require access to a simulator for generating free i.i.d. contexts or achieve a sub-optimal regret no better than $\widetilde O(T^{5/6})$. The authors significantly improve on these methods by achieving a regret of $\widetilde O(\sqrt{T})$ without the need for a simulator, while maintaining computational efficiency when the action set per round is small. The results affirmatively answer an open question about the existence of a polynomial-time algorithm with $\mathrm{poly}(d)\sqrt{T}$ regret. The approach allows for the case where the loss is linear up to an additive misspecification error, showing a near-optimal dependence on the magnitude of this error. The paper also presents a computationally inefficient algorithm that provides an improved $\widetilde O(d\sqrt{T})$ regret without a simulator, expanding on the EXP4 algorithm.

Strengths

The paper presents an algorithm that innovatively instantiates an individual Follow-The-Regularized-Leader (FTRL) algorithm on each action set. This development has a wide range of implications for fields. It promises more efficient and effective optimization algorithms in the face of adversarial environments. One of the major strengths of the paper lies in the construction of loss estimators and feature covariance matrix estimators, especially in challenging situations where the learner has no knowledge about the context distribution and there is no simulator available. The paper manages to achieve an improved bound on the bias. The techniques used surpass those used in previous studies [DLWZ23, SKM23], leading to a significant improvement in the algorithm's performance and making it more effective in practice. The introduction of feature centralization is also a notable strength. By centralizing the features by $\hat x_t$, an estimation of the mean features under the current policy, the bias $y_t - \hat y_t$ appears in a nice form that can be compensated by a bonus term. This approach provides an elegant solution to handling bias in the estimator. The paper successfully handles the issue of strong dependence between the policy and the empirical context distribution, Dˆt, which typically rules out canonical concentration bounds. The paper addresses the need for high efficiency in reusing the context samples, which is a challenging problem often not adequately handled by existing methods.

Weaknesses

A notable weakness of the paper is the absence of empirical or numerical experiments to validate the theoretical results. Without practical implementation and testing of the proposed estimators on different datasets, it becomes difficult to assess their efficacy and performance.

Questions

How might these estimators handle high-dimensional data? Are there potential ways to improve their performance in such scenarios in future work? Could the presented estimators be generalized to other types of bandit problems, such as non-linear contextual bandit problems?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations as I could tell. ================================================== Post-Rebuttal: I appreciate the response from the authors, and it addresses most of my concerns. Though I still believe the empirical experiments and the comparisons to other existing methods should be an important part of the work, this is still a strong theoretical paper.

Reviewer f8CU6/10 · confidence 2/52023-07-06

Summary

The paper studies an adversarial linear bandit problem. At each round, the adversary selects a hidden loss vector $y_t$, the environment 'stochastically generates' an action set $A_t$, and the learner chooses an action in the action set. The learner is able to observe noisy loss $\ip{y_t}{a_t}$.

Strengths

Update after rebuttal. This paper proposed an optimal algorithm for the adversarial contextual linear bandit problem. The algorithm consists of OMD with logdet barrier as well as the lifting technique.

Weaknesses

Update after rebuttal. Can dependence on d be improved?

Questions

First, I did not fully understand the problem setting studied in this paper and would need some clarifications. 1. In this work context vectors = actions, I have never seen this formulation in previous literature. The action set $A_t$ is generated from the distribution $D$. What exactly is distribution $D$? Is it defined over the power set of $B_2^d$? What assumption do you make on the size of $A_t$? 2. The formulation does not align with previous work (e.g. NO20). In NO20, they assume the action is chosen by an adversary for each arm in $[K]$ (action for arm $a$ is $\theta_{t,a}$), and the context $x_t$ is drawn from (known) distribution and revealed to the learner. The learner chooses one of the arms and observes the noisy loss centered at ${x_t^T}\cdot{\theta_{t, A_t}}$. Regret is compared to best policy, which is mapping from context to action. This current work assumes the loss vector is chosen adversarially and the action set is drawn 'stochastically' (which I don't completely understand, see point above. ) Regret is compared to best policy, which is mapping from action set to a distribution over action set. Given the different formulations, I do not believe the claims in this work, stating they improve over prior work, are accurate. Based on my understanding of this work, if the action set $A_t$ were constant each round, then this is the bandit linear optimization problem, as in (competing in the dark, abernethy et al. ). Hence this paper really seems to be solving a bandit linear optimization problem where the action set is changing each round, and not the 'adversarial linear contextual bandit' problem as introduced in the intro section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The author did not address limitations.

Reviewer V7BF7/10 · confidence 4/52023-07-07

Summary

Summary: This paper presents a near-optimal computationally efficient simulator-free algorithm in contextual bandits setting with i.i.d contexts and adversarial reward functions. Most of the past work on this setting either requires a simulator that allows them to draw a large number of contexts from a distribution, or is computationally inefficient. Existing algorithms that do not require a simulator and are computationally efficient only achieve a regret of O(T^{5/6}). This work greatly improves the dependence on T, achieving a regret of O(d^2 \sqrt{T}), while still computationally efficient without a simulator. Their algorithm is based on FTRL with a log-determinant barrier. They also present an algorithm that improves the dependence on d, achieving a regret of O(d \sqrt{T}) but is only computationally efficient for a finite policy class.

Strengths

I think this paper makes a solid theoretical contribution to the field. The improvement from O(T^{5/6}) to O(\sqrt{T}) is substantial and simulator-free computationally efficient algorithms are important. The authors exploit the algorithmic ideas of ZL22, the ghost sample trick of NO20, along with novel analysis to provide an improved regret bound. I am slightly concerned about the originality of the algorithmic approach - but this is made up for with the improved analysis and improved regret bound The paper was clearly well written and the contributions clearly stated. I enjoyed how the authors decomposed the regret and the analysis of the estimator in 3.2-3.4.

Weaknesses

The main weakness is that the paper was challenging to read inplaces, especially for readers not already familiar with the ideas of ZL22. There were other places where a bit of extra detail could have helped the exposition greatly. I have tried to remark on these in the list below. Remarks/Comments 1. On bottom of page 5, you mention the loss in the “original space”, can you explain precisely what that loss is? For readers not familiar with ZL22, Line 2 in the algorithm may be opaque - in particular the sequence H_s^{-1} is chosen to control the bias. Moving a bit of the discussion from 3.4/3.5 up - or even simply remarking that it will be explained later may help. 2. The discussion on line 188-189 is similar to 193-194. You can drop one. 3. In equation 2, you use a \hat{Sigma} - however the point of this equations seems to be that Sigma is known! Maybe remove the empirical hat? 4. Is it possible to elaborate a bit more in the paper how your methods “surpass those in [DLWZ23,SKM23]” as described in line 246? 5. \gamma_t should be defined in algorithm for readability

Questions

1. To analyze the first term |\hat{Sigma}_t - H_ty_t|_hat{Sigma}_t^{-1} in line 242 did you need to use any kind of matrix concentration? If so, did you need control of the minimal eigenvalue of H_t? If not, how were you able to avoid it? 2. The linear EXP4 algorithm looks very much similar to the original EXP4 algorithm. Could you provide more explanation of why we cannot directly use other computationally inefficient benchmark algorithms to get the same regret, i.e. why it’s necessary to present and prove linear EXP4? 3. The computationally efficient algorithm, Algorithm 1, achieves a O(d^2\sqrt{T}) regret while having a squared dependence on d; while the algorithm that achieves O(d\sqrt{T}) regret is optimal but computationally inefficient. Do you believe algorithm 1 could achieve a O(d\sqrt(T)) regret?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

N/A

Reviewer He1a8/10 · confidence 4/52023-07-07

Summary

This paper studied the contextual bandits with i.i.d. contexts and adversarial linear loss functions without simulators and proposed a follow-the-regularized-leader with log-determinant barrier (Logdet-FTRL) algorithm with a carefully designed covariance matrix estimator that is computationally efficient given small contexts set and achieves order-optimal (in T) regret \tilde{O}(d^2 \sqrt{T}). Previous algorithms achieving \sqrt{T} regret either require the knolwedge of contexts distribution or a simulator from which the learner can learn the contexts distribution by drawing free i.i.d. contexts. The proposed algorithm estimates the key covariance matrix via previously observes contexts with centralization and ridge regularizer. The setting contains stochastic sleeping bandits as special case and the proposed algorithm answers the open problem in stochastic sleeping bandits affirmtively. A computationally inefficient algorithm that achieves regret with improved d factor is also proposed.

Strengths

The analysis is novel and the contribution is significant to linear adversarial bandits with i.i.d. contexts without simulator. Multiple related scenarios and extentions are discussed. I went over the proofs of the main regret analysis in App. D and several technical lemmas. The proofs are sound to me.

Weaknesses

I do not see perticular weaknesses of the current version.

Questions

Minor typos: Line 164, has been taken The second equality between line 796 and line 797, U^{\Ac_0} instead of \bar{U}^{\Ac_0}.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Yes.

Reviewer V7BF2023-08-15

Thanks for the detailed response and walking me through the minimum eigenvalue concerns. I will leave my score where it is.

Authorsrebuttal2023-08-18

We thank the reviewer for the time and effort spent on reviewing our paper. As the discussion phase is about to end, we want to ensure our responses have adequately addressed your inquiries. We look forward to your feedback.

Reviewer f8CU2023-08-19

Thanks for the response, it addressed my questions. I was initially confused since there are different formulations for 'linear bandit' vs 'linear contextual bandit', but now it is clear. I have adjusted my score.

Authorsrebuttal2023-08-19

We thank the reviewer for the update and the positive evaluation. For the new question in the updated review, we answer it as below. $\textbf{Q3:}$ Can dependence on d be improved? $\textbf{Reply:}$ It is possible that the additional $d$ dependence can be removed, but it requires new techniques. The additional $d$ dependence comes from a union bound over $(1/\epsilon)^{O(d^2)}$ policies when dealing with the covariance matrix estimation error, since the policy used by the algorithm is parameterized by $O(d^2)$ parameters. Due to the strong dependency between the contexts received in previous rounds and the the current policy, we do not believe that such a union bound can be avoided. We suspect that it might be possible to improve the bound to $\tilde{O}(\sqrt{d^3T})$. Recall that the reason we have $(1/\epsilon)^{O(d^2)}$ is because we lift the problem to $(d+1)^2$ dimension and the bonus term is a $(d+1)\times(d+1)$ matrix. However, there are other ways of imposing bonus in *linear bandits* (without contexts) that allows the bonus term to be in $(d+1)$ dimension ([LLWZ20] and [ZL22]), which potentially only requires a union bound over $(1/\epsilon)^{O(d)}$ policies. To achieve the optimal $\tilde{O}(d\sqrt{T})$ bound, we might need a very different algorithm design that does not rely on concentration arguments. [LLWZ20] Chung-Wei Lee, Haipeng Luo, Chen-Yu Wei, Mengxiao Zhang. Bias no more: high-probability data-dependent regret bounds for adversarial bandits and mdps. NeurIPS 2020. [ZL22] Return of the bias: Almost minimax optimal high probability bounds for adversarial linear bandits. COLT 2022.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC