NeurIPS 2020

Optimal Algorithms for Stochastic Multi-Armed Bandits with Heavy Tailed Rewards


Review 1

Summary and Contributions: This paper studies the stochastic multi-armed bandit problem with rewards with bounded moment p in (1,2]. This work 1) provides an lower bound on the Robust UCB algorithm, introduced by Bubeck et. al. and show that the lower bound matches the upper bound in Bubeck et. al. 2) introduces an influence function and defines a robust mean estimator, using this function. 3) introduces an algorithm based on this estimator and prove upper and lower bound on the regret. Regret of new algorithm improves the result of Bubeck et. al. by eliminating log T factor and eliminates the assumption that the upper bound on the p-th moment is known. Algorithm can be implemented efficiently.

Strengths: The paper considers interesting problem and I found techniques, developed in this work, potential useful for other problems.

Weaknesses: Some distributions for perturbation don't give optimal problem-dependent bound.

Correctness: The analysis seems correct to me.

Clarity: The paper is well written, analysis is very detailed.

Relation to Prior Work: Could you compare the role of perturbations in iid and adversarial setting?

Reproducibility: Yes

Additional Feedback:


Review 2

Summary and Contributions: This paper studies multi-armed bandits with heavy-tailed rewards. The paper develops a perturbation-based algorithm and derives both lower and upper regret bounds.

Strengths: Compared to the existing algorithms, the proposed algorithm does not require prior knowledge of the moment bound and still achieves optimal regret bound. Theoretically, the paper removes the $ln(T)$ factor in the regret bound of the existing robust UCB algorithm. Empirically, the proposed algorithm behaves comparable to and even better than the existing algorithms with prior knowledge of the moment bound. Furthermore, since most of the existing work on heavy-tailed bandits with structures (such as linear bandits, Lipschitz bandits) is based on the robust UCB algorithm and require prior knowledge of the moment bound, the perturbation based method proposed in this paper can sever as the first step towards parameter-free algorithms for heavy-tailed bandits with structures and may inspire subsequent work.

Weaknesses: As it is stated in the paper that the perturbation-based analysis extends the framework of [8], it would be better to provide some discussions on the technical challenge and novelty in the extension. ***Post rebuttal*** Thanks for the response. My score remains the same.

Correctness: The technical content of the paper seems to be correct.

Clarity: The paper is well-written and well-organized.

Relation to Prior Work: I would like to suggest adding a citation of "No-Regret Algorithms for Heavy-Tailed Linear Bandits", since it is the first work to study linear bandits with heavy-tailed rewards.

Reproducibility: Yes

Additional Feedback:


Review 3

Summary and Contributions: The paper proposes algorithms for stochastic multi-armed bandits with heavy-tailed rewards that do not rely on prior information about bound on p-th moment. The algorithm is shown to achieve minimax optimal dependency on T but has worse dependency on K for gap-independent regret bound compared to robust UCB. Gap-dependent regret bound is also showed and compared with that of robust UCB. Finally, simulation experiments demonstrate that the proposed algorithm performs better than two benchmark algorithm.

Strengths: The problem is very relevant for the NeurIPS community. The theoretical claims look sounds. The simulation result is encouraging.

Weaknesses: The significance and novelty of the contribution is limited. As authors noted in the related works section, the idea of removing the need on prior information about bound on p-th moment has been explored in previous paper. The theoretical guarantee in this paper is only an incremental improvement over the previous works. Compared with robust UCB, the gap-independent bound is improved by a poly-logarithmic factor on T but also worsened by a poly-logarithmic factor on K. The gap-dependent bound of robust UCB and the gap-dependent bound of the proposed algorithm are not directly comparable, and one could be better than the other depending on T,p and \Delta_a. By the way, I think there is a typo for the gap-dependent bound of robust UCB on line 259, where it should be ln(T) instead of ln(T)^{p/(p-1)}. Otherwise, the bound of the proposed algorithm is always smaller than robust UCB.

Correctness: The theoretical claims look correct, even though I didn’t check the proof in the appendix. The simulation experiment results make sense.

Clarity: There is a typo on line 259 which confuses me for a couple of minutes. Other than this, the paper seem well-written.

Relation to Prior Work: Relation to prior work is clearly discussed.

Reproducibility: Yes

Additional Feedback: The figures in the main paper is very small and hard to read. Some experiments using real data instead of synthetic data would be more convincing. ------------------ After author response, I realized that the approach to remove the need on prior information of the bound on the p-th moment is different from previous ideas. Hence, I decided to raise my score.


Review 4

Summary and Contributions: This paper addresses stochastic multiarmed bandit problems with heavy tailed noise distributions. The authors provide a method that only requires knowledge of the degree of the largest finite moment, and not the bound on that moment. This algorithm is based on perturbing a robust mean estimate with an equally heavy tailed distribution. They provide a novel mean estimator and then use it in a perturbed exploration MAB algorithm. They provide several theorems, such as lower bound on the performance of prior algorithms, a concentration bound on their mean estimator, and a regret analysis of their MAB algorithm. ---------------------------- I have read the rebuttal, and appreciate the author's candor with the novelty of this work. My score is unchanged.

Strengths: There are two components that are novel and interesting to this paper. The proposed robust mean estimator, which is a generalization of the Cesa estimator, and the perturbation based MAB algorithm. The new estimator is required because other robust estimators will require knowledge of the moment bound, while this does not. The perturbation technique is also difficult to analyze because you need to show that it provides sufficient exploration.

Weaknesses: It is not clear if being agnostic to the moment bound is a large gain, while leaving the dependence on the moment degree. These are both presumably quantities that need to be calibrated, so what is the main driver for this research from a practical perspective. The motivation seemed to be that this was a line of research that was not completed, so they did it.

Correctness: The proofs seem correct, which I have gone through lightly. For the MAB portion, the error event (not pulling the optimal arm) is decomposed into three terms based on whether the arm is pulled, how large the mean estimator is, and how large the perturbed estimate is. They bound these in the lemmata in turn. Lemma C.2 is particularly involved, and it seems to be where the conditions on the Hazard function is used, with a very clever trick in (C.6) - (C.8) - not sure of it’s originality. The experiments are a nice complement, although they are not extensive, but they do convince me that there is not some hidden massive constant lurking around. One other minor issue is that they claim that (7) is an equivalent assumption, when I think that it only implies the assumption.

Clarity: The paper is well written with only minor typos.

Relation to Prior Work: They clearly motivate the directions that they take in order to achieve their goal of a moment bound agnostic optimal MAB algorithm. Unfortunately, they were somewhat dismissive of work in the more general contexts of contextual and linear bandits. Both can be reduced to the MAB setting, and in particular the work of [11, Shao et al.] can be applied to this setting by letting x_t = e_a. However, looking at that paper, I do not think that the results will pass the criticism of similar MAB works.

Reproducibility: Yes

Additional Feedback: