Thompson Sampling For Combinatorial Bandits: Polynomial Regret and Mismatched Sampling Paradox

We consider Thompson Sampling (TS) for linear combinatorial semi-bandits and subgaussian rewards. We propose the first known TS whose finite-time regret does not scale exponentially with the dimension of the problem. We further show the"mismatched sampling paradox": A learner who knows the rewards distributions and samples from the correct posterior distribution can perform exponentially worse than a learner who does not know the rewards and simply samples from a well-chosen Gaussian posterior. The code used to generate the experiments is available at https://github.com/RaymZhang/CTS-Mismatched-Paradox

Paper

References (25)

Scroll for more · 13 remaining

Similar papers

Peer review

Reviewer yJpu7/10 · confidence 5/52024-06-19

Summary

The paper explores CTS for the linear combinatorial semi-bandit problem with subgaussian rewards. It introduces a novel TS algorithm that avoids exponential regret scaling with problem dimensionality. Theoretical bounds and experiments are provided.

Strengths

The paper addresses a significant limitation in existing CTS algorithms by avoiding exponential regret scaling with problem dimensionality, which is a notable contribution to the field. It establishes new regret bounds that improve upon previous results in terms of finite-time performance. This is further shown experimentally. The paper is well-organized and clearly presents its ideas.

Weaknesses

The paper places significant emphasis on the "mismatched sampling paradox," showcasing situations where a divergence between assumed and actual reward distributions can unexpectedly improve performance. While intriguing, this phenomenon isn't entirely novel, as it's known that Thompson Sampling (TS) can adapt to various posterior distributions. The paper's focus on this paradox, including its prominence in the title, may overshadow what could be considered the core contribution: a little boost on the exploration of CTS to better control the regret constant term. The paper does not sufficiently clarify how the exploration boost is integrated into the analysis. The proof strategy is not well synthesized. [minor] Your main term scale with log(m)^2, but it is now possible to get a log(m) term instead. For example, use Lemma 4 in https://arxiv.org/pdf/2302.11182. I think this could be good to at least cite this paper mentioning that this could be done. line 203-204, I think the event D_t should contains the norm infty.

Questions

Can you explain better what makes your analysis work exactly ? What are the proof techniques that you used, and how this differs from [18] ? I put a good score for this paper as I think it worth it. But I can consider putting a lower score if a not satisfying answer is given to the questions.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The analysis only work for linear reward. Can you discuss why this is the case and how would it be possible to extend it the more general cases?

Reviewer 4UPn6/10 · confidence 3/52024-07-03

Summary

This paper addresses Thompson Sampling for stochastic combinatorial bandits with sub-Gaussian rewards. In this area, previous work has identified the interesting phenomenon that some versions of Thompson sampling incur a per-instance regret which is exponential in the (maximum) number of arms pulled per round, m. This is not experienced by UCB-style algorithms, although those tend to have a worse computational complexity. This paper furthers the discussion around the performance of TS in combinatorial bandits by identifying a variant which incurs only polynomial dependence on m. This is achieved by using a distribution with inflated variance to draw Thompson samples, rather than the 'true' posterior. The paper contains non-trivial theoretical work to derive this bound, and contains an experiment demonstrating the superiority of the new TS approach over 'natural' variants and its comparable performance to ESCB, a state-of-the-art UCB-based algorithm.

Strengths

I find the research questions considered in this paper very interesting, and think that the multi-armed bandits community at NeurIPS will value this work. Thompson Sampling is a very popular algorithm for all kinds of bandit problems, and the fact that its most natural extension to combinatorial bandits has poor performance is an important issue. The theoretical results in this paper meaningfully contribute to the understanding of where 'natural' Thompson sampling is necessary and where it is only 'some notion of adding randomness' that is needed. The theoretical work is based on state-of-the-art tools, does include some novel steps and is likely to be of its own interest. It does not appear to be a routine translation of existing tools. The paper is mostly well-written and clear, and for someone acquainted with combinatorial bandits it is quite easy to follow.

Weaknesses

The main weakness of the paper is that more could be done to articulate exactly where the results improve over existing bounds. The exponential and polynomial dependence on m is, as the paper states, in the constant order (wrt T) term. The main comparison between the results of the present paper and [18] is in terms of order results, and some constants are potentially large. It leaves the reader uncertain as to where in the space of T, m and d, the bound of this paper improves upon the bound of [18]. Adding some clarity around this would improve the paper, and help to establish the extent of its contribution. There are some further points where I feel clarification could be made around theoretical and experimental results. I ask questions about these in the section below, and they are also central to me assessing the extent of the contribution. There are some parts where the explanations are probably a bit too brief to be accessible to those without expert knowledge of the field - e.g. lines 52-64 discussing related literature use undefined technical terminology and discuss papers very quickly.

Questions

1. For which values of m and T specifically does the bound of this paper improve upon the bound in [18]? I appreciate this is problem dependent, but if this could at least be answered for the example considered in the experiments, that would be beneficial. 2. The situation in the experiments seems as though it could also be tackled by a 2-armed bandit policy that receives full-bandit feedback on [0,d/2] and essentially ignores the combinatorial structure. If those rewards were scaled to [0,1] would the combinatorial algorithms be outperformed by algorithms for 2-armed bandits? 3. When introducing B-CTS and G-CTS could you make clear if these are the variants to which the results of [18] and [22] apply? 4. The conclusion that using a mismatched variance term outperforms the natural variance is an interesting one, and seems to mirror some findings in the bandit literature that randomisation strategies that are not TS-like or based on a posterior distribution (explicitly) can perform well in bandit problems (e.g. Perturbed History Exploration and bootstrapping based approaches, Kveton et al. (2019) and subsequent work). Can you comment on whether this polynomial dependence relies on using something close to the posterior, or whether other randomisation strategies would seem to be sensible candidates? 5. Can you add to section 5 some clarity on what steps are novel and what are inspired by previous theoretical work? Some of the construction of a clear run seems to bear resemblance to (now) classic proofs for Thompson sampling (Agrawal and Goyal (2012), Kaufmann et al (2012), etc.) and the idea of using sample paths may have roots in other work too? (I'm not certain on that) Agrawal and Goyal (2012) http://proceedings.mlr.press/v23/agrawal12/agrawal12.pdf Kaufmann et al (2012) https://arxiv.org/pdf/1205.4217 Kveton, Szepesvari, et al. (2019) https://arxiv.org/abs/1902.10089

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

yes

Reviewer npKN6/10 · confidence 3/52024-07-07

Summary

The authors present a new Thompson Sampling algorithm for linear combinatorial stochastic semi-bandits. The algorithm provably achieves a better finite-time regret than previous works, and specifically without an exponential dependency on the dimension of the problem. The authors also present a "paradox" that shows using posterior knowledge is not always beneficial.

Strengths

* The paper is presented clearly and the theory is sound. * The experiments suggest the algorithm is useful in practice. * The presented paradox is original and interesting

Weaknesses

* The main result improves the known regret only for the term that does not depend on the time horizon, which is usually less interesting. * Code is not provided for the numerical experiments.

Questions

* In line 26, do you mean $X(t)$ is uniformly distributed? * Can you provide summary of the algorithms as figures? It is hard to follow in the text what is the exact algorithm * Can you provide any intuitive explanation for the presented "paradox"?

Rating

6

Confidence

3

Soundness

4

Presentation

3

Contribution

2

Limitations

The authors address the limitations of the paper.

Reviewer PpKq5/10 · confidence 3/52024-07-10

Summary

This paper proposes a modified version of posterior sampling that achieves optimal asymptotic regret bound, providing an algorithm through the methodology of Thompson Sampling that achieves such a bound.

Strengths

This is a technical paper, and the message is clear and intriguing. This paper validates the methodology of Thompson Sampling in the context of combinatorial bandits in terms of achieving polynomial dependence on the number of dimensions. The technique is novel, to my knowledge.

Weaknesses

I am exactly doing this line of research, so I am not able to follow all the proof details. The paper is notation-heavy; for instance, there are nine events defined, eight of which are some sort of deviations. It is hard to parse and keep track of for a person outside the exact line of research. I appreciate the authors' effort to explain, but I do believe in the main paper, the authors should write a more sketchy proof to highlight the key technique to prove novelty than writing a semi-rigorous proof: for instance, out of the eight deviation events, which are standard and bounded by conventional results? which are novel contributions? Of all the techniques involved, which is the key to this new result? In all, I feel it would be of great help if a clear logic line for the proof is demonstrated for people outside this line of research. At the current state, it is rather hard for me to evaluate the technical contribution.

Questions

What is the guarantee for G-CTS?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

adequately addressed

Reviewer npKN2024-08-09

Thank you for the response, I will keep my score

Reviewer 4UPn2024-08-10

Response to Rebuttal

Thanks very much for your response to my comments, all of these responses are convincing, I would just like to know if and how/where you intend to incorporate these responses into the updated paper?

Authorsrebuttal2024-08-12

Yes, we will incorporate those responses into our paper. We will emphasise the heuristic of the proof and especially the comparison with [18], Agrawal and Goyal (2012), Kaufmann et al (2012). We will also try to incorporate the exponential vs the polynomial scaling if the space constraint allows us.

Reviewer 4UPn2024-08-13

Thank you, I will in that case raise my score, I'd encourage that even if the space prohibits including in the main text some remark should be included in an appendix, or some reorganisation done to allow this inclusion.

Reviewer PpKq2024-08-13

Reply to rebuttal

The authors have addressed my concern. I would like to maintain my score.

Reviewer yJpu2024-08-13

I read the authors’ rebuttal and skimmed the comments of other reviewers. I would like to keep my score unchanged.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC