Regret Minimization via Saddle Point Optimization

A long line of works characterizes the sample complexity of regret minimization in sequential decision-making by min-max programs. In the corresponding saddle-point game, the min-player optimizes the sampling distribution against an adversarial max-player that chooses confusing models leading to large regret. The most recent instantiation of this idea is the decision-estimation coefficient (DEC), which was shown to provide nearly tight lower and upper bounds on the worst-case expected regret in structured bandits and reinforcement learning. By re-parametrizing the offset DEC with the confidence radius and solving the corresponding min-max program, we derive an anytime variant of the Estimation-To-Decisions (E2D) algorithm. Importantly, the algorithm optimizes the exploration-exploitation trade-off online instead of via the analysis. Our formulation leads to a practical algorithm for finite model classes and linear feedback models. We further point out connections to the information ratio, decoupling coefficient and PAC-DEC, and numerically evaluate the performance of E2D on simple examples.

Paper

Similar papers

Peer review

Reviewer frUC4/10 · confidence 5/52023-07-04

Summary

This paper focuses on regret minimization in sequential decision-making under uncertainty. It introduces the average-constrained decision-estimation coefficient, a saddle-point objective that characterizes the worst-case regret, enabling optimization of the information trade-off directly by the algorithm. Moreover, the paper presents a version of the Estimation-To-Decisions (E2D) algorithm (ANYTIME-E2D), with practical implementation details, improved bounds for high-dimensional linear bandits, and the first empirical results of the E2D algorithm.

Strengths

The paper offers several results. The regret results are new since they are based on (3).

Weaknesses

It is unclear how innovative the paper is. The brunt of the mathematics in the paper is drawn from existing prior results. Many mathematical derivations rely on standard Lagrangian (equations (3)-(7) and the proof of Lemma 1).

Questions

What is the importance and impact of the new algorithm? Why is it better, more effective?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

NA

Reviewer zR5j7/10 · confidence 3/52023-07-05

Summary

The authors consider a framework to solve the bandit problem by means of the minimax problem, which has been rapidly developed in recent years. Among them, the paper focuses on the decision-estimation coefficient (DEC). The DEC was developed in a series of studies by Foster+ and is known to characterize the upper and lower bound of the worst-case regret in a general class of bandit problems. The Estimation-to-Decisions (E2D) framework, an algorithm using DEC, determines action selection probabilities based on DEC. However, the algorithm based on offset DEC, which is an existing DEC, is expected to have conservative performance because it needs to determine the tradeoff parameter based on the regret upper bound. Furthermore, the recently developed constrained DEC does not require a tradeoff parameter, but has the problem that the strongly duality of a Lagrangian saddle point fails. To solve these problems, the authors propose average-constrained DEC, a formulation in which the feasible region of the max player in the constrained DEC formulation is relaxed. This eliminates the need to determine the tradeoff parameter based on the regret upper bound while maintaining the good properties of the optimization problem, resulting in an anytime algorithm. Furthermore, by considering a variant of average-constrained DEC, the authors show that an improved rate can be obtained for high-dimensional linear bandits, and this is confirmed through actual numerical experiments.

Strengths

- Overall, the paper is written in a very clear manner. In particular, it clearly summarizes the characteristics of the DECs proposed so far. - It clearly points out the problems of existing algorithms and presents simple solutions to solve them. - The authors devise a specific procedure for obtaining a regret upper bound not only when the model is discrete, but also when linear bandits ( possibly with side observations) are used. - While no numerical experimental evaluation of the DEC has been done so far, the effectiveness of the DEC is demonstrated through numerical experiments in the high dimensional linear bandits.

Weaknesses

- Abstract and introduction are slightly misleading, as it appears that the average-constrained DEC results can be used directly to obtain results for high-dimensional linear bandits. - Although partial monitoring is mentioned in l120-125, the authors are actually solving the problem in a setting where rewards are also observed (or can be assumed to be included), as described in Example 1.2. The authors do not explain to what extent there is a gap with partial monitoring, and it is not clear to what extent the proposed algorithm is applicable to partial monitoring. - As mentioned by the authors, there are other approaches to characterizing sequential decision problems by minimax problems, such as IDS and ExpByOpt. A brief explanation of why the focus is on DEC would be desirable. Minor issues and typos: - l31: $M_f$ should be $M_{f^*}$ - the meaning of "localization" was unclear and it would be better to explain the meaning in short. - In footnote 3, $\max_{f \in co(\mathcal{M})}$ is $\max_{f \in \mathcal{F}}$?

Questions

- While the original DEC is formulated based on the Hellinger distance, this paper uses KL divergence for simplicity. What are the advantages and disadvantages that arise from these? - (e.g.,l169) The meaning of telescope was unclear. Can you explain the meaning of this? - (after l244) For the part of inequality (i), shouldn't the second term of max contains $\mathrm{dec}^{ac}_{\epsilon_t}(\hat{f}_t)$? (The order and the last equation are correct.) - (l226) Why is the assumption "$\max_f \{ \epsilon^2 \mathrm{dec}^{ac}_{\epsilon} (f)\}$ is non-decreasing in $\epsilon$" reasonable? Is there any exception? - Foster+2022 also devised DEC in the adversarial setting. Is the same kind of reparameterization possible in adversarial setting as in the paper? Is there any exception?

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

NA

Reviewer j2sL6/10 · confidence 2/52023-07-07

Summary

This paper studies the estimation-to-decisions framework for sequential decision-making problems with structured observations. They propose the ANYTIME-E2D algorithm, improving precious approaches with a novel bound for linear bandits. Numerical simulations are presented to show the performance of their algorithm against the baselines.

Strengths

- This paper is in general well-written and the overall structure is nicely organized. The proofs seem to be theoretically rigorous. - The significance of the paper is relatively high, since the proposed algorithm can be applied to linear bandit settings and is achieving improved rates in the high-dimensional regime, which is getting increasing attention in the literature.

Weaknesses

- My first concern is the originality and significance of this work, over the existing E2D papers. It is claimed that they introduce the new objective $\operatorname{dec}_\epsilon^{a c}$, which enables the optimization of the information trade-off directly instead of the regret bound. However, no clear evidence is presented to validate the superiority of the new objective. I wonder if there are any special cases where the original objective fails to balance the trade-off, while the new one could. - Secondly, the authors could provide more compelling motivation for the notion of 'anytime' algorithms. As a new reader of the E2D literature, I don't get the advantage of an anytime algorithm over its original version.

Questions

As is discussed in the Weakness part, I hope the authors could provide more explanations and examples about the advantage of the proposed new objective and the anytime variant algorithm.

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

I don't see any limitations or potential negative societal impact of this work.

Reviewer ye2a6/10 · confidence 2/52023-07-07

Summary

This paper focuses on regret minimization in sequential decision-making through min-max optimization. The authors introduce an anytime variant of the estimation-to-decisions algorithm that utilizes the average-constrained decision-estimation coefficient. The proposed algorithm is shown to effectively balance exploration and exploitation. The algorithm achieve a slightly improved regret performance in the context of high-dimensional linear bandits.

Strengths

The paper proposed anytime variant of the estimation-to-decisions algorithm and show the algorithm could potentially improve the linear bandits for a given fixed period.

Weaknesses

1) A major concern is the contribution of the paper compared to the existing literature, particularly the work by Foster et al. in 2021 and 2023. The concept of average constrained DEC appears to be very similar to the (constrained) DEC introduced in Foster et al.'s papers. It is unclear what distinguishes the introduction of the average constrained DEC in this paper and what advantages it brings. 2) The min-max problem in E2D algorithm is inefficient to solve in general, where the paper needs to find the dual variable via grid search and the choice of stepsize could be an issue. 3) While the paper states that the formulation of the average DEC leads to a practical algorithm, it would be more convincing to provide numerical results to support this claim. Comparisons with Foster et al.'s work from 2021 and 2023, as well as classical methods like UCB, Thompson sampling, and Information Direct Sampling, would justify the practicality and effectiveness of the proposed algorithm.

Questions

Please see the weakness.

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

No

Reviewer zR5j2023-08-15

Thank you for your response. I have checked the replies to the questions and they are all reasonable. > Example 1.2 does not require the reward to be directly observed: It is sufficient that the reward vector is in the span of the observations. The assumption in Example 2.2 is $\phi_{\pi} \phi_{\pi}^\top \preceq M_{\pi}^\top M_{\pi}$. If this is related to the assumption that the reward vector is a span of observations, why not use it as an assumption in Example 2.2? In the reviewer's view, it would be easier to understand for readers as it matches the definition of observability used in partial monitoring.

Authorsrebuttal2023-08-17

We do agree that introducing observability conditions would be ideal. For the current work, we focused on a simple condition that allows us to related the DEC to the decoupling coefficient and the PAC-DEC. There are trivial ways to relax the condition, e.g. by bounding the ratio$\\|\phi_\pi \phi_\pi^\top\\|/\\|M_\pi^\top M_{\pi}\\|$ in the spectral norm. It is less clear if existing analysis in finite/linear partial monitoring imply bounds on the DEC. There are subtle technical differences on how the information ratio is defined in the frequentist setting, which prevent us from directly applying the existing results, e.g. [1]. Currently, we do not know if this is just a technical obstacle, or some deeper insight is required. [1] Kirschner, Johannes, Tor Lattimore, and Andreas Krause. "Linear Partial Monitoring for Sequential Decision-Making: Algorithms, Regret Bounds and Applications." arXiv preprint arXiv:2302.03683 (2023).

Reviewer zR5j2023-08-18

I thank the authors for their responses. I have checked the answers and they seem to be valid. The score will remain the same as present.

Reviewer ye2a2023-08-16

I want to thank the authors' response. As the paper argued that anytime E2D is the main contribution of "the anytime upper bound deviates from the proof by Foster et al (2021) in a non-obvious way." Can the authors discuss in detail the technical challenges with "replacing T with t" and how this paper addresses them? As acknowledged in the response, it seems we can just use the "anytime argument" in Foster et al (2021) to establish similar results.

Authorsrebuttal2023-08-17

> As acknowledged in the response, it seems we can just use the "anytime argument" in Foster et al (2021) to establish similar results. Foster et al (2021) do **not** present an anytime analysis; nor do we say so in our response. What is true, however, is that for any **fixed** horizon, both algorithms recover the same results. The point is, however, to achieve optimal regret uniformly over time, which matters, as we show and perhaps surprisingly, even in (high-dimensional/kernel) linear bandits. > Can the authors discuss in detail the technical challenges with "replacing T with t" and how this paper addresses them? The main insight is to use the $\epsilon$-parameterized average-constrained DEC in the analysis. The technical steps are as follows: 1) re-parameterize the saddle point problem in terms of $\epsilon$, 2) Use Sion's theorem to swap the inner min/max 3) modifying the regret upper bound to include the new average-constrained quantity (which perhaps is a more natural quantity to analyze than the offset DEC; while being computationally more tractable than the hard-constrained DEC from Foster et al (2023)). 4) Use Lemma 1 to bound $\lambda^*$ in terms of the average-constrained DEC. This result is not directly implied by the existing work. Second, there are several more contributions in the paper (efficient computation for linear bandits, new bounds on estimation error, experiments, improved bounds for high-dimensional linear bandits) as we have already detailed above.

Reviewer ye2a2023-08-17

I'd like to thank the authors' response, which addresses my confusion and most of my concerns. I increase my rate.

Reviewer frUC2023-08-20

Thanks for providing the answers to comments. I am staying with my score since I still believe there isn't much innovative in the paper. For example, the entire derivation of Lagrangian is standard.

Reviewer j2sL2023-08-21

I thank the authors for addressing my concerns. I will keep my score because the superiority of the proposed algorithm versus Foster et al (2021, 2023) is only validated in a numerical way.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC