Maximize to Explore: One Objective Function Fusing Estimation, Planning, and Exploration

In online reinforcement learning (online RL), balancing exploration and exploitation is crucial for finding an optimal policy in a sample-efficient way. To achieve this, existing sample-efficient online RL algorithms typically consist of three components: estimation, planning, and exploration. However, in order to cope with general function approximators, most of them involve impractical algorithmic components to incentivize exploration, such as optimization within data-dependent level-sets or complicated sampling procedures. To address this challenge, we propose an easy-to-implement RL framework called \textit{Maximize to Explore} (\texttt{MEX}), which only needs to optimize \emph{unconstrainedly} a single objective that integrates the estimation and planning components while balancing exploration and exploitation automatically. Theoretically, we prove that \texttt{MEX} achieves a sublinear regret with general function approximations for Markov decision processes (MDP) and is further extendable to two-player zero-sum Markov games (MG). Meanwhile, we adapt deep RL baselines to design practical versions of \texttt{MEX}, in both model-free and model-based manners, which can outperform baselines by a stable margin in various MuJoCo environments with sparse rewards. Compared with existing sample-efficient online RL algorithms with general function approximations, \texttt{MEX} achieves similar sample efficiency while enjoying a lower computational cost and is more compatible with modern deep RL methods.

Paper

Similar papers

Peer review

Reviewer J9HB6/10 · confidence 3/52023-06-10

Summary

This paper studied reinforcement learning with general function approximation setting. The authors proposed a new MEX framework, which unifies the exploration and exploitation within a unconstrained optimization objective. Under the structural assumptions about low GEC, they established regret upper bound for learning with their framework. Moreover, they conducted experiments in Mujoco under both model-based and model-free setting and achieved promising results.

Strengths

Computational efficiency is indeed an crucial shortage of existing provably efficient algorithm frameworks. I'm glad to see some effort to investigate how to close those gaps. The idea is clean, the proof is sound to me. The experiments look interesting and promising.

Weaknesses

(1) For the model-free setting, the objective (Eq. 3.1) may not be as easily to optimize as it looks like. By Eq. 3.3, the definition of L_h^k, solving Eq. 3.1 requires to solving a mini-max optimization problem. On the other hand, those constrained optimization objectives in previous literatures (like [1] and [2]) can also be directly converted to a Lagrangian form, which can also be regarded as one objective (a with minimax optimization problem). I don't think it is easy to conclude the objective in this paper is indeed easier to implement than the others. (2) There is not much technique novelty. The proof techniques can be found in previous literatures. The maximization objective is quite straightforward, which just convert the confidence interval constrained optimization objective in previous literatures to a Lagrangian style objective, although there are indeed some difference (here the dual parameter is fixed). [1] Jin et. al., Bellman Eluder Dimension: New Rich Classes of RL Problems, and Sample-Efficient Algorithms [2] Du et. al., Bilinear Classes: A Structural Framework for Provable Generalization in RL

Questions

(1) for the model-free setting, does there exist a general choice of $l$ function in Assumption 3.1? When Assumption 4.2 is satisfied, does $l$ function always exist?

Rating

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

N.A.

Reviewer wpGc7/10 · confidence 2/52023-06-24

Summary

This theory paper proposes an algorithmic framework where the "hypothesis" (model for MB methods, Q function for MF methods) for each iteration is chosen by maximizing one objective: the hypothesis likelihood (i.e., NLL for MB methods, TD error for MF methods) plus the expected returns (value function at the initial state). The paper provides regret guarantees matching prior work. The paper proposes a practical variant of the method (building on top of MBPO and TD3) that shows some promise on sparse versions of benchmark continuous control tasks (while matching performance on the standard, non-sparse versions).

Strengths

**Originality**: To the best of my knowledge the proposed frame work is original. * One prior paper that's related to the model-based version of this paper is https://arxiv.org/abs/2110.02758. This that paper lacks the theoretical contributions of this paper, and comes at the problem from a different perspective, the resulting algorithms look surprisingly similar: both learn a model that is optimized to (1) have high likelihood and (2) produce high-return policies. I'd be curious to see if the current paper also performs well on the stochastic gridworlds used in the prior paper to showcase the benefits of the "optimistic" model. **Clarity**: I generally found the writing clear, which is especially impressive for a large theory project packed into 9 pages. The bottom of this text block has a few suggestions for improving clarity. A few high-level things * The related work section is generally well written, including a pretty thorough review of prior work. One suggestion is to explain the differences from prior work. What are the limitations that this paper will address? * I didn't get much from reading Section 5; I'd prefer that the space be spent on more intuition for the previous section, or more empirical results. **Significance**: Overall, it seems like the proposed framework is a novel and potentially useful way of designing better RL algorithms. I especially appreciate the simplicity of the proposed approach. I think that the paper may be a bit closer to prior work that the paper makes it out to be (see comment under Originality, and the comment about Gumbel), but it still seems to represent somewhat of a departure from the conventional ways of thinking about RL algorithms. Like most new algorithmic frameworks, it seems like there may be a few kinks that have to be worked out empirically (how to choose $\eta$; is there a way of implementing the model-free version without the extra CQL term). Nonetheless, it seems like the paper could inspire more work in this direction. **Strengths** * The paper is generally well written, includes a thorough review of much prior work. * The proposed framework is derived for both model-free and model-based algorithm. * An empirical version of the proposed method is applied to to reasonable continuous control tasks. This is really nice to see in a theory paper.

Weaknesses

* In a few spots, I felt like the paper claims were not entirely substantiated: * "One Objective to Rule Them All" in the title -- Without comparisons to "all" prior methods, I don't think this is a fair claim. Perhaps a more accurate title would be "An Objective Fusing Estimation and Planning for Exploration" * "algorithms predominantly undertake three tasks" -- This seems to be describing methods based on Thompson sampling and posterior sampling, but I'm not sure that R-MAX-style methods fit this mold * "outperforms MBPO by a stable margin ... showcases greater sample efficiency" -- I'm a bit unsure what these statements are referring to. My read of Figure 1 is that there might be statistically significant gains on 2 / 8 tasks (cheetah-vel-sparse, ant-vel-sparse). I would still vote for accepting the paper even if it only outperforms baselines on 2/8 tasks (but I think that accurately portraying the results is very important); this is especially true if we see large benefits on some tabular settings (see suggestion under Originality). * The paper alludes to "data-dependent level-sets" in a number of places, but never defines what these are. I'd recommend either including a few sentences describing similarities/differences, or moving this discussion to the appendix. * I have a few questions about some of the technical details (see below). * From a practical perspective, the model-free version of the method seems to violate conventional wisdom: many practical TD methods (e.g., TD3, MBPO) choose the Q function that has lowest Q-values, whereas the proposed method chooses the one with highest Q values. Not only does this make be a bit concerned that the theoretical framework might not always work well in practice (as evinced by the need for the additional CQL regularizer), but it also makes me wonder exactly how the model-free version of the method is implemented when combined with TD3 (does it take the min or max of the Q functions)? At least for the model-free version, one thing that might help would be to sure that using the CQL regularizer allows the _same_ method to achieve good results in both online and offline settings. --------------------- **Minor comments**: * In the abstract, mention that this is going to be a theory paper. I expected this to be an empirical paper until partway through the intro, when I realized that all of the citations were referring to other theory papers. As such, I found the claim that prior methods "involve impractical algorithmic components" a bit strange, as prior _practical_ methods cannot require impractical components. * L45 -- L47: I found this part a bit unclear because it was unclear what the optimization variable is (a policy, a model, data). This point is clarified later. * L50: I found this part a bit unclear because it was unclear what the hypothesis class was over (policies, models, etc). * Some small grammar errors throughout. An automated grammar checker should catch these (e.g., dropped articles) * L80 -- This definition could be cut, as it's already included a few paragraphs earlier. Cutting this would make it possible to remove some of the whitespace hacks. * L101 -- What is "BE"? * Empirically, in many TD algorithms the TD error _increases_ as the policy improves. The standard explanation is that collecting non-zero rewards results in higher TD errors. This seems potentially related to the analysis in this paper. * L260 -- Including a table for comparing these prior methods might help.

Questions

1. How important is the value of $\eta$ for both the theoretical and empirical results? 2. L147 "Without loss of generality" -- Can the authors elaborate on this point? My guess is that we can ignore stochastic rewards because we could define a new MDP that has a copy of each state for each possible reward that that state could have (and update the transition probabilities accordingly). I'm not sure why it's OK to assume that the reward function is known. 3. There's a sense in which the proposed method resembles posterior sampling methods, by drawing a connection with Gumbel-softmax-style sampling. One way to sample a random variable is to _deterministically_ choose the value that maximizes likelihood + noise, for an appropriately chosen noise distribution (e.g., the Gumbel distribution for categorical distributions). I'm curious if we can interpret posterior sampling methods as a version of Eq. 3.1 where the first term ($V(x)$) is replaced by an appropriate noise distribution (which has nothing to do with the value). 4. L206: Does $V_f^\pi$ make sense for both the model-based and model-free versions of the method? For the model-based version, I'm interpreting this as the expected returns of policy $\pi$ under model $f$. But for the model-free version, I'm unsure if this is supposed to be the expected returns of $\pi$, or the estimate from (value function) $f$. 5. L206, "reduced to vanilla actor-critic": Is this true for the model-based version of the method? 6. Is the second term in Eq. 3.3 required?

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

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

Limitations are not discussed in the main paper. The appendix has a cursory paragraph on limitations. I would recommend including the limitations in the main text, and including a transparent discussion of benefits/limitations of the proposed method relative to prior methods. For example, the first term in Eq. 3.1 seems to sometimes cause optimization problems. Ensemble-based methods might be more stable.

Reviewer vhgd8/10 · confidence 3/52023-07-05

Summary

The paper presents a unified objective for optimizing regret in online RL algorithms with theoretical justification and experimental evaluation. The objective crucially dos not dependent on posterior estimation or bi-level optimization, making it easier to implement in practice than previous theory motivated algorithms. The authors provide a thorough theoretical analysis of their framework, covering both model-based and model-free approaches.

Strengths

The paper is very well crafted, and provides an algorithm that is both theoretically sound and empirically impactful, which is always a great achievement. Even though the presentation is necessarily dense at times, given the theoretical nature of the main contribution, the paper is fairly accessible, with the core intuitions mostly presented well. The results and the practical algorithm presented in the empirical section at the end are surprisingly intuitive and seem straightforward to implement on top of several model-based RL algorithms, not just MBPO. Caveat: I am not necessarily qualified to comment on the correctness of all mathematical details or the impact of the theoretical contributions. I am overall familiar with most concepts used (i.e. Bellman and Eluder rank) but not a core RL theory researcher, so I might be missing some details.

Weaknesses

Given the theoretical nature of the paper, the presentation is necessarily dense, which makes it hard to grasp a couple of intuitive concepts. While the authors do their best to present all main components, here are some suggestions for improvement: - The authors use generic functions several times, most prominently in Section 3, 3.2 and 3.3. They defer concrete realizations to later, but it would be helpful to provide these examples earlier so that readers unfamiliar with the exact approach have an easier time understanding the role that these function play. (Similarly in Section 5, 5.1) - The authors use a unified notation for model-free and model-based hypotheses. While this makes generic statements easier to provide, it led to some confusion on my part because I lost track of whether a model-free or model-based hypothesis was meant at some points. Potentially a specialized notation would clarify these whenever a concrete algorithmic framework is discussed. - While I know that this is not always the main stylistic choice in theory papers, I want to encourage the authors to explain why each assumption is used, i.e. what role they play in proofs. This is especially important for 3.1, which is very abstract and hard to relate to anything this early in the presentation. The role of the exploration policy is not elaborated on in the concrete implementation, etc. This is promised in line 199, but I cannot find a note on the exploration policy in Section 5. The full loss presented in 6.3 does not clarify how the additional term fulfills all the requirements introduced in Section 4. While I was able to piece together most of it, an explicit explanation here would greatly improve the ease with which the role of the assumptions can be understood. In addition, some intuitive description of what this objective optimizes would be nice to have. The experimental section is missing a discussion on how $\eta'$ is set in practice (also, why is $\eta'$ used instead of $\eta$?). 211: for seek of -> for the sake of?

Questions

Assumption 4.3 and Proposition 5: Are there requirements on the exploration policy to guarantee that sufficient coverage is obtained for the supervised learning guarantee? I might be misunderstanding the exact details here, but for the supervised guarantee to hold, it feels like some assumptions on the data is needed? Section 5 (line 279) makes it seem like the choice of $l$ depends on characteristics of the MDP. If so, what aspects of the MDP need to be know a priori to make this choice? Line 298: Is the choice of the Hellinger distance crucial, or would other choices work here as well? 6.2 If a differentiable model of the environment is learned, would it be possible to obtain the gradient wrt to the value function directly, without resorting to a PG style estimator? This would be directly implementable in MBPO, as the model is fully differentiable using reparametrization. The experimental section results miss some standard information: what implementation of MBPO was used (this is also important to obtain hyperparameters), how where the environments chosen (especially why is Humanoid missing), and what is denoted by the shaded area in the plots. Several times "specific MDP structure" is mentioned (lines201, 275, 304 [I assume here it refers to the GEC assumption?]), but the authors do not fully clarify what is meant by this. What is meant by this?

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

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

3 good

Contribution

4 excellent

Limitations

The authors do not address the limitations of their work. Potential discussion points are: the reliance on structural assumptions in the MDP are not provided in-depth, and the tuning of the parameter $\eta$.

Reviewer 44XZ7/10 · confidence 4/52023-07-06

Summary

This paper proposes an RL framework, maximize to explore (MEX), that requires solving an unconstrained optimization which encompasses estimation, planning, and exploration. It proposes both model-free and model-based algorithms and further extends this framework to zero-sum Markov game setting. All these algorithms are proved to be efficient. It also designs practical algorithms based on existing deep RL baselines. Experimental results show that the proposed algorithms outperform baselines in sparse reward environments.

Strengths

This paper proposes a general framework that is adapted to various settings while enjoying favorable theoretical guarantee. Its result generalizes previous ones and is important. It is technically sound, clearly written, and well organized.

Weaknesses

The idea of maximizing to explore is not novel, and previous work are not adequately cited. This line of research is first developed by [Kumar & Becker (1982)](https://ieeexplore.ieee.org/document/1102878) who proposes an estimation criterion that biases maximum likelihood estimation with the cost (or value) of decision. In recent years, several work ([Liu et al., 2020](https://proceedings.mlr.press/v119/liu20g.html); [Mete et al., 2021](https://proceedings.mlr.press/v144/mete21a.html); [Hung et al., 2021](https://ojs.aaai.org/index.php/AAAI/article/view/16961); [Mete et al., 2022](https://ieeexplore.ieee.org/document/9751189/); [Mete et al., 2022](https://proceedings.neurips.cc/paper_files/paper/2022/hash/3c601cd5866099648c6dc783e7f39858-Abstract-Conference.html); [Wu et al., 2022](https://proceedings.neurips.cc/paper_files/paper/2022/hash/5bcb807ae43ad0851a6ba6162a866404-Abstract-Conference.html)) have applied this method to different settings and proved some theoretical results. In particular, [Wu et al. (2022)](https://proceedings.neurips.cc/paper_files/paper/2022/hash/5bcb807ae43ad0851a6ba6162a866404-Abstract-Conference.html) gives a theoretical result similar to the one for model-based MEX in this paper. Rather than completely ignoring previous efforts, it is more appropriate to propose MEX as a further generalization of previous research. The proof of regret bound in this paper relies on the low generalized eluder coefficient assumption. However, it is not discussed whether this assumption hold in practice especially when the hypothesis class is parameterized by neural networks.

Questions

While the theory indicates that the coefficient balancing value and evidence should change over time, it is set to be a constant in practice. Why is it the case?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

This paper has not discussed its limitations, especially about optimization aspects of the learning objective and the practicality of assumptions.

Reviewer 44XZ2023-08-10

Re: Rebuttal by Authors

Thank you for your detailed and comprehensive rebuttal addressing my concerns. Your clarifications have sufficiently addressed my concerns. Taking into account your thorough rebuttal and the improvements you've outlined for your paper, I agree to raise the score from 5 to 7 and look forward to seeing the updated version of your paper.

Authorsrebuttal2023-08-10

Re:Re: Rebuttal by Authors

Dear Reviewer 44XZ, Thank you for your review and support. We will incorporate your valuable suggestions into our paper as we revise it based on the feedback from all reviewers. Your comments greatly assist us in strengthening the overall quality of our work. Best regards, Authors

Reviewer J9HB2023-08-10

Thanks for the response by authors. My concerns are addressed and I would like to keep the score for acceptance.

Authorsrebuttal2023-08-10

Reply by Authors

Dear Reviewer J9HB, Thank you for your review and support. We will incorporate your valuable suggestions into our paper as we revise it based on the feedback from all reviewers. Your comments greatly assist us in strengthening the overall quality of our work. Best regards, Authors

Reviewer wpGc2023-08-10

Reviewer response

Dear authors, Thanks for the detailed response and new experiments. This definitely helps clarify my understanding of the paper. I have one request: please revise the title to be more precise. I title that implies that the proposed method is "ruling over estimation" seems presumptuous and will likely be inaccurate in a few years. Also, please make sure to revise the paper to address the discussion with many reviewers about discussing the differences/similarities with prior work. Best, Reviewer

Authorsrebuttal2023-08-10

Reply by Authors

Dear Reviewer wpGc, Thank you for your review and support. We will incorporate your valuable suggestions into our paper as we revise it based on the feedback from all reviewers. As a part of the revision, we will change the title to a more precise one and add detailed discussions with prior works. Your comments greatly assist us in strengthening the overall quality of our work. Best regards, Authors

Reviewer vhgd2023-08-10

Reply

Thanks for the reply. This covers all my remaining concerns. Thank you for your contribution!

Authorsrebuttal2023-08-10

Reply by Authors

Dear Reviewer vhgd, Thank you for your review and support. We will incorporate your valuable suggestions into our paper as we revise it based on the feedback from all reviewers. Your comments greatly assist us in strengthening the overall quality of our work. Best regards, Authors

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC