A Theoretical Analysis of Optimistic Proximal Policy Optimization in Linear Markov Decision Processes

The proximal policy optimization (PPO) algorithm stands as one of the most prosperous methods in the field of reinforcement learning (RL). Despite its success, the theoretical understanding of PPO remains deficient. Specifically, it is unclear whether PPO or its optimistic variants can effectively solve linear Markov decision processes (MDPs), which are arguably the simplest models in RL with function approximation. To bridge this gap, we propose an optimistic variant of PPO for episodic adversarial linear MDPs with full-information feedback, and establish a $\tilde{\mathcal{O}}(d^{3/4}H^2K^{3/4})$ regret for it. Here $d$ is the ambient dimension of linear MDPs, $H$ is the length of each episode, and $K$ is the number of episodes. Compared with existing policy-based algorithms, we achieve the state-of-the-art regret bound in both stochastic linear MDPs and adversarial linear MDPs with full information. Additionally, our algorithm design features a novel multi-batched updating mechanism and the theoretical analysis utilizes a new covering number argument of value and policy classes, which might be of independent interest.

Paper

References (47)

Scroll for more · 35 remaining

Similar papers

Peer review

Reviewer Eaqc6/10 · confidence 3/52023-06-26

Summary

In this paper, the authors extend the theory of proximal policy optimization-based methods in the linear mixture MDPs and propose an optimistic variant PPO algorithm (OPPO+) for stochastic linear MDPs and adversarial linear MDPs with full information. The proposed algorithm adopts a multi-batched updating rule from bandit literature, where the policy is updated at regular intervals of batches, instead of updating it every episode. During the policy improvement step, the policy is updated using proximal policy optimization, while the policy evaluation step involves estimating the value using least square value iteration with the average reward from the previous batch. The proposed algorithm provides a regret guarantee for $\tilde{O}(d^{3/4} H^2 K^{3/4})$ and has a sample complexity of $\tilde{O}(d^3H^8/\epsilon^4 + d^5H^4/\epsilon^2)$. This regret bound is tighter than the existing PPO-based algorithms proposed for stochastic linear MDPs.

Strengths

- The proposed algorithm can be applied not only to stochastic linear MDPs but also to adversarial linear MDPs with full-information feedback. In order to cover the case of adversarial linear MDPs, the average reward technique is used, and for the analysis, result for bound on the gap between the value of the policy from the previous batch and the value of the policy from the current batch is suggested. - When applying the existing covering number theory from linear MDPs, the regret may depend on the size of the action space. To address this, a method using novel covering theory is proposed to reduce it to a logarithmic scale. - Compared to existing policy optimization algorithms, it provides a tighter regret guarantee in both the stochastic linear MDP and adversarial linear MDP with full-information feedback settings.

Weaknesses

- The computational cost of policy improvement is not discussed in detail. - It remains unclear what advantages the proposed policy optimization algorithms offer compared to value-based algorithms, apart from their theoretical purposes.

Questions

1. What does it mean for the proposed algorithm to be "optimistic"? What I mean is that for example, in value-based algorithms, optimism refers to the estimated values of the algorithm being more optimistic than the true optimal value (Lemma B.5 in [17]). 2. Can you explain the computational cost of policy improvement in OPPO+ compared to the policy improvement in value-based algorithms [17]? 3. Besides the ability to learn stochastic policies, what other advantages does OPPO+ have compared to value-based algorithms? 4. In line 203, the author argued that value-based algorithms cannot handle adversarial rewards. Could you provide more details about this? 5. By performing policy optimization infrequently, the algorithm achieves a regret of order $O(K^{3/4})$. What are the drawbacks or trade-offs associated with this approach? In other words, in line 81, it was mentioned that if policy optimization is performed every episode, the regret would have a linear dependence on $K$. What benefits can be obtained from this approach?

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

4 excellent

Contribution

3 good

Limitations

As this is a theoretical paper, it does not seem to have any negative societal impact. However, the authors have not mentioned the limitations of this algorithm. (If they have mentioned any, please let me know, and I will check.) Nonetheless, they have provided an introduction to future work in Remark 3.3.

Reviewer r8ZK7/10 · confidence 3/52023-07-03

Summary

The paper proposes a theoretical analysis of the PPO algorithm. Some novel techniques such as batch-wise update are proposed so the algorithm can work on the adversarial setting of linear MDPs. A regret bound is given, which is better then or comparable to previous results.

Strengths

The paper is well written, with a clear presentation and is easy to follow. The topic studied in the paper is important. The proposed technique is novel and the theoretical result is solid.

Weaknesses

1. See Question 1. One concern is the similarity between the proposed method and NPG. And additional comparison is needed so the readers could understand the novelty and contribution of the paper more easily. 2. The algorithm design of OPPO+ is similar to OPPO. The difference is OPPO+ replaces the step-wise update with a multi-batched update, and considers the linear MDP setting so the state-action feature $\phi$ can be directly obtained instead of doing integration. The authors claimed the weakness of OPPO, or linear mixture MDP, is that the integration is computationally expensive. But I would rather just regard it as a separate model setting, instead of a weakness. Therefore, when doing comparisons with previous work, besides the literature on linear MDP, it could be great if the authors can also give a more detailed explanation on the setting of linear mixture MDP. In particular, given that the algorithm formulations are so similar, it might be helpful if the author can explain why it's not straightforward to migrate the proof of OPPO to the linear MDP setting. For example, it can be added to the challenge & novelty section, which will greatly strengthen their argument.

Questions

1. The updating rule of PPO has a very similar formula to Natural Policy Gradient (NPG). There also have been literatures that apply NPG to linear MDPs, can the authors make a comparison between their approach and these literatures? For example, in section 4.2 of [1], the authors applied NPG to linear MDPs, and obtained a convergence rate of order $d^2H^6/\varepsilon^3$, which should imply a regret bound of order $d^{1/2}H^{3/2}K^{3/4}$. [1] Liu et al., Optimistic Natural Policy Gradient: a Simple Policy Optimization Algorithm for Online Learning in Linear MDPs, https://arxiv.org/pdf/2305.11032.pdf 2. Is it possible to extend the analysis to the general function approximation setting? 3. One novelty of the paper is its ability to handle adversarial rewards. Can the authors explain which part of their algorithm is crucial to achieving this goal? Still take [1] as an example, I think the algorithms have similar formulation, and the difference of [1] is it doesn't use a multi-batched update. Does that mean the multi-batched updating rule is the crucial part for the adversarial setting?

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

Since the paper is focused on a theoretical side, it's unlikely to have potential negative social impact. And some limitations and future research directions are mentioned in the Conclusion section.

Reviewer c9AM6/10 · confidence 3/52023-07-05

Summary

This paper studies the theoretical performance of an optimistic variant of PPO in episodic adversarial linear MDPs with full-information feedback (i.e., without assuming the reward functions are linear in the feature map), and establishes a regret bound of O(d^3/4 H^2 K^3/4) that matches the optimal regret bound in both stochastic linear MDPs and adversarial linear MDPs. The authors also introduce a new multi-batched updating mechanism to enable a new covering number argument of value and policy classes in their theoretical analysis.

Strengths

Existing theoretical studies of PPO mainly focus on linear mixture MDPs with full-information feedback, which are implemented in a model-based manner and require an integration of the individual base model. In comparison, this work studies another set of linear MDPs that have low-rank representations and proposes a new optimistic variant of PPO that is provably efficient in both stochastic linear MDPs and adversarial linear MDPs. In particular, this work exhibits several promising results: 1. From the algorithmic perspective, the proposed algorithm involves the novel design of a multi-batched updating mechanism and a policy evaluation step via average rewards. 2. Regarding the performance guarantee, the authors establish the optimal regret bound of O(d^3/4 H^2 K^3/4) with two fundamental findings. Instead of using the existing covering argument in linear MDPs, this work presents a new covering number argument for the value and policy classes. In addition, to ensure the sublinear regret, careful analysis has been done to analyze the drift between adjacent policies to control the error arising from the policy evaluation step. 3. Apart from the regret guarantee, it also provides a PAC guarantee as sample complexity, which allows fair comparison with existing works in this line. This work is well-organized and clearly articulates each part with the corresponding motivation, challenges, as well as technical novelties in its solutions. Overall, this paper is technically sound and demonstrates considerable technical novelties. It provides a better understanding of PPO in a class of MDPs with function approximation, which potentially benefits policy optimization in practice.

Weaknesses

While this paper provides insights into the PPO algorithm in linear MDPs, it does have several drawbacks: 1. One of the claimed novelties is the multi-batched updating mechanism, which coincides with the similar idea of "policy switch" in literature. However, there is no discussion of the existing works that involve "policy switch", and thus it is not clear whether the computational efficiency is solely brought by the policy switch scheme. 2. Right now, the regret bound requires the batch size to be O(\sqrt{d^3 K}), with leads to the number of batches being O(\sqrt{K}). But whether this choice of batch size and the number of batches is optimal remains to be unknown. It would be worthwhile to study and discuss the balance between the batch size and the number of batches, the optimal choice, how the balance will affect the regret. 3. Right now, this is solely theory-based work. As PPO is applied widely in practice, it will be beneficial to include simple benchmark empirical studies to demonstrate its effectiveness.

Questions

1. In the introduction, the authors mentioned both Cai et al. and He et al. that study linear mixture MDPs are implemented in a model-based manner, whereas existing theoretical studies in linear MDPs are typically value-based methods. However, from the algorithmic perspective, algorithms in Cai et al. and He et al. also perform regularized least-square approximation on value functions in the policy evaluation step. Do you treat all algorithmics that tries to explicitly learn /approximate the transition model (i.e., \hat{p}) as model-based methods in MDPs with function approximation? Could you explain whether your approach falls into the model-based category? If not, which step makes the difference? Is the multi-batched update the main reason that makes the proposed algorithm more computationally efficient compared to the direct extension of the existing OPPO method to linear MDPs from the algorithmic perspective?

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

This is a theory paper with no potential negative social impact under the discussed context.

Reviewer t7wE7/10 · confidence 3/52023-07-07

Summary

This work resolves the known issue for generalizing the policy-based algorithm proposed in [Cai el al.] for linear mixture MDPs to linear MDPs, by multi-batch updating and a bew covering number argument. The proposed model-free policy optimization algorithm advances the theoretical study of PPO in adversarial linear MDPs with full-information feedback.

Strengths

1 This work resolves the issue in generalizing the policy-based algorithm in [Cai. el al.] for linear mixture MDPs to linear MDPs. 2 Besides stochastic linear MDPs, the proposed algorithm can handle adversarial rewards with full-information feedback.

Weaknesses

1 Full-information feedback instead of bandit feedback is considered. 2 The proposed novel techniques developed in this work are insufficient for the policy-based algorithm to achieve the minimax optimal regret.

Questions

1 in Remark 3.3, the authors claim that they achieve the state-of-the-art regret bound for adversarial linear MDPs with full-information feedback. does it mean their result matches or improves the best existing result in linear MDPs? It would be better to give the reference of the best known exisiting result here. 2 Equipped with the new proposed techniques, is it possible to improve the result by refined analysis? 3 It seems the proposed novel techniques are specialized for policy-based algorithm for linear MDPs. Can the authors comment on the impact of the new techniques? For example, by applying those techniques, there is any existing result that can be improved or any hard problem that now can be solved.

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

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer r8ZK2023-08-14

Thank you for your response. I have raised the score accordingly.

Reviewer Eaqc2023-08-15

Thank you for providing the detailed clarification. It is helpful in understanding the paper. As a result, I have adjusted the scores accordingly.

Reviewer t7wE2023-08-16

Thanks for the detailed response and I am keeping my score.

Reviewer c9AM2023-08-17

Thanks for your detailed reply and the further information to provide a better understanding. I am keeping my score and vote for accept. Good luck.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC