Learning General Parameterized Policies for Infinite Horizon Average Reward Constrained MDPs via Primal-Dual Policy Gradient Algorithm

This paper explores the realm of infinite horizon average reward Constrained Markov Decision Processes (CMDPs). To the best of our knowledge, this work is the first to delve into the regret and constraint violation analysis of average reward CMDPs with a general policy parametrization. To address this challenge, we propose a primal dual-based policy gradient algorithm that adeptly manages the constraints while ensuring a low regret guarantee toward achieving a global optimal policy. In particular, our proposed algorithm achieves $\tilde{\mathcal{O}}({T}^{4/5})$ objective regret and $\tilde{\mathcal{O}}({T}^{4/5})$ constraint violation bounds.

Paper

Similar papers

Peer review

Reviewer UyCM7/10 · confidence 3/52024-07-12

Summary

The paper explores RL in Infinite-horizon Ergodic Average-reward constrained MDPs under general policy parametrization. In this setting, the authors propose a primal-dual based policy gradient algorithm that simultaneously optimizes the policy parameter and a constraint violation penalty term $\lambda$. They also equip their method with techniques to address challenges due to the average-reward MDP setup, constraints on the MDP and general policy parametrization. They prove global convergence of their resulting method in terms of the Lagrange function and from this derive bounds on the expected regret as well as constraint violation. Furthermore, when the policy class is expressive enough to approximately contain an optimal policy of the constrained MDP so that the un-expressivity parameter $\varepsilon_{bias}$ is zero or negligibly small, the authors prove that the expected average regret and constraint violation of their method decreases at a rate of $O(T^{-1/5})$.

Strengths

1. The authors consider the task of reinforcement learning in average-reward constrained MDPs and propose novel techniques to address underlying challenges. 2. This is a complete piece of work that explores provably efficient RL in average reward constrained MDPs. Their claims are backed by theoretical analysis, and they highlight the strength as well as weaknesses of their work. 3. This submission is clearly written. The authors highlight the required assumptions and adequately discuss the nature of most relevant parameters. 4. This work significantly contributes to existing theory on reinforcement learning when it is more desirable to optimize the average return, rather than the typical discounted return, and in addition to return optimization, the policy is required to adequately adhere to additional constraints on the MDP.

Weaknesses

See questions.

Questions

1. In Lines 243-246, the authors claim that there are scenarios under which the un-expressivity parameter $\varepsilon_{bias}$ is zero or negligible. I could reason about this claim when $\pi^{\*}$ is an optimal policy in the unconstrained MDP, but unsure that the same holds when $\pi^{*}$ is an optimal policy in the constrained MDP (see Line 240). Can the authors kindly provide exact pointers in the referenced papers that verify this claim, or a proof in the appendix? 2. In Lines 323 and 324, the authors claim that the primal-dual approach to policy optimization is known to give worse results than in the unconstrained version, even for the discounted setting. Could this be related to the fact that the method is policy, rather than value based? In the offline setting, [1] takes a value-based primal-dual approach and their result implicitly highlights that the constrained setting may not be worse off than the unconstrained setting. 3. Can the authors kindly elaborate on the bias of single trajectory-based estimations in this setting? Reference [13] in the paper seems to achieve better performance even with this biased estimation strategy. 4. How strong is Assumption 5? [1] Hong, K., \& Tewari, A. (2024). A Primal-Dual Algorithm for Offline Constrained Reinforcement Learning with Low-Rank MDPs. arXiv preprint arXiv:2402.04493. **Comments** 1. Line 265 makes reference to equation 21 twice rather than the expressions after lines 263 and 264. 2. Expectation is sometimes expressed without the parenthesis. For example, see Equations 20, 22, as well as Lines 265, 266 e.t.c.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

None. This is a purely theory paper.

Reviewer waAx6/10 · confidence 2/52024-07-12

Summary

This paper tackles the infinite horizon average reward Constrained MDP setting and proposes an analysis of regret and constraint violation under a general policy parametrization. They devise a primal-dual policy gradient algorithm achieving global optimality while ensuring sublinear bounds on the regret and constraint violation.

Strengths

The paper is interesting and well-written and places itself well in the related literature by filling the gap about infinite horizon average reward constraint MDPs under general parametrized policies. They employ similar techniques that can be found in the related unconstrained setting but also highlight and overcome the challenges encountered while facing the constrained and average reward setting. In particular, they take inspiration from already existing techniques developed for the discounted setting and adapt them to cope with the average reward MDP.

Weaknesses

From the theoretical perspective, I do not recognize specific weaknesses. However, I would like to highlight that the authors assume the knowledge of the mixing and hitting time of the MDP. I recognize that this is a standard assumption in this type of work, however, I wonder if they are necessary to achieve these results or if the authors are aware of similar techniques that can provide sublinear guarantees without this knowledge. Recently, the work of [1] highlighted some techniques that allow achieving global convergence results without this knowledge. Would it be possible to apply these techniques to this setting as well? From the simulation perspective, I believe that some experimental results showing the validity of the presented approach may be beneficial for the presented work. [1] Patel B. et al., Towards Global Optimality for Practical Average Reward Reinforcement Learning without Mixing Time Oracles, ICML 2024.

Questions

See the weakness section.

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer g5Lw7/10 · confidence 2/52024-07-12

Summary

This paper studies learning in constrained MDPs with the long-run average-reward objective. It gives the first sample complexity, regret bound and constraint violation bound in this setting with general parameterization, whereas all prior work is restricted to either tabular or linear parameterizations.

Strengths

The result and analysis look solid. The presentation is also clear.

Weaknesses

1 The assumption that all policies induce aperiodic irreducible Markov chain is a bit strong. To what extend can you weaken the assumptions such that the current analysis in the paper still go through? Is it enough if we only assume the Markov chain is weakly-communicating? 2 Claiming that the regret is O(T^{4/5}) in table 1 is a bit misleading, because in the regret and constraint violation bounds in Theorem 2, there are actually terms linear in T whose coefficients depend on the transfer error \epsilon_{bias}. I understand that the linear-in-T error terms due to the transfer error are common in the analysis of RL with general parameterizations, and the transfer error is zero when the parameterization is tabular. However, it is still worth discussing whether the coefficients of the linear-in-T terms in Theorem 2 are tight and optimal, how they compare to the prior work, and whether they are truly negligible when the policy is parameterized by a neural net of a reasonable scale. At the very least, how the regret and constraint violation are actually calculated should be clarified in a prominent part of the paper.

Questions

See my questions in the weaknesses section.

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

There is no negative social impact. There are some discussion of limitation in the conclusion section, but not sufficient.

Reviewer nhM46/10 · confidence 4/52024-07-12

Summary

This paper propose a primal dual policy gradient method for solving average reward constrained MDPs. The primal problem is minimizing the usual RL objective with a averaged reward, plus a penalty induced constraint violation term. The dual problem is to find appropriate Lagrangian multiplier that balances the two objectives to reach equilibrium. The proposed method has reached the claimed regret bound and constraint violation bounds with general policy parameterization.

Strengths

The paper is well written and components are clearly explained. The presentation is good and easy to understand. The investigated topic is interesting and important for real life applications, and average reward is harder to study in general w.r.t to its discounted counterparts due to the absence of contraction of the Bellman operator under the infinity norm. The analysis is easy to follow and plainly laid out. The basic flow of thought is easy to follow and nicely connected. The analysis itself is solid in terms of its assumptions made in the paper.

Weaknesses

There are a few points that made this paper limited in its technical contribution. 1. The claimed the general parameterization seems to rely on an accurate policy model such as neural network. In the paper the authors assumes somehow we can obtain an accurate policy for free. In reality, this is far from the truth. In this sense, the claimed contribution is not much of importance. 2. Also related to the first point, it seems that for policy evaluation (algorithm 2), the values functions are still in tabular form. Additionally, this is closely following the proposed policy evaluation method in [17] referred in the paper, which seems to the hardest part in average-reward MDPs. Existing literature shows that for policy optimization alone, it is not so different from discounted setup. See Li, Tianjiao, Feiyang Wu, and Guanghui Lan. "Stochastic first-order methods for average-reward markov decision processes." arXiv preprint arXiv:2205.05800 (2022). 3. There is no numerical experiments to validate the proposed algorithm. There is a growing important for RL researcher to bridge the gap between theories and practice.

Questions

It seems uniform ergodicity is crucial in the paper. I wonder one can relax the assumption for the MDP to be unichain.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer dyHb5/10 · confidence 3/52024-07-13

Summary

The paper presents a primal-dual policy gradient algorithm for infinite horizon average-reward Constrained Markov Decision Processes with general policy parametrization. It uniquely addresses minimizing regret and managing constraints simultaneously, providing the first analysis showing sub-linear bounds of O(T^{4/5}) for both regret and constraint violations. This approach extends the applicability of reinforcement learning to complex, large-scale decision-making problems under practical constraints.

Strengths

- The paper is the first to apply a primal-dual policy gradient framework to this particular setting of CMDPs, addressing a gap in the literature. - The problem is well-motivated. - The paper offers theoretical sub-linear bounds for both regret and constraint violations.

Weaknesses

- The assumption of ergodicity is quite strong and may limit the applicability of the proposed algorithm in practical scenarios where such conditions are not met. This contrasts with the Upper Confidence Reinforcement Learning (UCRL) framework, which operates under the more relaxed condition of weakly communicating MDPs. - The literature review table omits the MDP modeling assumptions. Specifically, the regret bound offered by the proposed algorithm is O(T^{4/5}), while tabular constrained RL algorithms under more general multichain MDP (communicating/weakly communicating) assumptions in [6] achieve O(T^{2/3}). Related facts were not mentioned in the paper. - The mixing time, t_{\text{mix}}, tends to depend unfavorably on the state space S and action space A, reflecting in the overall diameter of the MDPs.

Questions

- Can you explain in more detail the dependence of the mixing time on the state space S and action space A is critical, especially in large-scale problems. Clarification on how t_{\text{mix}} scales with S and A and its impact on the algorithm’s performance would provide a better understanding of the method’s applicability to complex environments. - Error propagation for estimated t_{\text{mix}} : Information on the numerical performance of the algorithm, particularly how t_{\text{mix}} influences runtime and convergence in practical scenarios, would substantiate the theoretical claims. Experimental results or simulations that highlight these dynamics could significantly enhance the manuscript.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors touched upon future work/limitations on theoretical complexity

Reviewer UyCM2024-08-08

I thank the authors for the rebuttal. I am satisfied with their response and have read other reviews and responses.

Reviewer nhM42024-08-11

Thank you for the response

I thank the authors for their response. Although I am not entirely satisfied by the author's rebuttal, the overall quality of this paper is acceptable.

Reviewer g5Lw2024-08-12

I appreciate the authors' informative response to my comments.

Reviewer waAx2024-08-12

I thank the authors for their detailed responses, I have no further questions for the moment.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC