Improved Communication Efficiency in Federated Natural Policy Gradient via ADMM-based Gradient Updates

Federated reinforcement learning (FedRL) enables agents to collaboratively train a global policy without sharing their individual data. However, high communication overhead remains a critical bottleneck, particularly for natural policy gradient (NPG) methods, which are second-order. To address this issue, we propose the FedNPG-ADMM framework, which leverages the alternating direction method of multipliers (ADMM) to approximate global NPG directions efficiently. We theoretically demonstrate that using ADMM-based gradient updates reduces communication complexity from ${O}({d^{2}})$ to ${O}({d})$ at each iteration, where $d$ is the number of model parameters. Furthermore, we show that achieving an $ε$-error stationary convergence requires ${O}(\frac{1}{(1-γ)^{2}ε})$ iterations for discount factor $γ$, demonstrating that FedNPG-ADMM maintains the same convergence rate as the standard FedNPG. Through evaluation of the proposed algorithms in MuJoCo environments, we demonstrate that FedNPG-ADMM maintains the reward performance of standard FedNPG, and that its convergence rate improves when the number of federated agents increases.

Paper

References (52)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ZDkh6/10 · confidence 4/52023-06-27

Summary

This paper proposes a communication-efficient algorithm, FedNPG-ADMM, for federated natural policy gradient by using a reformulation of quadratic problem. It reduces the communication complexity from $\mathcal{O}(d^2)$ to $\mathcal{O}(d)$. The convergence analysis is provided accordingly.

Strengths

1. The paper is well-written and clearly presented. 2. The reformulation of the original quadratic problem into a distributed manner (eq 14) is a smart move. 3. Using ADMM-based global direction estimation to reduce the communication complexity by $\mathcal{O}(d)$. 4. Stationary convergence in non-convex case is provided, which is the same order as the standard FedNPG.

Weaknesses

As it is in the federated learning (FL) setting, I am curious about main difference between the FL setting considered in this paper and the classic distributed learning. 1. data heterogeneity is one of the key features in FL, but how it impact the model performance does not clearly discussed in the paper, neither in theory or in experiments. 2. To reduce the communication costs, local updates are often used within each agent for FL algorithms. But in the FedNPG-ADMM algorithm proposed in this paper, there is no local update step for each agent. Could local steps be directly applied in the FedNPG-ADMM and further reduce the communication cost while maintaining the same convergence? Given above, to be precise, the setting for this paper is a distributed learning setting, rather than FL with data heterogeneity and local update. The major contribution is the communication reduction as claimed in the paper, so I expect the experiments could have more comparison of direct communication costs. 1. Directly comparison in terms of the communication cost should be given. In other words, the results with communication cost as the x-axis are also desired. 2. Appreciate it if other baselines could be included other than FedNPG. I am curious about the actual communication costs among first-order methods, second-order methods (FedNPG), and approximated version of second-order methods (FedNPG-ADMM).

Questions

see above

Rating

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

The authors have adequately addressed the limitations.

Reviewer pYuC4/10 · confidence 4/52023-07-03

Summary

The paper studies how to train a global policy using distributed data in reinforcement learning. The authors propose a distributed natural policy gradient method by employing ADMM to approximately compute a natural policy gradient direction. The communication complexity is linear in the dimension of policy parameters. The authors also prove the sublinear iteration complexity for achieving a stationary point. The effectiveness of the proposed method is finally verified in MuJoCo experiments.

Strengths

**originality** High communication complexity is a bottleneck for implementing many RL algorithms in federated training. I believe the novelty of this work lies in a new application of ADMM to a popular RL algorithm: natural policy gradient. This method achieves better communication complexity than one for the naive method. I am not aware of this method in the literature. **quality** - The authors provide the stationary-point analysis of the proposed method in a distributed training setting. - The authors verify the use of the proposed method in experiments. **clarity** The proposed method and theory are clearly explained. **significance** - The distributed training algorithm is important for applying RL in distributed systems. - The stationary-point convergence guarantee is provided.

Weaknesses

- The naive method of fererated natural policy gradient only has quadratic communication complexity with respect to the dimension of policy parameter. This does not seem to be a real bottleneck. - Directly extending centralized RL algorithms is not very significant, since fault-tolerance and robustness issues are more important to distributed systems. - It is important to compare communication complexity with the literature, e.g., Fault-Tolerant Federated Reinforcement Learning with Theoretical Guarantee. - Some assumptions can be strong in practice, e.g., the invertibility of Hessian. It is useful to provide examples when Hessian is invertible. - The provided convergence property can be arbitrarily sub-optimal, and it is not clear how to check all assumptions. The stationary-point analysis has a gap with the convergence rate analysis of natural policy gradient in the literature. - The experimental setting is artificial since it is not originally designed for federated learning. There are no comparable baselines in experiments.

Questions

Some questions are raised in Weaknesses. Here are some other questions. - Can the authors provide robustness or fault-tolerance guarantees? - Can ADMM-based update be applied to other policy gradient methods? It is useful to discuss generalizability. - Since the global convergence exists in many policy gradient methods, can the authors strengthen the convergence guarantees? - Can the authors compare the proposed method with other baselines, e.g., Fault-Tolerant Federated Reinforcement Learning with Theoretical Guarantee?

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

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Yes.

Reviewer GLbD7/10 · confidence 2/52023-07-05

Summary

This paper applies the ADMM technique to the Fed-NPG algorithm in reinforcement learning and reduces the communication cost from $O(d^2)$ to $O(d)$ where $d$ is the number of parameters, which nearly maintains the convergence results of Fed-NPG. Empirical results verify the theoretical analysis.

Strengths

The communication cost reduction is impressive since in the federated learning setting, the communication cost is one of the bottlenecks, and the convergence results are nearly the same as Fed-NPG. The idea to combine natural policy gradient with ADMM style methods (or the penalty functions) is interesting (and at least novel to me).

Weaknesses

I am not very familiar with reinforcement learning, and thus may not find potential weaknesses. However, I suggest comparing with the original policy gradient methods in the empirical section and even in the main content (show that Fed-NPG-ADMM converges much fast than the policy gradient variation in federated learning) and plotting all the communication cost for Fed-NPG, Fed-NPG-ADMM, and policy gradients.

Questions

See the weakness section.

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

n/a

Reviewer q8Jc6/10 · confidence 4/52023-07-23

Summary

The work proposed a new algorithm for federated policy optimization. The proposed work uses prima-dual update to replace the primal update of FedNPG, so that the communication cost reduces from $d^2$ to $d$. The proposed work enjoys same rate of convergence as FedNPG under certain assumptions, and numerical experiments were proposed to show the efficacy of the algorithm.

Strengths

The paper is well-rounded and the presentation is clear. The proposed method clearly enjoys the supremacy in terms of communication error. The numerical experiments were proposed to show the efficacy of the algorithm.

Weaknesses

(Please reply to the Questions section directly) The motivation of the proposed algorithm seems a bit less significant. Moreover, the convergence theory requires policy gradient bounded, which seems quite strong. In the numerical experiments, more results could be included to further show the efficiency of the proposed algorithm.

Questions

Here are the my questions and suggestions: 1. Concerns on the motivations: the idea of combining reinforcement learning and federated learning is interesting, which is also largely un-explored in the literature. It is good to see the exploration on this direction. An intriguing question is that can we consider existing works on reformulating the TRPO, such as the KL-penalitzed objective in [1], to federated setting? Since if we get rid of the constraint, we don’t need to communicate the Hessian in the first place (and FedAvg-type algorithm may also enjoys similar rate of convergence in theory). In summary, the authors propose to study the policy optimization in federated setting, but eventually ended up at a less significant improvement over FedNPG (and this improvement comes from essentially the constrained federated learning, not from RL or policy optimization), which makes me feel less confident about the contribution in the RL field; 2. In terms of the assumptions for convergence analysis, the authors proposed bounded policy gradient and Lipschitz continuous (Assumption 4.1) and claimed that they are fairly standard in the existing literature. I’m actually not very sure about the boundedness of the policy gradient. Could the authors just include the specific assumptions in the existing literatures? Same question for Assumption 4.2. 3. The penalty parameter $\rho$ is the penalty to make sure that each local $y_i$ consensus with the global $y$. In the experiment $\rho$ is simply a small number as 0.1 or 0.01. My concern is that during the update, is $y_i - y\rightarrow 0$ true or not? The authors could also plot the consensus errors. 4. In the experiments, for example Figure 2 and 3, FedNPG-ADMM seems not better than FedNPG in terms of the rewards. It would be interesting to see why this is the case. I know that the main claim of FedNPG-ADMM is its commutation efficiency, and I think the authors could also plot the curves wrt the CPU times. References: [1] Schulman, John, et al. "Proximal policy optimization algorithms." arXiv preprint arXiv:1707.06347 (2017).

Rating

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

2 fair

Limitations

The limitation is well stated in weakness and question sections. The authors also include the discussions in limitations at the end of the work. I’m not aware of any potential negative social impact of this work.

Reviewer q8Jc2023-08-13

Thank you for your detailed responses to my comments and questions. I believe that paper is useful contribution to the field and the rebuttal addressed my biggest concern. I appreciate it if the authors could modify the paper accordingly. I'll raise my evaluation to 6.

Authorsrebuttal2023-08-14

Thank you

Thank you for your consideration. We will be sure to modify the paper accordingly.

Reviewer GLbD2023-08-14

Thanks for your response. Since I am not very familiar with the RL background, I will maintain my score and confidence.

Reviewer ZDkh2023-08-14

I would like to thank the authors' rebuttal and I will keep the score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC