DIFFER: Decomposing Individual Reward for Fair Experience Replay in Multi-Agent Reinforcement Learning

Cooperative multi-agent reinforcement learning (MARL) is a challenging task, as agents must learn complex and diverse individual strategies from a shared team reward. However, existing methods struggle to distinguish and exploit important individual experiences, as they lack an effective way to decompose the team reward into individual rewards. To address this challenge, we propose DIFFER, a powerful theoretical framework for decomposing individual rewards to enable fair experience replay in MARL. By enforcing the invariance of network gradients, we establish a partial differential equation whose solution yields the underlying individual reward function. The individual TD-error can then be computed from the solved closed-form individual rewards, indicating the importance of each piece of experience in the learning task and guiding the training process. Our method elegantly achieves an equivalence to the original learning framework when individual experiences are homogeneous, while also adapting to achieve more muscular efficiency and fairness when diversity is observed.Our extensive experiments on popular benchmarks validate the effectiveness of our theory and method, demonstrating significant improvements in learning efficiency and fairness.

Paper

Similar papers

Peer review

Reviewer ddQi5/10 · confidence 5/52023-06-30

Summary

This paper proposes to train a mixing network and agent networks separately for multiagent reinforcement learning when the system is based on the centralized training and decentralized execution (CTDE) paradigm. The authors derive an individual reward that maintains invariance between the gradients of the mixing network's loss function and the sum of the loss functions of the agent networks. Then, the authors apply Prioritized Experience Replay to select samples for training the agent networks. The proposed method, DIFFER, is compared with QMIX, COMIX, QPLEX, and MASER on several tasks, such as SMAC, GRF, and MAMujoco. The experimental results show that DIFFER outperforms the baseline methods and demonstrate that a fair experience replay mechanism works efficiently.

Strengths

- Originality: The idea of fair experience replay based on the constraint between the gradients of the mixing and the agent networks is novel. However, I found the following paper that may be related to this study: D. Hostallero et al. (2020). Inducing Cooperation through Reward Reshaping based on Peer Evaluations in Deep Multi-Agent Reinforcement Learning. In Proc. of AAMAS. - Quality: The experimental results support the claims and the proposed method. - Clarity: The manuscript is written very well and easy to follow. - Significance: While the CTDE approach becomes popular in MARL, many previous studies do not pay attention to the difference between individual experiences. This study sheds light on its importance.

Weaknesses

- Although the experimental results show that the proposed method works well, there is no theoretical analysis. In particular, the gradient of the loss function of the agent network is independent of the target networks as described later. It would be better to discuss the loss function in detail. - In addition, the off-policyness of the proposed method should be discussed.

Questions

Major comments: - The individual reward (2) is interesting, but I am unsure how the TD error is calculated. I think that the TD error is given by $$ (R + \gamma \tilde{Q}_{\mathrm{tot}} - Q_{\mathrm{tot}}) \frac{\partial Q_{\mathrm{tot}}}{\partial Q_i} $$ because the term $\gamma \tilde{Q}_i – Q_i$ cancels out. Is the target network $\tilde{Q}_i$ necessary? Is my understanding correct? - I would like to know whether the used learning algorithm is off-policy or not. Experience replay buffer technique is typically applied to off-policy reinforcement learning algorithms, but the loss functions (6) and (7) suggest that the action value function is trained in an on-policy manner, like SARSA. Is on-policyness necessary for the proposed method? - Is the input of the mixing network consistent with the outputs of the agent networks when the mixing and agent networks are trained separately? - Does QMIX-divide adopt uniform sampling instead PER for training the agent networks? If so, the performance difference between QMIX-DIFFER and QMIX-divide can be discussed based on Fujimoto et al. (2020). S. Fujimoto et al. (2020). An Equivalence between Loss Functions and Non-Uniform Sampling in Experience Replay. NeurIPS 33. Minor comments: - Line 87: $A^n$ -> $A^N$. - Line 96: $r$ -> $R$. - Line 154: Since $S$ denotes the state space, it would be better to use a different symbol. - Eq.(7) and Line 191: $\chi^{\mathrm{indi}}_j$ -> $\chi^{\mathrm{ind}}_j$.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited 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

2 fair

Contribution

3 good

Limitations

Minor comments: The authors do not discuss this work's potential negative social impacts.

Reviewer U3kJ7/10 · confidence 3/52023-07-02

Summary

This is a very interesting and strong paper on MARL that seeks the strategy of learning individual policies rather than team policies in order to break symmetry and learn identity and contextual influences on the individual policies. In order to do so, it solves the technical problem of decomposing a team reward into individual reward, and provides a proof that their method works. They then employ a replay buffer that selects individual experiences to train the agent network akin to prioritized replay. The authors demosntrate their method on a variety of benchmarks, showing improvement.

Strengths

The method is novel; the theory and experiments solid. The writing was very clear and flowed very naturally. The authors demonstrate their method not just multi agent Mujoco, but even the difficult StarCraft II micromanagement tasks.

Weaknesses

I do not have any concerns.

Questions

N/A

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

4 excellent

Limitations

The authors have a limitations section that is fair and does not weaken the strengths of the paper.

Reviewer f7XM6/10 · confidence 4/52023-07-06

Summary

This paper introduces a method for decomposing shared rewards into individual rewards within value-based multi-agent reinforcement learning (MARL) methods. By decomposing the rewards, the method enables the prioritization of crucial individual experiences. Experimental results demonstrate the effectiveness of the proposed approach across a range of value-based MARL methods.

Strengths

1. The paper provides theoretical evidence for the optimization equivalence between value-based MARL methods and those employing decomposed rewards. 2. The concept of decomposing shared rewards into individual rewards is innovative and appears to yield positive results, showcasing the potential for significant impact. 3. The authors have made their code available, ensuring reproducibility of the experiments.

Weaknesses

1. The paper lacks a discussion of policy-based MARL methods, which is a significant gap in the related work. It would be valuable to include a comprehensive overview and comparison with policy-based approaches. 2. The absence of a comparison with policy-based methods diminishes the completeness of the evaluation. Including such a comparison would further strengthen the paper's contribution. Some related papers are [1][2].

Questions

1. Is there any intuitive visualization or illustration available that can help understand the decomposed individual rewards more clearly? 2. How does this method relate to policy-based MARL methods? It would be beneficial to discuss the potential application of individual reward decomposition to policy-based approaches.

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 paper primarily focuses on cooperative MARL tasks. It would be intriguing to explore the applicability of reward decomposition in mixed cooperative-competitive settings.

Reviewer Jgve6/10 · confidence 3/52023-07-21

Summary

The paper presents a new algorithm DIFFER to help individual agents to learn individual TD-error such that it improves the overall (team) model performance. The paper's proposed method is different from an existing method MASER in that it is maintaining the overall team objective. Overall model performance comparison was made against QMIX which uses the traditional team TD-error only.

Strengths

The paper adequately points out the limitation of the proposed method in that additional complexity thus more resources is required. The paper also demonstrates how each complexity (i.e. individual TD-error computation and the selection phase) contributes to the final model performance compared to a prior art. This helps the reader make an informed decision based on the tradeoff between complexity and utility demonstrated. The paper seems overall well-written such that it was fairly easy to follow.

Weaknesses

I would like to see more comparisons of DIFFER's performance and complexity against MASER's, a closely related work.

Questions

See the weakness section 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

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

The authors adequately listed the potential limitation in the paper.

Authorsrebuttal2023-08-17

Be Glad to Tell Us Any Concern

Dear Reviewer: We sincerely appreciate your valuable feedback. We wonder whether our response adequately addresses all of your concern. We are more than willing to provide any additional clarification or answer any questions you may have regarding our work. Best regards, The authors of DIFFER

Authorsrebuttal2023-08-17

Be Glad to Tell Us Any Concern

Dear Reviewer: We appreciate your comments and time! We wonder whether our response have response all of your questions about our work. And we are most glad to discuss anything with you about our work. Best regards, The authors of DIFFER

Authorsrebuttal2023-08-17

Be Glad to Tell Us Any Concern

Dear Reviewer: We genuinely value your suggestions and comments. We address your comments point by point and try our best to respond to them. If there are further suggestions or questions, please let us know them. We are eager to engage in a discussion with you. Best regards, The authors of DIFFER

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC