Multi-Agent Coordination via Multi-Level Communication

The partial observability and stochasticity in multi-agent settings can be mitigated by accessing more information about others via communication. However, the coordination problem still exists since agents cannot communicate actual actions with each other at the same time due to the circular dependencies. In this paper, we propose a novel multi-level communication scheme, Sequential Communication (SeqComm). SeqComm treats agents asynchronously (the upper-level agents make decisions before the lower-level ones) and has two communication phases. In the negotiation phase, agents determine the priority of decision-making by communicating hidden states of observations and comparing the value of intention, obtained by modeling the environment dynamics. In the launching phase, the upper-level agents take the lead in making decisions and then communicate their actions with the lower-level agents. Theoretically, we prove the policies learned by SeqComm are guaranteed to improve monotonically and converge. Empirically, we show that SeqComm outperforms existing methods in various cooperative multi-agent tasks.

Paper

References (43)

Scroll for more · 31 remaining

Similar papers

Peer review

Reviewer gF4i5/10 · confidence 4/52024-06-27

Summary

This paper proposed a multi-level sequential communication framework that has two communication phases. It also proved that the policies learned in this way are guaranteed to improve and converge. Instead of the observations, this paper focus on make agents communicate about the action selections. Each agent is assigned a priority of decision-making first and an equilibrium is set up to be the learning objective for all agents working together.

Strengths

This paper is well-written and well motivated, it borrows ideas from game theory and try to bring the equilibrium into multi-agent decision-making process.

Weaknesses

1. If the authors formulate this problem as multi-agent sequential decision making problem, the Markovian property required for formulating the problem as an MDP is no longer satisfied and there is a need proving an alternative policy gradient theorem before simply inserting the Seqcomm into MAPPO. 2. In Figure 2's caption, the authors stated that 'in the launching phase, the agents who hold the upper-level positions will make decisions prior to the lower-level agents. Besides, their actions will be shared with anyone that has not yet made decisions', then how to change the structure of MAPPO model to make it fit this change? And how to adjust the time step based on this change? 3. There is no code available in this submission, the details of the models' structures could not be seen. The reproducibility remains unsure. 4. Only comparing to TarMAC is not enough, this MARL with communication algorithm was proposed in 2019, there are more recent MARL communication algorithms outperforming TarMAC.

Questions

See the weakness section above.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer LzSK7/10 · confidence 4/52024-07-12

Summary

This paper introduces SeqComm, a novel multi-level communication scheme for multi-agent coordination in reinforcement learning. The key contributions are: 1. A new approach treating agents asynchronously with two communication phases: negotiation and launching. 2. Theoretical analysis proving monotonic improvement and convergence of learned policies. 3. Empirical evaluation on StarCraft multi-agent challenge v2 (SMACv2) showing superior performance over existing methods. The paper addresses the coordination problem in multi-agent settings by allowing agents to communicate actual actions and determine decision-making priorities dynamically.

Strengths

The paper provides a solid theoretical foundation with proofs for policy improvement and convergence (Propositions 1 and 2, Theorem 1). The experimental methodology is sound, using the challenging SMACv2 benchmark and comparing against relevant baselines. The ablation studies support the importance of the proposed components. The paper is generally well-written and organized. The introduction clearly motivates the problem and contributions. The method section provides a detailed explanation of SeqComm. Figures and algorithms help illustrate the approach. The paper presents a novel and significant contribution to multi-agent reinforcement learning. The idea of asynchronous decision-making with dynamic priority determination is original and addresses an important challenge in the field. The theoretical guarantees and strong empirical results on a challenging benchmark demonstrate the value of the approach. Overall the strengths are: - Novel multi-level communication scheme addressing the coordination problem - Theoretical analysis providing performance guarantees - Fairly strong empirical results on SMACv2, outperforming existing methods - Ablation studies demonstrating the importance of key components - Addresses both full and local communication scenarios I'd also like to point out that this was run on a GTX 1050, which makes the model approachable to a general audience.

Weaknesses

- The reliance on homogeneous agents with parameter sharing may limit the applicability of SeqComm in real-world scenarios where agents often have diverse capabilities. However, it is a commonly accepted assumption. - The approach heavily relies on a world model, but there's limited discussion on how the quality of this model affects performance, especially in more complex environments. Also the effect of the attention module is not analysed thoroughly. Attention mechanisms can be computationally expensive, especially as the number of agents increases. The paper doesn't address how this affects the overall scalability of SeqComm. It's not entirely clear how the attention mechanism is adapted for the local communication scenario, where agents only communicate with nearby agents. Furthermore, attention weights could potentially provide insights into which agents or information are most important for decision-making, but this aspect isn't explored in the paper. - Overall, I believe a table reporting IQM values over all environments would strengthen the paper and help with reproducibility. Furthermore, plots showing wall-clock time in comparison to other models would be insightful, especially since final performance only appears to be significantly better in p_10v10 and t_10v10. - While the paper addresses local communication, there's limited discussion on the trade-offs between communication frequency, accuracy, and overall system performance. - Though not standard in related work either, the paper doesn't thoroughly explore the robustness of SeqComm to noisy observations, communication failures, or adversarial agents, which are common challenges in real-world multi-agent systems. - While TarMAC is a strong baseline, the paper could benefit from comparisons with more recent state-of-the-art methods in multi-agent communication.

Questions

1. How does the computational complexity of SeqComm scale with the number of agents, particularly in the negotiation phase? Is there a point where the overhead becomes prohibitive? 2. Have you explored the performance of SeqComm in heterogeneous agent scenarios where parameter sharing may not be applicable? 3. How might the method be adapted for such settings? 4. How sensitive is the approach to the quality of the world model, especially in more complex environments? What happens if the world model is significantly inaccurate? 5. The paper mentions that SeqComm can provide a proper priority of decision-making. How does this prioritization mechanism perform in highly dynamic environments where optimal decision order may change rapidly? 6. Have you investigated the potential emergence of behavioral patterns or strategies across agents due to the asynchronous decision-making process? Are there any interesting emergent behaviors? 7. How does the attention mechanism perform compared to simpler aggregation methods? An ablation study could illuminate this. 8. Does the attention mechanism learn meaningful patterns of inter-agent importance? Visualizing attention weights could provide insights into learned coordination strategies. 9. How does the computational cost of the attention mechanism scale with the number of agents, especially in the full communication scenario? 10. Could more advanced attention mechanisms (e.g., multi-head attention) provide further improvements? 11. How is the attention mechanism modified to handle the local communication scenario effectively?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

- The evaluation is limited to one type of environment (SMACv2). It's unclear how well the method generalizes to other multi-agent scenarios with different dynamics or objectives. - The heavy dependence on a world model could be a significant limitation in environments where accurate modeling is challenging or computationally expensive. - While the paper addresses local communication, there's limited discussion on the trade-offs between communication frequency, accuracy, and overall system performance. - The empirical results are not significantly better in a lot of environments and changing the sight range from SmacV2 makes comparisons to previous work harder.

Reviewer bJxi7/10 · confidence 3/52024-07-12

Summary

This paper introduces SeqComm, a novel multi-level communication scheme for multi-agent reinforcement learning. SeqComm enables agents to coordinate asynchronously, with upper-level agents making decisions before lower-level ones. The approach involves two communication phases: negotiation and launching. In the negotiation phase, agents communicate hidden states to determine decision-making priority, while in the launching phase, upper-level agents lead in decision-making and share actions with lower-level agents.

Strengths

1. SeqComm introduces a novel communication scheme that significantly improves coordination in multi-agent settings. 2. This paper demonstrates theoretical guarantees of monotonic improvement and convergence. 3. This method outperforms existing approaches in various cooperative tasks, showcasing its effectiveness in challenging environments.

Weaknesses

1. The assumptions regarding local observations and communications might not be realistic for all applications.

Questions

None

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

refer to the "weaknesses" part.

Reviewer iQrt7/10 · confidence 3/52024-07-26

Summary

This paper introduces SeqComm, a novel multi-agent reinforcement learning (MARL) method that addresses coordination issues through sequential decision-making and multi-level communication. The main contributions include: 1. A new asynchronous perspective on MARL, allowing agents to make decisions sequentially. 2. A two-phase communication scheme: negotiation phase for determining decision priorities, and launching phase for explicit coordination. 3. Theoretical guarantees of monotonic improvement and convergence for the learned policies. 4. Empirical evaluation on SMACv2 demonstrating superior performance over existing methods.

Strengths

(1) Innovative methodology: The paper proposes a novel approach to MARL by introducing sequential decision-making and multi-level communication, which effectively addresses coordination issues. (2) Theoretical foundation: The authors provide rigorous theoretical analysis, including proofs of monotonic improvement and convergence for the learned policies. (3) Comprehensive empirical evaluation: The method is thoroughly evaluated on multiple maps in SMACv2, demonstrating consistent performance improvements over existing baselines. (4) Ablation studies: The paper includes detailed ablation studies that validate the importance of dynamic decision priority determination. (5) Practical considerations: The authors provide both a full communication version and a local communication version, addressing potential limitations in real-world applications.

Weaknesses

(1) Computational complexity: The paper lacks a detailed analysis of the computational overhead introduced by the multi-level communication scheme, especially for large-scale multi-agent systems. (2) Sensitivity analysis: There is no discussion on the sensitivity of the method to hyperparameters, such as the number of sampling trajectories (F) or the length of predicted future trajectories (H). (3) Potential for deadlocks: The paper lacks a thorough discussion on the possibility of deadlocks in the proposed asynchronous mechanism, which is a critical consideration for any asynchronous system.

Questions

(1) Could you elaborate on your choice of SMACv2 as the primary benchmark for evaluating SeqComm? Are there specific characteristics of SMACv2 that make it particularly suitable for demonstrating the advantages of your method? Additionally, do you believe the results from SMACv2 would generalize well to other MARL environments, and if so, why? (2) Have you considered the possibility of deadlocks in your asynchronous mechanism? What measures, if any, have been implemented to prevent or resolve potential deadlocks? (3) Can you provide an analysis of the computational complexity of SeqComm compared to existing methods, especially for large-scale multi-agent systems? (4) How sensitive is SeqComm to the choice of hyperparameters, particularly F and H? Are there any guidelines for selecting these values?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have made an effort to address some limitations of their work, which is commendable. They acknowledge that the assumption of access to other agents' local observations might not be realistic in all applications. This transparency is appreciated and aligns with the NeurIPS guidelines on discussing limitations. However, there are several areas where the discussion of limitations could be expanded: Scalability: The authors could further discuss how SeqComm's performance and computational requirements might change as the number of agents increases. Generalizability: While SMACv2 is a valuable benchmark, the authors could address how well they expect their results to generalize to other MARL environments. Potential for deadlocks: Given the asynchronous nature of the method, a discussion on the possibility of deadlocks and how they are prevented or resolved would be beneficial. Hyperparameter sensitivity: The authors could elaborate on how sensitive their method is to the choice of hyperparameters, particularly F and H. Regarding societal impact, the authors have not explicitly discussed potential negative consequences. While the immediate applications of this work may seem benign, it would be beneficial to consider and discuss potential misuse scenarios or unintended consequences of more efficient multi-agent coordination.

Authorsrebuttal2024-08-11

This paper aims to demonstrate SE is better than NE. In nature, under SE where agent makes decisions sequentially, parts of agents can obtain the extra information (actions of upper-level agents). Intuitively, if the agent has more valuable information, it will help make the decision. For other potential equilibrium, if it allows the agent to get extra valuable information and does not break the fundamental dynamic of the environment, it also can benefit the performance. Assumption issues. 1. If information asymmetry which means communication is not allowed, then the SeqComm is degraded to MAPPO. If only parts of the information are asymmetrical, it will degrade to the local-communication version. 2. In our view, we will recompute the decision-making order based on the current situation during the execution phase. It will ensure stability. 3. When perfect information about other agents' strategies is not available, we can do the opponent modeling, in more detail, we can train a policy that predicts the actions of other agents. Since we can get the observations of others via communication, the opponent modeling can be precise.

Reviewer iQrt2024-08-11

Thank you for your comprehensive rebuttal. Your explanations have effectively addressed my concerns and substantially improved my assessment of the paper. As a result, I have decided to increase my score to 7. This decision is based on several key strengths of your work: Your SeqComm approach offers an innovative solution to multi-agent coordination through multi-level communication, firmly grounded in Stackelberg Equilibrium theory. The method's robust performance on SMACv2, coupled with a clear analysis of its communication efficiency, demonstrates its practical value in complex MARL scenarios. I am confident that your contribution has the potential to make a significant impact on the MARL field.

Authorsrebuttal2024-08-12

Thank you for your approval, we will revise the text in conjunction with the rebuttal.

Reviewer LzSK2024-08-07

I thank the authors for the extensive rebuttal and running additional results. Your replies certainly improved my outlook on the paper. One more weakness that I'd like your comment on is: > Furthermore, plots showing wall-clock time in comparison to other models would be insightful, especially since final performance only appears to be significantly better in p_10v10 and t_10v10. I do not expect the wall-clock plots for each experiment, given that time is limited. However, I do believe it's an insightful comparison with baselines, as different communication techniques come with different computational costs. In the extreme, your method could take 100x the compute to achieve slightly better results. Alternatively, your method could be 100x faster, which would make the paper that much stronger. I agree that the ultimate metric is final training performance and wall-clock time is secondary, so this would not be a dealbreaker.

Authorsrebuttal2024-08-08

Thank you for your feedback. We are sorry we missed this important question. Since the training speed between protoss, terran, and zerg are very small within the same method, we take the results in protoss maps to illustrate the difference. SeqComm (full comm) fps (frame per second during training): protoss 5v5: 64.32 10v10: 32.75 10v11: 29.82 SeqComm (local comm) fps: protoss 5v5: 108.38 10v10: 64.03 10v11: 59.55 MAPPO fps: protoss 5v5: 166\~167 10v10: 146\~147 10v11: 136\~137 Note that all results are tested on NVIDIA A100. Compared with MAPPO, we only need 2x to 3x the compute to achieve better results.

Reviewer LzSK2024-08-08

Thank you for the quick turn-around. If the authors are willing to add a more extensive wall-clock comparison for the camera-ready version, I'd be happy to improve my score to a 7. A 7 expects a high impact on the community. I think this is justified as the authors provide convincing empirical and theoretical results that their proposed communication scheme is a valid research direction to pursue. The authors also claim to open-source the code upon acceptance, enabling reproducibility. Furthermore, the conceptual simplicity of the method in combination with the improved empirical results is usually a good indicator of impact. I think the 2x-3x wall-clock time is justifiable, especially with increasing computational efficiency of MARL environments. For example, SMAX provides a 40000x efficiency improvement over Smacv2 (33mins vs 44h) when training IPPO, which would make the 2x-3x slow down negligible [1] [1] Rutherford, Alexander, et al. "Jaxmarl: Multi-agent rl environments in jax." arXiv preprint arXiv:2311.10090 (2023).

Authorsrebuttal2024-08-08

Thank you for the comment. We are willing to add a more extensive wall-clock comparison for the camera-ready version. We also guarantee that we will open-source the code upon acceptance, enabling reproducibility.

Reviewer iQrt2024-08-08

Thank you for the authors' comprehensive rebuttal and additional results. The response has indeed improved my view of the paper. However, I still have some concerns regarding the Stackelberg Equilibrium (SE) as the foundational assumption for agent coordination in the proposed SeqComm scheme. Specific Points for Further Consideration: Rationale for SE in MARL: While the authors have addressed the application of SE in SeqComm, I would appreciate a more rigorous justification for choosing SE over other potential equilibrium concepts. It would be beneficial to explore how SE compares to alternative concepts that may be more naturally aligned with the decentralized and dynamic nature of MARL environments. A comparative analysis highlighting the specific advantages of SE in this context would strengthen the paper's theoretical foundation. Assumptions of SE: The paper assumes that agents can achieve and maintain an SE, which may not always be realistic in scenarios where agents have limited or imperfect information about others' strategies. I encourage the authors to discuss the robustness of their approach when these assumptions are violated. Specifically: a) How does the SeqComm scheme perform under conditions of information asymmetry? b) What mechanisms, if any, are in place to ensure the stability of the SE in dynamic MARL environments? c) How does the approach handle scenarios where perfect information about other agents' strategies is not available?

Reviewer gF4i2024-08-09

I thank the authors for addressing my concerns, I like the idea that the code will be released. I suggest that the authors should include the discussions about the new MDP in the main body of the paper.

Authorsrebuttal2024-08-11

Thank you for your suggestion. We will include the discussions about the new MDP in the main body of the paper and release the code. If you find our clarifications satisfactory, we kindly ask for your consideration in adjusting your scores accordingly.

Reviewer gF4i2024-08-14

As stated in the authors' rebuttal, they formed a new MDP, and the formation of this new MDP necessitates the proof of the new alternative policy gradient theorem, which is currently absent in the manuscript. This proof is essential as it forms the foundational support for the proposed theorem. Additionally, the NeurIPS submission guidelines strongly recommend including code at the time of paper submission. The absence of accompanying code not only deviates from these guidelines but also undermines fairness, especially when compared to other submissions that included their codes. Therefore, I believe that my evaluation is justified and fair for this submission.

Authorsrebuttal2024-08-12

Thank you for your approval, we will revise the text in conjunction with the rebuttal.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC