RiskQ: Risk-sensitive Multi-Agent Reinforcement Learning Value Factorization

Multi-agent systems are characterized by environmental uncertainty, varying policies of agents, and partial observability, which result in significant risks. In the context of Multi-Agent Reinforcement Learning (MARL), learning coordinated and decentralized policies that are sensitive to risk is challenging. To formulate the coordination requirements in risk-sensitive MARL, we introduce the Risk-sensitive Individual-Global-Max (RIGM) principle as a generalization of the Individual-Global-Max (IGM) and Distributional IGM (DIGM) principles. This principle requires that the collection of risk-sensitive action selections of each agent should be equivalent to the risk-sensitive action selection of the central policy. Current MARL value factorization methods do not satisfy the RIGM principle for common risk metrics such as the Value at Risk (VaR) metric or distorted risk measurements. Therefore, we propose RiskQ to address this limitation, which models the joint return distribution by modeling quantiles of it as weighted quantile mixtures of per-agent return distribution utilities. RiskQ satisfies the RIGM principle for the VaR and distorted risk metrics. We show that RiskQ can obtain promising performance through extensive experiments. The source code of RiskQ is available in https://github.com/xmu-rl-3dv/RiskQ.

Paper

Similar papers

Peer review

Reviewer EN1R6/10 · confidence 2/52023-06-09

Summary

This paper studies risk-sensitive action selection in cooperative multiagent reinforcement learning (MARL) under the centralized training but decentralized execution (CTDE) paradigm, based on the value decomposition class of methods. The principle underlying value decomposition is Individual Global Max (IGM), with its variant under distributional learning being called Distributional IGM (DIGM). This paper generalizes IGM and DIGM for risk-sensitive behavior, resulting in the Risk-sensitive IGM (RIGM). The sources resulting in risk are environmental uncertainty (stochasticity in state dynamics and/or rewards), non-stationary policies of agents (all agents are learning/evolving), and partial observability (perceptual ambiguity: observations have missing information in comparison to the Markov states). Given that the problem setting is cooperative MARL, the notion of risk is one that minimizes risk on the global or collective action, and not individual consequences per agent.

Strengths

- The contribution is novel (as far as I know) - Problem is well motivated and exposition is done sufficiently well - Experiments are thorough, with a good collection of environments/tasks and key baselines included

Weaknesses

Limitations are not thoroughly discussed (my opinion could change subject to authors' response)

Questions

- Are there any none value decomposition baselines for dealing with risk for CTDE? - Isn't there a simpler way to achieve risk-aversive sub-policies by having a risk-aversive central critic update (as in Dabney et al. (2018b))? Maybe the issue there is not being able to ensure individual policies are also risk aversive? - What are the limitations of this approach and potential future works?

Rating

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

Not discussed in good detail (my opinion could change subject to authors' response)

Authorsrebuttal2023-08-19

Dear Reviewer EN1R, We have answered the question regarding value decomposition baselines with risk, thoroughly discussed the limitations (in the global response), and discussed future directions. We have studied the performance of having a risk-aversive central critic update; the results are shown in the response PDF. We are looking forward to your reply. Please let us know if you have any other concerns. Thanks and Best Regards, Authors of Submission 6208

Reviewer rLaN6/10 · confidence 1/52023-06-15

Summary

The authors introduce the Risk-sensitive Individual-Global-Max (RIGM) principle as a generalization of the Individual-Global-Max (IGM) and Distributional IGM (DIGM) principles and propose RiskQ satisfies the RIGM principle. The method is evaluated by both theoretical analysis and experiments.

Strengths

1. The paper is easy to follow. 2. The idea of considering risk in multi-agent reinforcement learning is sound. 3. The method is evaluated by both theoretical analysis and experiments.

Weaknesses

1. It is possible that considering risk could potentially be detrimental to solving tasks that suffer from the relative overgeneralization problem. Agents may exhibit a tendency to select less risky actions, leading to convergence towards a conservative policy. It would be beneficial if the authors could analyze the representation capacity of their proposed method.

Questions

1. Is this work also related to safe RL? 2. Which behaviors in standard SMAC scenarios can potentially involve risk?

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer X7fQ5/10 · confidence 3/52023-07-06

Summary

The authors propose RiskQ, a value function factorization method for MARL that considers the risk factor. Risk is a concept present in reinforcement learning, and the authors propose a generalisation of the two popular principles IGM and DIGM: the RIGM, that considers the factor of risk in MARL.

Strengths

In general, the content of the paper is well structured and organized. - Provides extensive theoretical analysis of the proposed method - Extends current popular principles in MARL: IGM and DIGM - Provides empirical results in a wide range of different environments

Weaknesses

- In section 2, the authors mention Dec-Pomdps bu they do not introduce it. Additionally, none of the notation is defined, making it harder to understand some of the concepts, although these can be understood. This should be improved for clarity and completeness purposes. - In definition 3, value at risk is defined and is stated before as being a common metric (stated in the abstract, for example). Yet, I cannot find a reference for this metric in the text when value at risk is mentioned. - In definition 4, the DRM is defined, but it is unclear for me what $g(\omega)$ exactly represents; this is not mentioned. - In line 161 it is stated that the RIGM principle is a generalization of DIGM and IGM. I believe it would be more accurate to say that RIGM is a generalization of DIGM since this is already a generalization of the IGM. - In section 5.4 the authors describe some scenarios that are risk-averse versions of SMAC introduced in other works. For completeness, the descriptions and specifications of these environments should also be part of this paper. - In line 33: "Risk refers to the uncertainty of future outcomes in multi-agent systems". I can see some parallelisms between this and curiosity-based approaches, that aim to improve exploration based on states that they don't know. In fact, from my understanding, risk is also a way to improve exploration. I believe it would be beneficial to include some discussions with this or other exploration methods. Minor: - line 80: "They models" - line 59 and line 61: "jointed" - in line 102-103: "In this work, we interchangeablt use the terms: stochastic value function and return distribution" but right ahead in lines 133-134 the authors distance again the two terms "global value function (rather than a value distribution)"; these ambiguities can be confusing for the reader Although it seems promising, I believe that this paper still contains some lack of clarity that needs to be improved. Please find some questions ahead.

Questions

- In section 4.3, in lines 231-232 it is stated how $\psi$ is integrated in the learning problem. According to line 158, $\psi_a$ is expected to return a reward. How is this integrated in the learning problem? It is unclear to me. - According to section 4.3 and more precisely lines 236-240, I don't see hoe the risk factor is also applied on $Z_{jt}$ as stated in Definition 5 (RIGM). From the description in these lines it sounds to be like RMIX, as described in line 192. Could the authors clarify this?

Rating

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

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Please find above.

Reviewer e2976/10 · confidence 4/52023-07-21

Summary

This paper propose a factored distributional value learning for multi-agent reinforcement learning. The paper also propose a theoretically desirable factored risk-distortion property to which the proposed method adhere.

Strengths

The theoretical argument establishes an intuitively sound criterion for evaluating and developing risk distorted methods. The manuscript is written clearly, the theoretical results are proved, and details are comprehensively provided in the appendix.

Weaknesses

The empirical improves are marginal and does not seem to validate the claim of “RiskQ can yield promising results in environments that require risk-sensitive cooperation”, especially considering the quantile loss update includes potentially many quantile samples, effectively increasing the batch-size; not to mention the distributional approach itself provides a richer learning target and potential performance gains.

Questions

RIGM seems very important on a theoretical stand point according to the manuscript, how important is it empirically? Any simple counterexamples that proves its usefulness? or deficiencies of DIGM? The correctness argument for the proposed method seems to assume the learned CDF, in one way or another, resemble or converge onto the true distribution. How sound is this assumption? Can it be validated in the experiments? If so, it can significantly strengthen the correctness argument. A related work, likelihood quantile networks (Lyu et al.), mention that risk-seeking may be beneficial in MARL exploration, but obviously it depends. In what way would you recommend risk distortion can actually be used for exploration-exploitation trade-off?

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 method has the same limitations as other factored value learning methods, such as scalability. In addition, another limitation is that the method does not readily extend to mixed-reward settings in a scalable way.

Reviewer rLaN2023-08-15

Thanks for the response.

Authorsrebuttal2023-08-16

Thanks for the reviewer's time and effort to read our response. We are eager to discuss with the reviewer regarding our work. Here, we describe our work further to give the reviewer a better overview of the value factorization methods. Our work is a value factorization approach. Value factorization (or value decomposition) methods learn a factored state-action value function (utility function) for each agent. Each agent acts greedily according to the state-acton function. For value factorization methods, the **individual-global-max (IGM)** principle is an important principle. It requires that the optimal joint action equal to the collection of individual optimal actions. Many approaches have been proposed. For example, VDN [1] models the joint state-action value function as the sum of each individual state-action value function. QMIX [2] models the joint state-action value function as a monotonic increasing function with respect to each individual state-action value function. QTran [3] can factorize any factorized state-action value function and does not suffer from representation limitations. However, it is sample inefficient. QPLEX [4] and WQMIX [5] use different methods to learn state-action value functions that do not suffer from representation limitations. DMIX [6] extends the individual-global-max (IGM) principle to the **distributional individual-global-max (DIGM) principle**, which is used for distributional MARL. DIGM principle requires that for distributional state-action value function (stochastic value function), the optimal joint action is equal to the collection of individual optimal actions. However, it suffers from representation limitations too. ResZ [7] uses a residual function-based approach, it satisfies the DIGM principle without representation limitations. In this work, we believe that for risk-sensitive MARL scenarios, to learn factored risk-sensitive state-action value functions for each agent, a new IGM principle should be developed. The **risk-sensitive individual-global-max (RIGM)** principle requires that for risk-sensitive MARL, the optimal risk-sensitive joint action must equal the collection of individual risk-sensitive optimal actions. In this work, we use distributional RL algorithms and use risk-metric (e.g., CVaR) to calculate the risk from the return distribution of a state-action pair. It is non-trivial to design an algorithm that satisfies the RIGM principle. As we show in Theorem 1, methods that satisfy the IGM principle **do not guarantee adhere to the RIGM principle**, and in Theorem 2, methods that satisfy the DIGM principle **do not guarantee adhere to the RIGM principle** for some risk-metrics. Further, we have shown that the risk-sensitive MARL algorithm RMIX [8] satisfies the RIGM principle only for a special case, and the risk-sensitive MARL algorithm DRIMA [9] does not satisfy the RIGM principle for the CVaR metric. We have developed RiskQ, which satisfies the RIGM principle for the VaR and distorted risk measures. In the paper and the response PDF, we have shown that RiskQ performs significantly better than the other algorithms in risk-sensitive MARL scenarios. Through ablation studies, we show that it is not satisfying the RIGM principle will lead to a performance drop. And RiskQ performs better than its representation-limitation-free variants. **Reference** [1] Sunehag et al. Value-decomposition networks for cooperative multi-agent learning based on team reward. AAMAS 2018. [2] Rashid et al. QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning. ICML 2018. [3] Son et al. QTRAN: learning to factorize with transformation for cooperative multi-agent reinforcement learning. ICML 2019. [4] Wang et al. QPlex: Duplex dueling multi-agent q-learning. ICLR, 2021. [5] Rashid et al. Weighted QMIX: expanding monotonic value function factorisation for deep multi-agent reinforcement learning. NeurIPS, 2020. [6] Sun et al. DFAC framework: Factorizing the value function via quantile mixture for multi-agent distributional q-learning. ICML, 2021. [7] Shen et al. ResQ: A residual q function-based approach for multi-agent reinforcement learning value factorization. NeurIPS 2022. [8] Qiu et al. RMIX: learning risk-sensitive policies for cooperative reinforcement learning agents. NeurIPS, 2021. [9] Son et al. Disentangling sources of risk for distributional multi-agent reinforcement learning. In ICML, 2022.

Reviewer X7fQ2023-08-15

Thanks

Thanks for your response. Most of my questions were addressed. However, I believe the paper still needs some improvements in terms of clarity, as there are several terms that are not easy to follow with the flow of the paper. Additionally, there is a lot of important information in the Appendix that is not referred to in the main paper. I suggest that at least the authors refer to the appendix more often in the main body. With this being said, I will raise my score a bit since by main concerns have been answered, but I still believe the paper could be a bit improved in terms of clarity.

Authorsrebuttal2023-08-16

Thanks for your time and effort to read our response. We are happy to see that most of the questions are addressed. We will improve the clarity of this paper. For example, we will describe the two terms return distribution and stochastic value function in a better way. And we will refer to the appendix more often in the revised paper. Are there any other questions that we do not address? We are eager to answer and address the drawbacks of this paper. Your opinion is valuable for us to improve the quality of this work.

Authorsrebuttal2023-08-19

Thanks for your quick response. We are happy that your concerns have been addressed and appreciate that the rating has increased.

Reviewer e2972023-08-19

I appreciate the response and clarifications made by the authors. My questions are mostly answered. Now I do not see any critical downside in this work, I'm increasing my score.

Authorsrebuttal2023-08-19

Thank you for taking the time to review our rebuttal. We are glad to hear that the majority of your concerns have been resolved and there are no significant drawbacks. We are grateful for the increase in the rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC