NeurIPS 2020

Promoting Coordination through Policy Regularization in Multi-Agent Deep Reinforcement Learning


Review 1

Summary and Contributions: The paper hypothesized and proves that coordination-promoting inductive biases on policy search helps discover successful behaviors more efficiently. Two approaches are introduced in this paper to help search for coordinated policies: TeamReg, predicts the teammate behaviors to promote coordination, CoachReg, enables agents to recognize different situations and synchronously switch to different sub-policies. Evaluation was done on continuous and discrete action control tasks. Results show that there is a significant performance improvement in almost all the domains.

Strengths: Novel idea of extending agent modeling to introduce inductive bias to promote coordination behaviours in policy search. Ability of the proposed work to run in a distributed fashion Sound evaluation proving the claims made in this paper Highly significant and relevant in MARL

Weaknesses: Would like to see a wider variety of domains for evaluation, performance on domains having higher number of agents. Results of how proposed algorithms perform when agents scale The ability of agents accurately predicting other agents actions might not be true in many domains

Correctness: Results seem to reflect the claims of this paper

Clarity: Well written paper, includes all the details necessary.

Relation to Prior Work: Prior work has been explained clearly and the paper does a good job of letting the readers know of the current state of art in the field and why and how they extended it.

Reproducibility: Yes

Additional Feedback: Why is performance of TeamReg poor, given that the soccer domain opponents are rule based (so presumably the predictability of the opponents is not necessarily a problem?)


Review 2

Summary and Contributions: Based on rebuttal and discussion: Upon reading all reviews, I recognize that we agree the article is well presented, and I stand by the concerns I raised. Note that I primarily criticized the absence of some relevant context in the original submission (which the authors admit in their rebuttal), rather than the contribution itself (albeit it may be smaller than proclaimed). The author rebuttal alleviated some concerns. Their refutation of it being a planning setting is fair. While I maintain that it is a self-play setting, this is implied by CTDE and thus not necessary to state again. A stale flavor remains from overselling their contribution’s novelty in the introduction [L36-45]. As the authors confirm in their rebuttal, it would be fair to say they extend the auxiliary task of policy prediction (established in the literature) with predictability. While the rebuttal concentrates on CTDE, my review demanded a positioning within “opponent-prediction based auxiliary tasks and intrinsic motivation”, which I see as complementary but not equivalent to CTDE. ----- This article extends the algorithm MADDPG with auxiliary tasks of (a) opponent policy prediction (this has been done before, understated or missed by the authors) and predictabilty (of the agent itself), and (b) prediction of a synchronised exploration signal used during centralised training. It argues for the need of coordinated exploration and presents empirical comparisons of the novel extended algorithms (a) and (b) against ablated versions and DDPG.

Strengths: The article addresses the critical and highly relevant challenge of joint exploration in multi-agent reinforcement learning. It is presented in good language.

Weaknesses: My concerns are mostly with the difficulty of evaluating novelty and significance given some key omissions of related work. I would expect at least a technical discussion on the difference from related work, and why this was chosen. At the very least, it should be stated that opponent policy prediction has been shown to be beneficial previously, and is here applied to MADDPG. The article also seems to carry a 'deep bias' in literature selection. The article falls below expectations in terms of related work embedding (listed below), reflection and transfer insights. The specific algorithmic extensions seem motivated a posteriori, but the method of extending is not easily transferable. The motivation section focusses on action re-encoding, but does not reflect on this, leaving a perceived gap towards the extensions. Results are not clear cut, and require a lot of exceptions in Section 7.1.

Correctness: Overall, the article's technical arguments seem correct. The claim of novelty is questionable given the omissions of some key related work, and the lack of technical juxtaposition with some of what is cited (see related work comments below).

Clarity: The presentation is good, both in language and layout. The arguments that are made are easy to follow, albeit the related work section comes a bit late and remains superficial. The abstract does not explicitly state that this article addresses the planning setting (having access to the model), and operates in self-play (controlling all learning agents, especially for CoachReg).

Relation to Prior Work: - One key omission is the reference "A Deep Policy Inference Q-Network for Multi-Agent Systems" by Hong et al, 2018, which employed a KL-divergence opponent policy prediction loss (compare Eq. (4), (5) of both papers). Also, while [11] is cited, their use of cross-entropy opponent policy prediction is neither mentioned in Section 4.1 nor 5. These two omissions put a questionmark to the purported novelty and originality of the first extensions (called (a) in the summary above). - The introduction should highlight the line of work this article falls into, which to me is opponent-prediction based auxiliary tasks and intrinsic motivation. Some links are drawn in related work, but I would expect the high level positioning in the introduction. - The algorithm "MAVEN: Multi-agent variational exploration", co-authored by Shimon Whiteson, deserve mentioning and demarcation, as it similarly addresses the fundamental challenge of joint exploration. - The link of predictability to intrinsic motivation is recognized but not fully pursued or exposed to the reader. - The policy mask is not embedded in previous work, while it has similarities with attention models and hierarchical models (e.g., "Feudal networks for hierarchical reinforcement learning").

Reproducibility: Yes

Additional Feedback: - It is unclear (from the main article) if negative regularization parameters lambda are searched. - KL divergence is first abbreviated and later written in full. - "MADDPG, the most widely used algorithm" is an unsubstantiated overstatement - "This through evaluation" maybe 'thorough'? - While the broader impact section is fine, it could be strengthened by highlighting that joint exploration is indeed a bottleneck challenge in MARL.


Review 3

Summary and Contributions: The paper proposes two novel techniques to encourage the discovery of coordinated strategies in the centralized training decentralized execution (CTDE) framework.

Strengths: The paper is excellent in clarity. The paper clearly presents and explains the proposed techniques. The details of the experiments are given to make the work highly reproducible. The proposed techniques are compatible with any MARL algorithm in the Centralized Training and a Decentralized Execution (CTDE) framework, and thus the paper might be interesting to the broad research community in MARL.

Weaknesses: The empirical evaluation would be much more convincing if the proposed algorithms are compared with something stronger than MADDPG (and its variants). To your knowledge, is there any other work that aims to improve the centralized training in the CTDE framework? The introduction of the two techniques are somewhat ad-hoc, even after section 3 is presented for motivation. In particular, the ending of section 3 emphasizes on learning “the same type of constraint”, while the TeamReg seems to focus on purely predicting others’ decisions. Moreover, in section 4.2, the idea of policy masks is introduced without enough motivation/intuition. Is the idea of policy masking itself novel?

Correctness: Yes.

Clarity: Yes, the paper is well written in terms of clarity.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: Given that both TeamReg and CoachReg reduce to auxiliary losses, is there any reason why the authors didn't combine these losses? --- I am lowering my score a bit, but I am still learning towards accepting the paper.


Review 4

Summary and Contributions: This paper addresses how to learn collective behavior across multiple agents without succumbing to the curse of dimensionality and without resorting to single agent exploration. The key idea in the paper is that by coordinating the agents’ policies in order to guide their exploration it is possible to train multi-agent systems to exhibit competitive coordinated behavior. This policy coordination (or regularization) is a form of inductive bias. The paper presents two methods: TeamReg, based on inter-agent action predictability and CoachReg that relies on synchronized behavior selection. Empirical evaluation is performed on continuous control tasks with sparse rewards and on the discrete action Google Research Football environment.

Strengths: A key strength of the paper are two methods for policy coordination in a multi-agent team. As the authors acknowledge, their focus is on agent behaviors that are predictable given teammate behavior. This is a reasonable basis on which to build. The second strength of the paper is to point out that the widespread assumption that centralized training always outperforms decentralized training may not be valid.

Weaknesses: The key limitation of the work is that the empirical evaluation is inconclusive. The results in section 7.1 illustrate that on the COMPROMISE and CHASE environments more work remains to be done. On the BOUNCE environment more episodes are needed to clarify asymptotic behavior. The results in section 7.4 are more promising (how did you decide that 80K episodes was a reasonable number?) though I'm puzzled why MADDPG + TeamReg does not perform somewhat better. While MADDPG + CoachReg does seem to learn policies that achieve positive return the numbers overall are still small and it's not clear whether these results imply anything for scaling up to scenarios more complicated than 3v2.

Correctness: The empirical results are a promising start but needs more work to be conclusive.

Clarity: The paper is well-written and easy to read and understand.

Relation to Prior Work: The paper clearly makes a connection to prior work in the field and how it differs from previous contributions. It is well-situated in the literature.

Reproducibility: Yes

Additional Feedback: The extension of the baseline training (inlined figure in rebuttal) is appreciated.