Guarantees for Self-Play in Multiplayer Games via Polymatrix Decomposability

Self-play is a technique for machine learning in multi-agent systems where a learning algorithm learns by interacting with copies of itself. Self-play is useful for generating large quantities of data for learning, but has the drawback that the agents the learner will face post-training may have dramatically different behavior than the learner came to expect by interacting with itself. For the special case of two-player constant-sum games, self-play that reaches Nash equilibrium is guaranteed to produce strategies that perform well against any post-training opponent; however, no such guarantee exists for multiplayer games. We show that in games that approximately decompose into a set of two-player constant-sum games (called constant-sum polymatrix games) where global $\epsilon$-Nash equilibria are boundedly far from Nash equilibria in each subgame (called subgame stability), any no-external-regret algorithm that learns by self-play will produce a strategy with bounded vulnerability. For the first time, our results identify a structural property of multiplayer games that enable performance guarantees for the strategies produced by a broad class of self-play algorithms. We demonstrate our findings through experiments on Leduc poker.

Paper

References (51)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer WEmp6/10 · confidence 2/52023-06-26

Summary

The paper studies theoretical performance guarantees for agents learned using self-play in multiplayer games. Self-play is a common approach of machine learning in multi-agent systems for generating unbounded quantities of training data, but agents trained using self-play may perform poorly against new agents whose behavior differ dramatically from those seen during training. Despite guarantees previously established for self-play agents in 2 player constant-sum games, these guarantees do not extend outside of two-player constant-sum games. To solve this problem, the authors identify a structural property of multiplayer, general-sum game and use it to establish guarantees on the performance of strategies learned via self-play against new opponents. They show that any game can be projected into the space of constant-sum polymatrix games, which enables performance guarantees for the strategies produced by a broad class of self-play algorithms. The findings are empirically demonstrated on Leduc poker.

Strengths

- I find the studied topic which extends the theoretical guarantees of self-play beyond two-player (constant-sum) games to be important to the multi-agent learning community. - The proposed method, by projecting a general game into the space of constant-sum polymatrix games, is novel and interesting. - The theoretical results are verified on 3-player Leduc poker

Weaknesses

The technical sections of the paper is slightly hard to follow for a non-expert in this field.

Questions

- Can the authors pleaes explicitly list the assumptions/conditions needed for the theoretical results? - Can the structural property identified in this paper extends to other application domains (outside of multi-agent self-play)?

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

2 fair

Contribution

3 good

Limitations

I don't see negative societal impact.

Reviewer wuxJ8/10 · confidence 3/52023-07-05

Summary

This paper asks and answers the question: "In what games does self-play (with regret-minimizing algorithms) in multiplayer imperfect-information games perform well?" The motivation behind the question is that such algorithms, such as multiplayer versions of CFR (which has theoretic guarantees in 2-player zero-sum games), have empirically performed well in some multiplayer games, despite no previous theory guaranteeing that they will do so. (The salient example is no-limit Texas Hold'em, in the work "Superhuman AI for multiplayer poker", Brown & Sandholm 2019.) This paper answers the question primarily by looking at the concept of the coarse correlated equilibrium (CCE), since that is what the policies of the regret-minimizing players will converge to. The authors show that if games can be decomposed into a bunch of 2-player constant-sum games, called a constant-sum polymatrix (CSP) game, and the CSP game fulfills some additional properties, then we can bound the worst-case performance of a player's CCE strategy. They also show that even in games which can't be factored into polymatrix games (which is most games), we can project them into the space of CSP games, and still bound the worst-case performance of each player's CCE strategy. They also refine the analysis further by examining worst-case performance only against other self-play strategies. Finally, they perform experiments on multiplayer Leduc poker, and show that the empirical worst-case performances are within the bounds predicted by their theory.

Strengths

- I think this is an excellent paper. The question it asks and answers is fascinating, and one that I think was begging to be examined for several years now. Therefore, I judge this paper to be a significant contribution to science. - The results of the paper are novel and nontrivial. - The paper is well-written. The writing is clear and understandable. The introduction explains the background and motivates the problem well.

Weaknesses

The paper mentions no-limit Texas hold'em as a motivating real-world positive example of self-play working in multiplayer games. I also suggest that the authors mention some real-world negative examples -- for example, FAIR's line of work on Diplomacy, particularly the self-play agents performing poorly against humans in "No-Press Diplomacy from Scratch", Bakhtin et al 2021 https://arxiv.org/pdf/2110.02924.pdf (the problem is the motivation of "Mastering the Game of No-Press Diplomacy via Human-Regularized Reinforcement Learning and Planning" Bakhtin et al 2022: "the resulting agent DORA does very well when playing with other copies of itself. However, DORA performs poorly in games with 6 human human-like agents.") typo on line 336: "with ... be the set"

Questions

It's not clear to me how important the caveat in Footnote 3 on page 8 (section 6) is. So CFR does not necessarily converge to a CCE. The experiments in Section 6 use CFR. Is my interpretation correct that you take the marginal strategies of the CFR iterates, and hand-wavingly say that we can assume they are marginal strategies of (a distribution converging to) a CCE, even though they're not?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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 adequately addressed the limitations.

Reviewer Wd367/10 · confidence 3/52023-07-07

Summary

In multiplayer games, the authors derive bounds for the vulnerability of marginal strategies trained via no-regret self-play against other, uncorrelated agents independently trained via no-regret self-play. This is done by projecting games onto the space of constant-sum polymatrix (CSP) games, which can be decomposed into a set of 2-player constant-sum games between individual players. The closer this projection is, the tighter the bound on self-play's vulnerability to other self-play-trained agents is. This claim is validated by demonstrating that 3-payer Leduc can indeed be closely approximated as a CSP game and that the calculated vulnerability bounds are relatively close to the empirically measured vulnerability among marginal CFR strategies from many seeds.

Strengths

- A bound on the vulnerability of marginal no-regret strategies to other uncorrelated no-regret strategies is novel and highly useful to the game theory/AI community. - The paper is well written and concepts are clearly explained.

Weaknesses

- I would have preferred to see a range of common tractable games examined to get an intuition on when and how often a game can be closely approximated as a CSP game (and thus provide a useful bound on vulnerability). Currently, experiments only include Leduc Poker. - The analysis method is not immediately transferable to large games, and this is clearly stated as a limitation. - Minor note: It's not explicitly stated until the very end of section 6 that the 3-player variant of Leduc Poker is used for experiments.

Questions

Are there any heuristics one may be able to look for in larger games that might indicate when a game has a high chance of being approximately decomposable as a CSP game? If there is intuition on this to be gained from explicitly analyzing smaller games, it would be a great thing to discuss.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

All limitations have been adequately addressed.

Reviewer hBPZ6/10 · confidence 4/52023-07-07

Summary

This paper explores the intriguing problem of why no-regret algorithms seem to approximate well in multiplayer games, a phenomenon that has been empirically demonstrated in multiagent poker. The authors identify a structural property in multi-player games that allows performance guarantees for strategies derived through self-play algorithms. They propose that multi-player games can be projected into a series of two-player constant-sum games, known as polymatrix games. The proximity of a game to this structure is hypothesized to diminish the effect of correlation issues on the removal of a mediator. The researchers take an algorithm-agnostic approach, which broadens the applicability of their analysis to a variety of game-theoretically inspired learning algorithms and MARL algorithms that converge to coarse correlated equilibria.

Strengths

The paper's strength lies in its novel approach to an important problem. The authors introduce theory that all multiplayer games can be projected onto the space of polymatrix games, offering a promising direction for further research. If a high subgame stability game exists within this space, no-regret learning algorithms are predicted to converge to low-exploitability strategies. Additionally, the authors' algorithm-agnostic approach ensures that the analysis remains broadly applicable to a variety of game-theoretically inspired learning and MARL algorithms.

Weaknesses

The paper's main weaknesses lie in its experimental section, which comes across as somewhat unclear. The section seems intended to demonstrate that CFR will always converge to a low-exploitable strategy due to the subgame stability of the game, but this is not clearly stated. If the goal is to show that no-regret algorithms will always converge, the experiment design should include many seeds with different hyperparameters and many different no-regret algorithms. Secondly, it's unclear how this theory can be practically applied. For instance, if a new game is introduced, can it be predicted ahead of time whether CFR or another no-regret algorithm will converge to a low-exploitability strategy? This potential application isn't clear and should be further highlighted if it is indeed feasible.

Questions

How can the theory be applied practically? Can it be used to predict whether a no-regret algorithm will converge to a low-exploitability strategy in a new game? Could the use of the term 'vulnerability' over 'exploitability' be explained or justified in this context? What are the detailed parameters and specifics of the experiments conducted, such as the number of seeds and the range of no-regret algorithms tested?

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

There are certain limitations in the terminology used in the paper. The authors use 'vulnerability' instead of 'exploitability', although 'exploitability' is the term most widely used and recognized in this field. In addition, the term 'self-play' can lead to confusion, as it can generally refer to methods where RL agents play against themselves, which are not no-regret. The term 'no-regret' would be clearer. Lastly, at the end of page 9: “and if there exists a game with this set with high subgame stability” do you mean “and if there exists a game in this set with high subgame stability”?

Reviewer wuxJ2023-08-10

response

Thanks for taking the time to respond. Glad to hear that my suggestion is good. The Hanabi experiments sound useful. Re: response to questions -- I think I understand now. Is this a correct take?: The distribution of joint strategies produced by CFR iterates does converge to a CCE. However, the footnote is explaining that if you take each player's marginal/average strategy, the strategy profile that you get from combining all of those is not a CCE. But this is fine, because the whole point of the paper is to analyze those marginal strategies anyways.

Authorsrebuttal2023-08-11

Yes that is correct! Just to add a couple details: CFR iterates are behavior strategies, whereas CCE are distributions over pure strategy profiles. This is why you need an extra step with CFR-JR to convert the behavior strategies to equivalent mixed strategies when you want to extract the empirical distribution of play. But you don't need to do this to extract the marginals in behavior strategy form. The marginals *could* be CCE themselves (if they are a Nash equilibrium) but this is not necessarily the case.

Reviewer WEmp2023-08-11

Thanks for taking the time to respond! The clarification of the assumptions is useful to know, and thanks for commenting on the extensions!

Reviewer hBPZ2023-08-14

Response to Authors

Thanks for answering my questions. After reading the rebuttal and other reviews I choose to keep my score at a 6.

Reviewer Wd362023-08-15

Thank you for your replies. Adding Hanabi is a great addition. That addresses the one meaningful weakness I had with this work. I also appreciate that you plan to clearer distinguish that, by 'self play', you mean 'no-regret self play' as hBPZ mentioned. This could lead to a small amount of initial confusion for some readers.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC