Extensive-Form Game Solving via Blackwell Approachability on Treeplexes

In this paper, we introduce the first algorithmic framework for Blackwell approachability on the sequence-form polytope, the class of convex polytopes capturing the strategies of players in extensive-form games (EFGs). This leads to a new class of regret-minimization algorithms that are stepsize-invariant, in the same sense as the Regret Matching and Regret Matching$^+$ algorithms for the simplex. Our modular framework can be combined with any existing regret minimizer over cones to compute a Nash equilibrium in two-player zero-sum EFGs with perfect recall, through the self-play framework. Leveraging predictive online mirror descent, we introduce Predictive Treeplex Blackwell$^+$ (PTB$^+$), and show a $O(1/\sqrt{T})$ convergence rate to Nash equilibrium in self-play. We then show how to stabilize PTB$^+$ with a stepsize, resulting in an algorithm with a state-of-the-art $O(1/T)$ convergence rate. We provide an extensive set of experiments to compare our framework with several algorithmic benchmarks, including CFR$^+$ and its predictive variant, and we highlight interesting connections between practical performance and the stepsize-dependence or stepsize-invariance properties of classical algorithms.

Paper

Similar papers

Peer review

Reviewer qS3q6/10 · confidence 4/52024-07-03

Summary

This paper introduces a method, predictive treeplex Blackwell, for using Blackwell approachability directly on the treeplex in order to perform regret minimization in extensive-form strategy sets. They show that their algorithm achieves $O(\sqrt{T})$ regret (where $O$ hides polynomial factors in the game size), and a smoothed version of their algorithm enjoys $O(1/T)$ convergence toward Nash equilibrium when used by both players in zero-sum game.

Strengths

The method is conceptually very interesting, and shows that Blackwell-based stepsize-invariant regret minimizers are not solely restricted to the simplex. The experiments are comprehensive and illuminate clearly the authors' message about the role of infoset-level stepsize invariance. Thus, despite some minor issues listed below, I am generally in favor of acceptance.

Weaknesses

My most significant issue is simple: although the method is conceptually interesting, it is unclear if it carries any advantages over CFR (see also questions below re. clairvoyant CFR). For example, in theory, by Corollary 4.3 it seems that the convergence rate is something like $O(d^{5/2}/\sqrt{T})$ (since $\hat\Omega \le \sqrt{d}; \lVert \boldsymbol M \rVert_2 \le d$ -- maybe tighter analysis is possible but my point would stand regardless), compared to the $O(d/\sqrt{T})$ achieved by CFR-based methods. Similarly, the method seems consistently outperformed or matched by PCFR+ in basically every game, and the per-iteration complexity is inferior by a logarithmic factor. Minor notes: 1. In Proposition 3.1, the regret of Algorithm 1 should be expressed in terms of the regret of the regret minimizer on $\mathcal C$. Also, technically, $\text{cone}(\mathcal T)$ is an infinite set, so by "regret of the regret minimizer on $\mathcal C$" I really mean "regret of the regret minimizer on $\mathcal C$ against vectors $\boldsymbol{\hat x} \in \mathcal T$". The proper regret bound holds for OMD/FTRL with gradient dynamics, but these details, I think, should be spelled out. See e.g. Proposition 2 and the discussion afterward in [12].

Questions

1. The paper claims that Smooth PTB+ is the first Blackwell-based $O(1/T)$-converging algorithm for EFGs. But doesn't the algorithm it is based on, namely the Clairvoyant CFR+ algorithm of [14], also achieve the same convergence rate? (see Appendix J of [14]). 1. It is interesting to me that RM+ does not coincide with TB+ when the domain is a simplex (Appendix E). What happens experimentally if you use (P)TB+ instead of (P)RM+ at every information set in CFR? 1. Is it possible to extend the fundamental ideas of this paper beyond treeplexes to other convex sets, to obtain more stepsize-invariant regret minimizers?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

yes

Reviewer PEWS7/10 · confidence 4/52024-07-11

Summary

This paper studies a Blackwell's approachability method for solving extensive-form game. Rather than applying it at each infoset as in the CFR approach, it applies it globally, which allows for a $\mathcal{O}(1/T)$ convergence rate with a predictive version.

Strengths

This paper is well written and well presented. The results are sound, and the authors are the first to obtain a $\mathcal{O}(1/T)$ regret . It is surprising that such an approach has not been published before.

Weaknesses

The fact that CFR+ obtains good practical results because no learning tuning is needed is more or less already known. The experiments are a bit hard to read.

Questions

The conclusion mentions that the algorithm gets worse practical results than the CFR approaches, despite the stepsize invariance. This implies that the invariance may be necessary at the infoset level. Do you think it is possible to change your approach to allow for such infoset invariance?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors mentioned the limit of their approach (mainly the invariance mentioned above). There is no potential negative societal impact behind their work.

Reviewer n7s44/10 · confidence 4/52024-07-14

Summary

This paper designs a new algorithm for computing minimax equilibria in two-player zero-sum extensive form games. It is well-known that one way to do this is to have both players run no-regret algorithms against each other and take the time-average of their strategies. The main innovation of this paper is designing a new class of no-regret algorithms for extensive form games via a reduction from Blackwell approachability. In particular, minimizing regret in a zero-sum extensive form game is equivalent to minnimizing external regret in an online linear optimization problem over a polytope called the treeplex. The authors show that if you cast this OLO problem as a Blackwell approachability problem, and then perform an approachability to OLO reduction (along the lines of Abernethy et al.), you can reduce the original OLO problem of minimizing external regret over the treeplex to minimizing external regret over the conical hull of the treeplex. If you use predictive online mirror descent (with Euclidean distance) to solve the resulting OLO problem, you get an algorithm for the original problem the authors call Predictive Treeplex Blackwell+ (PTB+). This algorithm is stepsize independent and has O(1/sqrt(T)) convergence to equilibria. If you “smooth” this algorithm by projecting onto a truncation of the cone, you get a different algorithm the authors call Smooth PTB+, which the authors show has O(1/T) convergence to equilibria (but is no longer stepsize invariant). The authors implement these algorithms (and some other transformations of existing algorithms, e.g. Adam and AdaGrad) and compare them experimentally to a range of existing algorithms. They find that PTB+ performs the best out of these new algorithms, but still is significantly worse than the state-of-the-art PCFR+ (predictive counterfactual regret minimization) on some games. The authors hypothesize this is due to the fact that PCFR+ has an even stronger form of stepsize invariance than PTB+.

Strengths

Extensive-form game solving is one of the big successes of the theory of learning in games (with tools like regret minimization being directly used to construct superhuman-level algorithms for games like poker). This paper proposes a class of novel algorithms for extensive form game solving and both theoretically and empirically analyzes their performance (showing that this class of algorithms can have theoretical convergence rates on par with the best known algorithms). This analysis is relatively thorough and the paper is well-written and easy to read.

Weaknesses

Overall, I am a little unimpressed by the results of this paper. While it is true to the best of my knowledge that this class of algorithms (as applied to extensive-form game solving) is novel, it doesn’t really seem like they unlock any new guarantees that were not previously achievable. Several times throughout the paper the authors emphasize that this is the first algorithm “based on Blackwell approachability” to achieve these guarantees, and that this resolves an interesting open question. But it is not clear to me that “based on Blackwell approachability” is really a well-defined concept (perhaps you could recover some existing algorithms via other applications of approachability) or even a desired one. I also feel that the use of Blackwell approachability here is a little superfluous. The authors use Blackwell approachability to reduce an OLO problem on T to an OLO problem on cone(T). The eventual reduction is very simple (it essentially boils down to “projectvizing” T by adding an extra coordinate) and it is easy to see directly that the regret of the cone(T) OLO algorithm bounds the regret of the overall algorithm (that said, it is a nice observation that OLO algorithms for cone(T) seem give rise to "stepsize invariant" algorithms for original problem). It should also be pointed out that there has been significant work on Blackwell approachability since the work of Abernethy et al., including several papers which resolve some of the deficiencies the authors point out in Appendix C (which mostly stem from the fact that Abernethy et al. only consider the L_2 norm). I would recommend the authors look at “Refined approachability algorithms and application to regret minimization with global costs” by Kwon or “Pseudonorm Approachability and Applications to Regret Minimization” by Dann et al. This would perhaps be okay if these new algorithms were shown to empirically significantly outperform state-of-the-art on some class of games, but this does not seem to be the case; if anything, they seem to underperform existing algorithms such as PCFR+. I think some of the resulting conjectures about the role of stepsize invariance (and different types of step-size of invariance) on the practical performance of these algorithms are interesting, but they are not very convincingly explored in this work.

Questions

Feel free to reply to any part of the review above.

Rating

4

Confidence

4

Soundness

4

Presentation

3

Contribution

2

Limitations

Limitations adequately addressed.

Authorsrebuttal2024-08-11

As the end of the discussion period approaches, we would like to ask if our responses address your concerns and comments. We remain available to provide further clarifications on our work.

Reviewer n7s42024-08-12

Thank you for the detailed response. After reading through it (and the other reviews and comments), I've decided to maintain my current evaluation of the paper.

Reviewer WD2M7/10 · confidence 2/52024-07-15

Summary

This paper studies computations via regret minimization of Nash equilibria in zero-sum extensive form games (EFG) with the perfect recall assumption. The actions of the players are a sequence of polytopes (treeplexes). In prior work (counterfactual regret minimization framework), this was solved with methods that run regret minimization locally for a phase in the game on the corresponding information sets. Common regret minimization algorithms used are regret matching that operates on the simplex and is based on Blackwell approachability, or online mirror descent (OMD). Using regret matching as a local optimizer showed a better empirical performance. This paper to develop Blackwell approachability-based algorithms directly for the treeplexes (instead of locally). The benefit of using Blackwell approachability is a property named "stepsize invariant", which means that stepsizes across different information sets, and the iterates of the algorithm main do not depend on them. This is different from running OMD on the treeplex. Several algorithm instantiations are also provided and tested via numerical experiments. The main message of the papers is to claim that information sets stepsize invariance seems like a crucial property for good empirical performance and shed light on the strong empirical performance with such property, such as CFR+ (Counterfactual regret minimization with regret matching+ as a local optimizer)

Strengths

The computation of Nash equilibria in zero-sum extensive form games via iterative methods is fundamental in "learning with games"/"self-play". This paper studies natural approaches to tackle this problem and perhaps sheds light on an interesting property for strong empirical performance. Although I'm not an expert in the field, the authors give enough background to explain how their approach is related to prior work. In my understanding, this paper makes a significant contribution.

Weaknesses

Couldn't find Weaknesses

Questions

-

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations are properly addressed.

Reviewer qS3q2024-08-08

Thank you. My opinion of the paper was and remains positive, and I will keep my score.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC