A Finite-Sample Analysis of Payoff-Based Independent Learning in Zero-Sum Stochastic Games

We study two-player zero-sum stochastic games, and propose a form of independent learning dynamics called Doubly Smoothed Best-Response dynamics, which integrates a discrete and doubly smoothed variant of the best-response dynamics into temporal-difference (TD)-learning and minimax value iteration. The resulting dynamics are payoff-based, convergent, rational, and symmetric among players. Our main results provide finite-sample guarantees. In particular, we prove the first-known $\tilde{\mathcal{O}}(1/\epsilon^2)$ sample complexity bound for payoff-based independent learning dynamics, up to a smoothing bias. In the special case where the stochastic game has only one state (i.e., matrix games), we provide a sharper $\tilde{\mathcal{O}}(1/\epsilon)$ sample complexity. Our analysis uses a novel coupled Lyapunov drift approach to capture the evolution of multiple sets of coupled and stochastic iterates, which might be of independent interest.

Paper

References (82)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer bFVC5/10 · confidence 4/52023-07-05

Summary

The authors propose a doubly smoothed best-response dynamics for two-player zero-sum Markov games, with matrix games as the degenerate case. Upper bounds of Nash Gap with bias and regularized Nash Gap without bias are presented.

Strengths

The technical sections are concrete and solid. The idea of simultaneously smoothly updating in the value (q) space and the policy (simplex) space appears novel. In such a way, each player is facing a much more stable environment than simply smoothing in the value space. As a consequence, the convergence results are natural outcomes of such an algorithmic design. Moreover, the Assumption 3.1 makes the results more appealing in a sense that it deviates from typically made assumptions in RL analysis.

Weaknesses

Some major concerns: (1) The "first form of smoothing" in the policy space is very similar to the mixed strategy in fictitious self-play but not exactly the same. In fictitious self-play, the policies are mixed in the way of having $1-\alpha_k$ vs $\alpha_k$ probabilities to select the old policy vs the new best response. In comparison, here the authors simply averaged two stochastic strategies. Even with the explanations provided in the paper, it is still not very clear to me why the authors would want to deviate from the classical practice in learning Markov games (or extensive form games), which is backed by enormous amount of practical implementations. (2) While the authors claim that the learning dynamics are independent for the two players, the learning rates of them are tied through a constant, which makes the claim somewhat questionable. Please elaborate more about your definition of being "independent". (3) The bound for Nash gap is somewhat weak. Apart from the term $\ell_\tau$ that is exponential in $\tau$, the constant $c_2$ also has both $O(\tau)$ and $O(\tau^{-1})$ terms, which requires balancing. Overall, I think a corollary with the best choice of $\tau$ is highly desirable to at least let readers to directly evaluate the strength of the bound. (4) The bound of regularized Nash gap is not at all surprising -- it appears rather standard in minimax (saddle point) optimization with regularized geometry. So the real emphasis should be the Nash gap itself. Although I see the authors discuss the natural difficulty for softmax policy algorithms to converge without bad dependencies on the temperature $\tau$, it still feels quite vague why this is the case -- I believe a deeper illustration with math will greatly strengthen the paper -- it won't make the Nash gap look as unsatisfying as it appears now.

Questions

Please see the Weakness section.

Rating

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

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

Yes, there is no potential negative societal impact.

Reviewer zACq6/10 · confidence 3/52023-07-07

Summary

This paper studied a best-response type learning dynamics in two-player zero-sum stochastic games called doubly smoothed best-response dynamics. The dynamics uses smoothed value function updates with softmax smoothed best-response, and combines minimax value iteration. Ths dynamics is payoff-based, convergent, rational, and symmetric. The author also provided the first finite-sample analysis of these type of dynamics, showing convergence to Nash equilibrium up to smoothing bias.

Strengths

1. Finite-sample analysis (or convergence rates analysis) for payoff-based (or noisy bandit feedback-based) learning dynamics in stochasitc games is an interesing question with practical importance. This paper contributes to this area by the first finite-sample analysis of a payoff-based best-response-type independent learning dynamics for zero-sum stochastic games. 2. This paper is well-written and easy to follow. I appreciated it that the authors provided detailed discussion and high-level ideas of the algorithm and proof sketch for the main results.

Weaknesses

The claim that the proposed dynamics is *convergent* and *rational* seems not rigorous since the smoothing bias $\tau$ persists over time and no convergence to exact Nash equilirium or best response is really proved. One possible fix might be introducing a decreasing parameter $\tau_t$ so that exact convergence can be shown. This might significantly slow down the convergence of other bias terms because of the exponetionally dependence on $\tau$. The proposed dynamics is thus very slow in terms of convergence to Nash equilirium and this is a weakness of the current work. Minor comments: 1. Line 277, $Q(s,a^{i}, a^{-i})$ has not been defined. 2. Under Line 284: one term should be $R^i(s, a^{i}, a^{-i})$.

Questions

1. Do Algorithm 1 and 2 provide sublinear regret guarantee for one player when the other player might be an adversary?

Rating

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

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer r3LS6/10 · confidence 3/52023-07-07

Summary

Authors propose algorithms for learning the Nash equilibrium in two-player games and two-player stochastic games. The algorithm for games is effectively a single-time scale algorithm (Doubly smoothed best response dynamics). While , the algorithm for stochastic games (Double smoothed best response dynamics with value iteration) is a two-time scale approach. They show that these algorithm is independently implemented by each agent without any communication requirement and the rewards are based on actual payoffs obtained after each each action, rather than the full information setting. Also, they also show that the players reach a policy that is a best-response of the other players' (stationary) policies.

Strengths

The aim to reach convergence in last-iterate in zero-sum games and stochastic zero-sum games through independent learning is a challenging problem and the authors provide not only a finite sample guarantee, but also in addition, that the learned policies are best-responses to the stationary policies of other agents and achieve "rational" learning (Bowling and Veloso ' 2001). Also, the assumption about existence of a joint policy that induces an irreducible and aperiodic Markov chain is relatively weaker than other existing assumptions, that assume any policy always induces such an irreducible and aperiodic Markov chain or any policy created by the algorithm's trajectory is uniformly geometrically ergodic. They use coupled Lyapunov drift inequalities to guarantee convergence and finite sample guarantees.

Weaknesses

1) If the game has a non-interior Nash, do you still get convergence? Lemma D.1 seems to say that all strategies are played with a probability lower bounded by a certain value. 2) Is there any hope of getting a stronger probabilistic guarantee for at least the two-player zero-sum games setting? Essentially to understand the convergence of the iterates (in distribution or w.h.p)? instead of an expected sense as derived in this paper). 3) Authors state that prior two-time scale approaches might require implicit coordination amongst players, however, their algorithm for the stochastic games is also two-time scale and it would help if they could clarify how they are able to avoid this in the algorithm and how they make up for this in the analysis.

Questions

See weaknesses.

Rating

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

They have mentioned the limitations in the paper.

Reviewer Tp2P7/10 · confidence 3/52023-07-13

Summary

The authors focus on the problem of finite-sample convergence analysis of independent best-response-type learning dynamics in two-player zero-sum stochastic games. The dynamics are payoff based and stimulate the Shapley operator (that is known to be a contractive). Under the assumption that for any pair of policies, the induced Markov chain is irreducible and aperiodic, it is shown that in expectation the Nash gap is at most epsilon after 1/eps steps of the algorithm. Finally the authors provide results rationality type results (i.e., on the regret when the one player follows the dynamics and the other plays a stationary policy).

Strengths

I think that the result of 1/eps rate is quite interesting and surprising (I would expect a rate of 1/eps^2). Moreover the techniques seem highly non-trivial (design of a novel Lyapunov type function).

Weaknesses

Sometimes the write-up is not self-contained, but this is because of the limited space.

Questions

Can you please explain if your results carry over for other settings like Markov potential Games?

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Not applicable

Reviewer zACq2023-08-11

Acknowledgment of Rebuttal

Thank you for the detailed reply! I have no further questions.

Authorsrebuttal2023-08-14

We thank the reviewer for the feedback.

Reviewer bFVC2023-08-14

Re: Rebuttal by Authors

Thanks for your responses and I am generally glad with the changes the authors promise to make in the next version. Since some weaknesses still stand, I'm raising my score to 5.

Authorsrebuttal2023-08-14

We thank the reviewer for the feedback and if you have any other questions, please let us know.

Reviewer r3LS2023-08-14

Thank you for the rebuttal

I thank the authors for their rebuttal and I understand this is mainly a theoretical study, but did the authors try their algorithms against well known algorithms (MWU, OMWU) in simple games such as Rock, Paper, Scissors etc? I think it might be useful to shed light on the smoothing bias, temperature and how this affects the convergence. Other than that my initial questions have been clarified.

Authorsrebuttal2023-08-14

We thank the reviewer for the great suggestion. We will include numerical simulations on Benchmark examples (such as the RPS game suggested by the reviewer) in our next version and compare our learning dynamics with existing algorithms such as MWU and OMWU. Moreover, we will also investigate the dependence on the temperature $\tau$ (which determines the smoothing bias) in our numerical simulations.

Authorsrebuttal2023-08-19

We thank the reviewer for the feedback.

Area Chair rHGt2023-08-21

Quick question

Dear authors, Could you clarify what is the difference between Algorithm 1 in your paper and the 2003 algorithm of Leslie and Collins, "Convergent multiple-times-scales reinforcement learning algorithms in normal form games", The Annals of Applied Probability 13 (2003), no. 4, 1231–1251 [Eq. (6) in p. 8]? I understand that you use a different tuning for $\alpha_k$ and $\beta_k$ but, given that Leslie and Collins establish convergence with probability $1$ in zero-sum games (not just in expectation), this raises the question why the standard choice $\beta_k/\alpha_k \to 0$ would not be more appropriate in your setting as well. Regards, The AC

Authorsrebuttal2023-08-21

Thank You for the Comments

Dear AC, **The Difference:** We thank the AC for pointing out Leslie and Collins (2003), which we will include in our literature review. The main difference between Algorithm 1 of this work and the one in Leslie and Collins (2003) is that our learning dynamics is *single time-scale* while the one in Leslie and Collins (2003) is *two time-scale* as they require $\lambda_n/\mu_n\rightarrow 0$. Technically, the fact that we do not have a time-scale separation between the two sets of iterates presents major technical challenges in establishing the finite-sample bound, which we overcome with our coupled Lyapunov approach. In terms of results, Leslie and Collins (2003) consider only the asymptotic convergence in matrix games while we establish the rate of convergence (*which is more informative and challenging compared with asymptotic convergence*) in both matrix and stochastic games. **Our Results Imply the Mean-Square Convergence of the Last Policy Iterate** Thanks to reviewer r3LS’s second comment, we realized that algorithm 1 does enjoy a stronger convergence guarantee (not just in expectation for the regularized Nash gap). Specifically, recall that the Nash distribution (denoted by $(\pi_\tau^1,\pi_\tau^2)$) is the unique minimizer of the regularized Nash gap defined right before Corollary 2.2.1, where the uniqueness part follows from the regularized Nash gap being a strongly convex function. Therefore, by the quadratic growth property of strongly convex functions, we have $\sum_{i=1,2}\lVert \pi_k^i-\pi_\tau^i \rVert_2^2\leq c RNG(\pi_k^1,\pi_k^2)$ for some constant $c>0$, which, after combining with Corollary 2.1.1., provides the mean-square convergence of the last iterate $(\pi_k^1,\pi_k^2)$. We will include this result as a corollary in our next version. **Single Time-Scale vs Two Time-Scale** The stepsize condition imposed in Leslie and Collins (2003) is standard for establishing the almost sure convergence (Borkar 2009) while our goal is to characterize the rate of convergence, where the stepsize choice might be different to obtain the best possible rate. While it is theoretically unclear if there is a fundamental gap between the convergence rates of two time-scale stochastic iterative algorithms (of which the algorithm in Leslie and Collins (2003) is a special case) and their single time-scale counterparts, in the existing literature for finite-sample analysis, the established upper bounds for two time-scale algorithms are mostly **sub-optimal**. For example, the convergence rate of two time-scale SGD established in Zeng et al (2022) was $k^{-1/4}$ while single time-scale SGD achieves $k^{-1/2}$. The convergence rate of the two time-scale natural actor-critic algorithm established in Khodadadian et al. (2022) was $k^{-1/6}$ while the lower bound for RL is $k^{-1/2}$, which is achieved with single time-scale algorithms such as $Q$-learning (Li et al. 2020). To obtain the best possible rate of convergence (cf. the $1/k$ rate convergence in Corollary 2.1.1 of Algorithm 1), we adopt a single time-scale framework. **Almost Sure Convergence vs. Finite-Sample Analysis** Asymptotic convergence in general does *not* imply any rate of convergence. In addition, the techniques and conditions to establish asymptotic convergence and finite-sample bound are fundamentally different. In particular, consider the requirements on choosing the stepsizes. Asymptotic convergence typically requires the stepsizes to be squared summable (Robbins and Monro 1951), recent results in Bhandari et al.( 2018); Chen et al (2022) show that mean-square finite-sample bound can be established with stepsizes of the form $1/k^z$ for all $z\in (0,1]$. Note that when $z\in (0,1/2)$, the stepsizes are not squared summable. Borkar, V. S. (2009). Stochastic approximation: a dynamical systems viewpoint (Vol. 48). Springer. Zeng, S., Doan, T. T., & Romberg, J. (2021). A two-time-scale stochastic optimization framework with applications in control and reinforcement learning. arXiv preprint arXiv:2109.14756. Bhandari, J., Russo, D., & Singal, R. (2018, July). A finite time analysis of temporal difference learning with linear function approximation. In Conference on learning theory (pp. 1691-1692). PMLR. Chen, Z., Zhang, S., Doan, T. T., Clarke, J. P., & Maguluri, S. T. (2022). Finite-sample analysis of nonlinear stochastic approximation with applications in reinforcement learning. Automatica, 146, 110623. Khodadadian, S., Doan, T. T., Romberg, J., & Maguluri, S. T. (2022). Finite sample analysis of two-time-scale natural actor-critic algorithm. IEEE Transactions on Automatic Control. Li, G., Wei, Y., Chi, Y., Gu, Y., & Chen, Y. (2020). Sample complexity of asynchronous Q-learning: Sharper analysis and variance reduction. Advances in neural information processing systems, 33, 7031-7043. Best, Authors

Area Chair rHGt2023-08-21

Dear authors, Thanks for confirming that the algorithm is the same as that of Leslie and Collins (LC in the sequel), only with a different schedule for $\alpha_k$ and $\beta_k$. For the rest, my question still stands: the LC schedule guarantees almost sure convergence, but its convergence rate is not known; your choice of schedule comes with rates, but it does not guarantee almost sure convergence (only in expectation). Given that the LC algorithm is quite classical by now, it is natural to ask whether you can get a rate of convergence under the LC schedule - or, rather, for general $\alpha_k$, $\beta_k$ - before settling on a particular schedule. On this point, the papers by Zeng et al. and Khodadadian et al. are not particularly relevant because they concern completely different settings and algorithms - for example, the time-scales in Khodadadian involve an explicit exploration step, which is well-known to slow down convergence. There are other papers in the game-theoretic literature where different time-scales *do* lead to better rates - for example, Hsieh et al., "*Explore aggressively, update conservatively: Stochastic extragradient methods with variable stepsize scaling*", NeurIPS 2020 - but it is impossible to make a meaningful comparison across different settings, so let us focus on the precise question at hand. Regards, The AC PS: This point is not particularly relevant in the discussion (so I'm not including it above), but square summability is not required to establish asymptotic convergence, see e.g., Propositions 4.2 and 4.4 in Benaïm, "*Dynamics of Stochastic Approximation Algorithms*" (1999). In the setting of your paper, you could have step-sizes of the form $1/k^z$ for all $z\in(0,1]$, or even of the form $1/(\log k)^{1+\epsilon}$ for some $\epsilon > 0$.

Authorsrebuttal2023-08-21

Dear AC, Thank the AC for the prompt feedback. We want to clarify that the algorithm is the same as the one in LC up to a time-scale separation *only* in the matrix-game setting. The fact that we only require the stepsizes to be different up to a multiplicative constant enables us to run our algorithm with a single set of stepsizes. Furthermore, our main focus is on the *stochastic-game* case as emphasized in our title. This requires the additional value iteration component in the algorithm, which is not present in the matrix-game setting, and presents technical challenges in handling the time-inhomogenous Markovian noise. As for the rate of convergence, in our proof of Theorem 2.1 and Corollary 2.1.1 (See Appendix D), we are, in fact, keeping the stepsizes as general as possible (i.e., using $\alpha_k$ and $\beta_k$ without specifying the decay rate) up until the last inequality of Line 1294, at which point imposing a time-scale separation clearly results in a **slower** rate of convergence. Specifically, in the last inequality of Line 1294, we establish the overall negative drift inequality of the combined Lyapunov function. Suppose that we impose a time-scale separation, which means that the stepsize ratio $c_{\alpha,\beta}$ is not constant but decays with $k$. Since $c_{\alpha,\beta}$ appears in the denominator in the last term on the right-hand side of the last inequality in Line 1294, the resulting rate from our proof technique will be slower compared with our current result without time-scale separation. Regards, Authors

Area Chair rHGt2023-08-21

Dear authors, > Furthermore, our main focus is on the stochastic-game case as emphasized in our title. There is no confusion on that - I am focusing on the normal form setting because this is where the overlap with LC lies. > Imposing a time-scale separation [in the last inequality of Line 1294] clearly results in a slower rate of convergence The trade-off in L1294 is clear: the two coefficients in the parentheses have to scale at the same rate in order to get sufficient decrease for your choice of Lyapunov function $V_R + \|\mathcal{R}\pi - q\|^2$. However, the definition of $V_R$ treats the Nash gap for the underlying game and the entropic smoothing differently: the standard Lyapunov function for this type of (smoothed) best reply algorithms would be $\sum_{i=1,2} \max_{\hat\mu^i}(\hat\mu^i - \mu^i) [ R^i \mu^{-i} - \tau \nabla\nu(\mu^i) ]$, i.e., the Nash gap for the modified game with payoffs $\mathcal{R} - \tau\nu$ (note that the Nash equilibria of this modified game are precisely the Nash distributions / logit equilibria of the original game). [Caveat: the sign of $\nu$ might be off depending on whether you include a "$-$" in the definition of the entropy or not; I'm not sure which convention you use in the appendix] This choice does not involve the two terms that you control separately in Lemmas D.3 and D.4, so it is not clear if the common scaling of $\alpha_k$ and $\beta_k$ is really necessary or if it is an artifact of the analysis. At any rate, I thank you for your replies, and I will be sure to reach out if any further points requiring clarification come up during the committee discussion stage. Regards, The AC

Authorsrebuttal2023-08-22

Dear AC, We thank the AC for pointing us to the new Lyapunov function. It is interesting to investigate whether using this Lyapunov function would result in an improved rate of convergence. On the other hand, our regularized Nash gap, i.e., $V_R$, is also a natural choice of the Lyapunov function, which was previously used to study the smoothed best-response dynamics in Hofbauer and Hopkins (2005) (see their Theorem 3.2). We will include a clear discussion (in words and in math) about the single time-scale vs two time-scale comparison and our choice of the Lyapunov functions in our next version, which will certainly improve this work. Hofbauer, J., & Hopkins, E. (2005). Learning in perturbed asymmetric games. Games and Economic Behavior, 52(1), 133-152. Best, The Authors

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC