Accelerated regularized learning in finite N-person games

Motivated by the success of Nesterov's accelerated gradient algorithm for convex minimization problems, we examine whether it is possible to achieve similar performance gains in the context of online learning in games. To that end, we introduce a family of accelerated learning methods, which we call"follow the accelerated leader"(FTXL), and which incorporates the use of momentum within the general framework of regularized learning - and, in particular, the exponential/multiplicative weights algorithm and its variants. Drawing inspiration and techniques from the continuous-time analysis of Nesterov's algorithm, we show that FTXL converges locally to strict Nash equilibria at a superlinear rate, achieving in this way an exponential speed-up over vanilla regularized learning methods (which, by comparison, converge to strict equilibria at a geometric, linear rate). Importantly, FTXL maintains its superlinear convergence rate in a broad range of feedback structures, from deterministic, full information models to stochastic, realization-based ones, and even when run with bandit, payoff-based information, where players are only able to observe their individual realized payoffs.

Paper

References (44)

Scroll for more · 32 remaining

Similar papers

Peer review

Reviewer sdz46/10 · confidence 3/52024-06-25

Summary

The paper studies the extension of Nesterov’s accelerated gradient algorithm to the solution of N-player games, named "follow the accelerated leader" (FTXL). The method is studied in both continuous and discrete time, and under various information oracles. The convergence of the algorithm is super-linear when started from an initialization close enough to the NE. Simulations are conducted which confirms the accelerated convergence numerically.

Strengths

The paper is very well-written and well-motivated. The core contribution, which is the accelerated algorithm with local super-linear convergence, is clearly presented, novel, and significant to my knowledge. That the method works under different information oracles, especially bandit feedback, is also important. Despite the simplicity, the simulations do numerically verify the accelerated rate established in theory.

Weaknesses

The convergence of FTXL in all settings requires sufficiently close initialization or has to do with the neighborhood $\mathcal{U}$. The exact definition of "sufficiently close" or this neighborhood is not given in the main paper. I believe this should be added in the statement of the theorems for clarity. The authors should also discuss how realistic it is that the initial point satisfies the requirement (and/or how likely the iterates of FTXL or FTRL may run into this "good neighborhood"). It seems that no assumption on the convexity of the payoff function is made. With a general non-convex payoff function, I wonder what is the problem structure and technical apparatus that allow the authors to derive the quadratic convergence (or even convergence at all). This is currently not sufficiently discussed in the main paper. While I am happy to study the detailed analysis, I would appreciate a sketch of the proof which highlights the main techniques.

Questions

Please see Weaknesses section.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Paper of theoretical nature. No negative social impact is foreseeable.

Reviewer 3QUL7/10 · confidence 4/52024-07-07

Summary

The submission studies the convergence of finite N-person games. Combining the idea of (NAG) and (FTRL), the submission proposes a continuous-time dynamic (FTXL-D) and its discrete-time scheme (FTXL). The novelty of integrating momentum and regularization into an algorithm allows the construction of quadratic rate methods.

Strengths

- A novel viewpoint that uses (HBVF)’s interpretation to link (NAG) and (FTRL-D) to form (FTXL-D). - Improved converge rates from linear to quadratic are shown in Theorems 1 – 4. Theorems 2 – 4 cover common practical scenarios. - Clear comparison between the proposed method and the previous methods, especially at the intuition level. - Pinpoint the critical term to focus on the technical details (U in Theorem 3). This greatly increases the time to understand the analysis. - Rigorous analysis is provided in the Appendix.

Weaknesses

- A smoother transition toward (FTXL-D) in lines 182–186 is desired. In a game, $x$ is the response and $y$ is an aggregated payoff. For (HBVF), the interpretation is for a moving object. - What is the correspondence of $x$ and $y$ of a game in (HBVF)? Do we correspond $x$ (position) in (HBVF) to $y$ (payoff) in a game? If so, what corresponds to the response ($x$ in a game) in (HBVF)? Is it the control $w$ (my naive guess)? - These analogs, with the aid of a clearer transition, would help us to better understand the first paragraph of Section 3.2. Maybe the response x in a game and the position x in (HBVF) introduces a bit of confusion, too? - There are two sets of notations, one for the continuous-time regime and the other for the discrete-time regime. A small table to distinguish them might help the readability. - Although in general, the submission is easy to read, a minor aspect of improving readability could be explaining the meaning $\dot{y}$ (lines 192 and 196) and $\ddot{z}$ (line 486), which could ease the reading barrier for people who are more comfortable with discrete-time analysis.

Questions

- In line 245, the friction parameter is set to zero ($r=0$) according to the discussion of Example 3. This setting, by direct substitution, would cancel out the desired effect of the momentum in (FTXL-D) in line 186 or (4) in line 197. However, in (FTXL) (line 258), the momentum appears to play an important role in the leftmost equality $y_{n+1} = y_n + \gamma p_n$. Did I misunderstand something? - In Figure 1, why does FTXL perform poorly in the early rounds? Is it due to the constants in the bound or is there another reason?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer qtGD8/10 · confidence 3/52024-07-10

Summary

The paper proposes a momentum-based follow-the-regularized-leader (FTRL) type algorithm for finding the Nash equilibrium in finite games. The paper first investigates the continuous second order ODE and a concrete game, and then devises the appropriate FTRL scheme. The convergence rate for the proposed algorithm in all three information settings of reward is quadratic (O( exp(-T^2))).

Strengths

1. A rich collection of results on momentum methods for finite-games. The paper has three sets of theoretical results. First, convergence results for the second-order ODE dynamics with vanishing friction (Theorem 1) and non-vanishing friction (Theorem B.1), both under the logit best response. Second, an analysis of a momentum FTRL under full information on a concrete one-player game with vanishing (Prop C.1) and non-vanishing friction (Prop C.2). And finally, analysis momentum FTRL with three kinds of info feedback. 2. The paper discovered an important phenomenon that, in the continuous second order ODE, "the friction term hinders convergence". This claim is backed by a concrete analysis of momentum FTRL on a concrete game. In my opinion this phenomenon is interesting to the game learning community. 3. The writing is flowing and smooth.

Weaknesses

See question.

Questions

Minor comments 1. Are lower bounds of learning a finite game with the three kinds feedbacks known? If so, it would be great to provide a survey. 2. In line 305 the phrasing "second-order in space" is a bit confusing. In my opinion, it would be better to say "using Hessian information of relevant functions". Typo 1. At Eq FTRL, near line 131, the second equation should be x_i,n = Q_i (y_i, n).

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

4

Limitations

The paper provides a extensive discussion of momentum method for finite game solving, and does not have major limitations in my opnion.

Reviewer yttq5/10 · confidence 3/52024-07-20

Summary

This paper primarily focuses on introducing a Nesterov's accelerated gradient (NAG) algorithm for online learning in games. Initially, the author shows that a continuous-time version of NAG converges to a strict Nash equilibrium at a rate that is quadratic. This rate of convergence is notably faster than that of standard FTRL algorithms. Following this, the paper demonstrates that the convergence rate of NAG is preserved even in a bandit feedback setting. The experimental results confirm that the proposed algorithm converges to an equilibrium more quickly than a FTRL algorithm.

Strengths

* Despite the significant importance of introducing Nesterov's accelerated gradient, a method highly esteemed in convex optimization problems, into the context of learning in games, such an approach has not been pursued until now. * The intuition behind the proposed method is thoroughly explained. Furthermore, the proposed algorithm appears to be straightforward to implement. * The derived convergence rates are confirmed to be tight by deriving the exact convergence rate in a simple instance.

Weaknesses

My primary concern is that the theoretical results are only applicable to games with a strict Nash equilibrium. This is somewhat problematic as even in a simple game like rock-paper-scissors, a strict Nash equilibrium does not exist. Moreover, the convergence in games that do not have a strict equilibrium has been the subject of extensive research. Therefore, I'm wondering how efficient NAG algorithms are in two-player zero-sum games and monotone games. Furthermore, Theorems 1, 2, 3, and 4 hinge on the presumption that the initial strategy is in close proximity to an equilibrium. Could you provide some insight into how close the initial point needs to be to the equilibrium?

Questions

My major concerns and questions are stated in Weaknesses. I also have the following additional questions: * Why is the update rule of $y_{i,n+1}=y_{i,n}+\gamma p_{i,n+1}$ in (FTXL) derived? Looking at the original NAG's update rule in (NAG), it seems more natural that $y_{i,n+1}=y_{i,n}+\gamma (p_{i,n+1}-p_{i,n})$. * Is it possible to recover FTRL algorithms by setting $r$ appropriately in FTXL?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

I see no negative societal impacts that need to be addressed.

Reviewer sdz42024-08-08

I thank the authors for the clarification and would like to maintain my score. I support the acceptance of the paper.

Authorsrebuttal2024-08-12

Thank you

Thank you for your time and your support! Kind regards, The authors

Area Chair y6MW2024-08-09

Small question

Hi, Can the authors comment on whether there are any connections between their FTXL method and the "linear coupling" method of Allen-Zhu and Orecchia [1], which also aims to combine mirror descent/FTRL and acceleration? The technique I mention is not analyzed in the context of games, but I am curious of there are connections nonetheless. Thank you! [1] Linear Coupling: An Ultimate Unification of Gradient and Mirror Descent

Authorsrebuttal2024-08-12

Dear AC, > Can the authors comment on whether there are any connections between their FTXL method and the "linear coupling" method of Allen-Zhu and Orecchia [1], which also aims to combine mirror descent/FTRL and acceleration? The technique I mention is not analyzed in the context of games, but I am curious of there are connections nonetheless. Excellent question, thanks for bringing it up! It's a complex issue, so we're giving two answers, one shorter and one (much) longer. --- ### **Short answer** The linear coupling approach of [1] is related to the momentum-based approach of Su, Boyd and Candès which forms the basis of (FTXL), but the two are not equivalent. In essence, they are both different approaches to acceleration – and, in turn, both different than the original "estimate sequences" approach of Nesterov – and we do not see a way of directly applying the techniques of [1] to our setting. --- ### **Long answer** Because [1] is not taking a momentum-based approach (see e.g., footnote 5 in p. 4 of the arxiv version of the paper), it is difficult to accurately translate the coupling approach of [1] to our setting and provide a direct comparison between the two methods. One of the main reasons for this is that [1] is essentially using two step-sizes: the first is taken equal to the inverse Lipschitz modulus of the function being minimized and is used to take a gradient step; the second step-size sequence is much more aggressive, and it is used to generate an ancillary, exploration sequence which "scouts ahead". These two sequences are then "coupled" with a mixing coefficient which plays a role "similar" - but *not equivalent* - to the friction coefficient in the (HBVF) formulation of (NAG) by Su, Boyd, and Candès. The above is the best high-level description and analogy we can make between the coupling approach of [1] and the momentum-driven analysis of Su, Boyd, and Candès and/or momentum analysis in Nesterov's 2004 textbook (which is itself quite different from Nesterov's original 1983 paper). At a low level (and omitting certain technical details and distinctions that are not central to this discussion), the linear coupling approach of [1] applied to our setting would correspond to the update scheme: \begin{align} x_t &= Q(y_t) \\\\ w_t &= \lambda_t z_t + (1-\lambda_t) x_t \\\\ y_{t+1} &= y_t + (1-\lambda_t) \eta_t \hat v_t \\\\ z_{t+1} &= \lambda_t z_t + (1-\lambda_t) x_{t+1} \end{align} with $\hat v_t$ obtained by querying a first-order oracle at $w_t$ – that is, $\hat v_t$ is an estimate, possibly imperfect, of $v(w_t)$. The first and third lines of this update scheme are similar to the corresponding update structure of (FTXL). However, whereas (FTXL) builds momentum by the aggregation of gradient information via the momentum variables $p_t$, the linear coupling method above achieves acceleration through the coupling of the sequences $w_t$, $z_t$ and $x_t$, and by taking an increasing step-size sequence $\eta_t$ that grows roughly as $\Theta(t)$, and a mixing coefficient $\lambda_t$ that evolves as $\lambda_t = 1 - 1/(L\eta_t)$, where $L$ is the Lipschitz modulus of $v$. Beyond this comparison, we cannot provide a term-by-term correspondence between the momentum-based and coupling-based approaches, because the two methors are not equivalent (even though they give the same value convergence rates). In particular, we do not see a way of linking the parameters $\eta_t$ and $\lambda_t$ of the coupling approach to the friction and step-size parameters of the momentum approach. In the context of convex minimization problems, the coupling-based approach of [1] is more amenable to a regret-based analysis – this is the "unification" that [1] referred to – while the momentum-based approach of Su, Boyd, and Candès facilitates a Lyapunov-based analysis. Ultimately, these are all different - though, of course, equally valid - approaches to acceleration. From a game-theoretic standpoint, the momentum-based approach seems to be more fruitful and easier to implement, but studying the linear coupling approach of [1] could also be very relevant. --- We will be happy to include a version of the above discussion at the first possible revision opportunity. In the meantime, please let us know if you have any follow-up questions on the above, and thanks again for your question and for handling our submission! Kind regards, The authors

Authorsrebuttal2024-08-14

Thank you again for your time, input, and positive evaluation! Kind regards, The authors

Authorsrebuttal2024-08-14

Thank you again for your time and input. Kind regards, The authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC