Last-Iterate Convergence Properties of Regret-Matching Algorithms in Games

Authors: Yang Cai, Gabriele Farina, Julien Grand-Clément, Christian Kroer, Chung-Wei Lee, Haipeng Luo, Weiqiang Zheng

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We start by showing numerically that several variants used in practice, such as RM+, predictive RM+ and alternating RM+, all lack last-iterate convergence guarantees even on a simple 3 3 matrix game. We then prove that recent variants of these algorithms based on a smoothing technique, extragradient RM+ and smooth Predictive RM+, enjoy asymptotic last-iterate convergence (without a rate), 1/ t best-iterate convergence, and when combined with restarting, linear-rate last-iterate convergence. 4. Experiments. We validate the last-iterate convergence of Ex RM+ and SPRM+ (including their restarted variants that we propose) numerically on four instances of matrix games, including Kuhn poker and Goofspiel.
Researcher Affiliation Academia Yang Cai Yale EMAIL Gabriele Farina MIT EMAIL Julien Grand-Clément HEC Paris EMAIL Christian Kroer Columbia EMAIL Chung-Wei Lee USC EMAIL Haipeng Luo USC EMAIL Weiqiang Zheng Yale EMAIL
Pseudocode Yes Algorithm 1 Regret Matching+ (RM+)... Algorithm 2 Predictive RM+ (PRM+)... Algorithm 3 Extragradient RM+ (Ex RM+)... Algorithm 4 Smooth PRM+ (SPRM+)... Algorithm 5 Restarting Ex RM+ (RS-Ex RM+)... Algorithm 6 Restarting SPRM+ (RS-SPRM+)... Algorithm 7 Alternating RM+ (alt. RM+)... Algorithm 8 Alternating Predictive RM+ (alt. PRM+)... Algorithm 9 Extragradient algorithm (EG)... Algorithm 10 Optimistic gradient descent (OG)
Open Source Code No No explicit statement or link for open-source code release is provided in the paper. The paper mentions the algorithms were coded in Python 3.8.8 for the numerical experiments, but does not offer access to the implementation.
Open Datasets Yes We use the 3x3 matrix game instance from Section 3, the normal-form representations of Kuhn poker and Goofspiel, as well as 25 random matrix games of size (d1, d2) = (10, 15)... Kuhn poker is a widely used benchmark game introduced by Kuhn (1950)... We use an imperfect-information variant of the standard benchmark game introduced by Ross (1971)... The coefficients of the matrices are normally distributed with mean 0 and variance 1 and are sampled using the numpy.random.normal function from the numpy Python package (Harris et al., 2020).
Dataset Splits No The paper analyzes algorithms for solving matrix games and mentions specific game instances (3x3 matrix game, Kuhn poker, Goofspiel, random matrix games) but does not describe any training/test/validation dataset splits, as these are typically not applicable to the problem of solving a single game instance. Evaluation is performed on the full game definitions.
Hardware Specification Yes All the algorithms were coded with Python 3.8.8, and we ran our numerical experiments on a laptop with 2.2 GHz Intel Core i7 and 8 GB of RAM.
Software Dependencies Yes All the algorithms were coded with Python 3.8.8... The coefficients of the matrices are normally distributed with mean 0 and variance 1 and are sampled using the numpy.random.normal function from the numpy Python package (Harris et al., 2020).
Experiment Setup Yes For the algorithms requiring a stepsize (Ex RM+, SPRM+, EG and OG), we choose the best stepsizes η {1, 10 1, 10 2, 10 3, 10 4}. We initialize all algorithms at ((1/d1)1d1, (1/d2)1d2). We present our results when choosing a stepsize η = 0.05 in Figure 7 and Figure 8.