Fast Last-Iterate Convergence of Learning in Games Requires Forgetful Algorithms

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

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Figure 1: Comparison of the dynamics produced by three variants of OFTRL with different regularizers (negative entropy, logarithmic regularizer, and squared Euclidean norm) and OGDA in the same game Aδ defined in (2) for δ := 10 2. The bottom row shows the duality gap achieved by the last iterates. The OFTRL variants exhibit poor performance due to their lack of forgetfulness, while OGDA converges quickly to the Nash equilibrium. Figure 2: Performance of OMWU on the game Aδ defined in eq. (2) for three choices of δ. In all plots, the learning rate was set to η = 0.1. As predicted by our analysis, the length of the flat region between iteration T1 (red star) and T2 (blue dot) scales inversely proportionally with δ. Our simulations were done within a few hours on an average consumer laptop.
Researcher Affiliation Academia Yang Cai Yale yang.cai@yale.edu Gabriele Farina MIT gfarina@mit.edu Julien Grand-Clément HEC Paris grand-clement@hec.fr Christian Kroer Columbia ck2945@columbia.edu Chung-Wei Lee USC leechung@usc.edu Haipeng Luo USC haipengl@usc.edu Weiqiang Zheng Yale weiqiang.zheng@yale.edu
Pseudocode No The paper provides mathematical definitions and update rules for algorithms (OOMD, OFTRL) but does not present them in a structured pseudocode or algorithm block format.
Open Source Code Yes Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: See Python code in the supplementary materials.
Open Datasets No The paper defines a specific 2x2 matrix game Aδ for its analysis and simulations ('The game s loss matrix Aδ is parameterized by δ (0, 1 2) and is defined as follows: Aδ = 2 + δ 1 2 0 1'). This is a synthetic instance created by the authors, not a publicly available dataset with external access information or citation.
Dataset Splits No The paper conducts simulations on a specific game instance and analyzes dynamics. It does not mention traditional training, validation, or test dataset splits.
Hardware Specification No Our simulations were done within a few hours on an average consumer laptop. The term 'average consumer laptop' is not a specific hardware detail (e.g., CPU/GPU model, memory, or clock speed).
Software Dependencies No The paper does not provide specific version numbers for any ancillary software dependencies or libraries used in the experiments. It mentions 'Python code in the supplementary materials' but without version details.
Experiment Setup Yes In all plots, the learning rate was set to η = 0.1. In this section we present our numerical results when OFTRL and OOMD are instantiated with adaptive stepsize [Duchi et al., 2011]: ηt = 1/ q ϵ + Pt 1 k=1 ℓk 2 k with some constant ϵ > 0. We present our numerical experiments in Figure 4, where we choose ϵ = 0.1.