On the Convergence of No-Regret Learning Dynamics in Time-Varying Games

Authors: Ioannis Anagnostides, Ioannis Panageas, Gabriele Farina, Tuomas Sandholm

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, although the focus of this paper is theoretical, in this section we provide some illustrative experimental examples. In particular, Appendix B.1 contains experiments on time-varying potential games, while Appendix B.2 focuses on time-varying (two-player) zero-sum games.
Researcher Affiliation Collaboration Ioannis Anagnostides Carnegie Mellon University ianagnos@cs.cmu.edu Ioannis Panageas University of California Irvine ipanagea@ics.uci.edu Gabriele Farina MIT gfarina@mit.edu Tuomas Sandholm Carnegie Mellon University Strategic Machine, Inc. Strategy Robot, Inc. Optimized Markets, Inc. sandholm@cs.cmu.edu
Pseudocode No No
Open Source Code No No
Open Datasets No No
Dataset Splits No No
Hardware Specification No No
Software Dependencies No No
Experiment Setup Yes In our first experiment, we first sampled two matrices A, P Rdx dy, where dx = dy = 1000. Then, we defined each payoff matrix as A(t) := A(t 1) + Pt α for t 1, where A(0) := A. Here, α > 0 is a parameter that controls the variation of the payoff matrices. In this time-varying setup, we let each player employ (online) GD with learning rate η := 0.1.