Independent Policy Gradient Methods for Competitive Reinforcement Learning

Authors: Constantinos Daskalakis, Dylan J. Foster, Noah Golowich

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Figure 1: Figures (a) and (b) display plots for one ratio game, and Figures (c) and (d) display plots for another; the games matrices are specified in Appendix D.1. Figures (a) and (c) plot the quantity sign( F(z),z z ) for z 2 2, parameterized as z = (x,1 x,y,1 y); yellow denotes negative and purple denotes positive. The red dot denotes the equilibrium z . Figures (b) and (d) plot convergence of extragradient with learning rate 0.01, initialized at z0 = (1,0,1,0); note that z0 is inside the region in which the MVI does not hold for each problem. The blue line plots the primal-dual gap maxy V (x(i),y ) minx V (x ,y(i)) and the orange line plots the primal gap maxy V (x(i),y ) V (x ,y ).
Researcher Affiliation Academia Constantinos Daskalakis costis@csail.mit.edu Dylan J. Foster dylanf@mit.edu Noah Golowich nzg@mit.edu Massachusetts Institute of Technology
Pseudocode Yes Algorithm 1: Two-timescale Stochastic Gradient Descent-A scent (SGD-A) for Two-player Zero-sum Games
Open Source Code No The paper does not contain any explicit statement about releasing source code for the methodology or provide a link to a code repository.
Open Datasets No The paper defines and analyzes specific game setups ('ratio game', 'stochastic game') directly within its text (e.g., equation (11) and Appendix D.1). It does not use or provide concrete access information for any external, publicly available datasets.
Dataset Splits No The paper does not provide specific details about training, validation, or test dataset splits. The experiments are conducted on theoretically defined game structures.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes Figures (b) and (d) plot convergence of extragradient with learning rate 0.01, initialized at z0 = (1,0,1,0);