Reproduce “DREAM: Deep Regret minimization with Advantage baselines and Model-free learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eric Steinberger, Adam Lerer, Noam Brown
DREAM enables model-free poker agents to minimize regret without a perfect simulator. It samples one trajectory per decision and uses learned Q-baselines to reduce feedback variance. In Leduc Hold’em, DREAM achieves two orders of magnitude better sample complexity than NFSP and matches SD-CFR. Its equilibrium guarantee covers historical average policies in two-player zero-sum games; FHP exploitability is an approximate lower bound.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.