Entropy regularized subgame solving sequential Bayesian games with public actions

Sobhan Mohammadpour, Samuel Sokota, Brandon Gary Kaplowitz, J Zico Kolter, Noam Brown, Gabriele Farina

Keywords: Poker, Game theory, reinforcement learning, search, GPUs, dynamic programming

TL;DR: We introduce generic primitive needed for solving imperfect information games

Abstract: Subgame solving is central to scaling equilibrium approximation in large imperfect-information games. We introduce a small set of reusable GPU primitives for high-performance, entropy-regularized subgame solving in public-action Bayesian games (PBGs). Instantiated in Liar’s Dice and Heads-up Hold’em, our approach achieves one to two orders of magnitude speedups and substantially lower memory usage compared to baselines.

Submission Number: 28