FPDou: Mastering DouDizhu with Fictitious Play
Combined Fictitious Play with deep reinforcement learning to master DouDizhu, a three-player card game where two farmers cooperate against the landlord.
Overview
DouDizhu is one of the most popular card games in China, with an estimated 600 million players worldwide. It is a three-player game where two farmers cooperate against the landlord. By treating the two farmers as a single player, the game can be modeled as a two-player zero-sum setting, enabling the application of game-theoretic algorithms like Fictitious Play.
Methodology
FPDou is a state-of-the-art DouDizhu AI agent built on Generalized Weakened Fictitious Play (GWFP) and deep reinforcement learning for the large-scale three-player imperfect-information card game DouDizhu. It eliminates multi-agent non-stationarity via alternating on/off-policy self-play, unifies two peasants into a regularized cooperative team under a two-player zero-sum formulation, adopts a tiny handcrafted-feature-free distributional CNN backbone, and achieves top-ranked performance across mainstream DouDizhu benchmark platforms.
Results
FPDou achieved state-of-the-art performance on DouDizhu, outperforming both open-source and closed-source models. It ranked first among 452 bots on the Botzone platform.
