Why classic Monte Carlo still wins at DouDizhu

What it does

DouZero is a reinforcement-learning framework built to play DouDizhu, the most popular card game in China. It trains agents from scratch via self-play to compete as either the Landlord or one of the two Peasants. The catch is that legal moves change drastically every turn, and the action space spans roughly 10^4 possibilities—far larger than what most standard RL algorithms comfortably handle.

The interesting bit

Rather than chasing the latest policy-gradient fashion, the authors revived classic Monte Carlo methods, augmenting them with deep neural networks, action encoding, and parallel actors. Starting from scratch on a single server with four GPUs, the system reached first place on the Botzone leaderboard among 344 agents after days of training.

Key highlights

• Tackles a three-player, imperfect-information game mixing competition with implicit cooperation (two Peasants versus one Landlord).
• Uses Deep Monte Carlo (DMC) with action encoding to wrangle an action space of approximately 10^4 legal moves.
• Ships pretrained models for two reward objectives: Average Difference Points (DouZero-ADP) and Winning Percentage (DouZero-WP).
• Offers an online demo, a local demo option, and a Google Colab notebook for experimentation.
• Community forks have added ResNet backbones and full-auto gameplay integrations.

Caveats

• Training expects GPUs; Windows users are restricted to CPU-only actors.
• CPU evaluation works, but the README notes GPU evaluation is “much faster.”

Verdict

Grab it if you research imperfect-information games, large action spaces, or mixed cooperative-competitive RL. Pass if you want a polished consumer card-game bot—this is a training framework and research artifact.