Case study - Laizi Dou Dizhu AI for a complex action space
An extension of research and open-source DouZero ideas to wildcard rules, with redesigned legal-action generation, state representation, action encoding, and policy evaluation.
- Client
- Anonymized game AI startup
- Service
- Game AI / Reinforcement learning / Decision systems
- Imperfect information
- Self-play
- Action space
- Wildcard rules
- Policy evaluation
Why the variant is difficult
Dou Dizhu already combines three-player competition, cooperation, imperfect information, and a large dynamic action space. Wildcards multiply possible combinations and make duplicate or invalid actions much more likely.
The model must preserve both the physical card and its substituted meaning while reasoning about future hand structure, teammate cooperation, and opponent pressure.
Solution
After studying the DouZero paper and implementation, we retained its strengths in large action spaces and self-play while rebuilding the rule-dependent modules.
The work covered wildcard hand enumeration, legal-action generation, normalization and deduplication, state encoding, action features, inference integration, and policy evaluation.
Model iteration
Automated self-play and real match analysis were used to identify weaknesses in bidding, splitting hands, preserving wildcards, suppressing opponents, supporting teammates, and end-game play.
These findings fed back into rule generation, state representation, training, and evaluation rather than relying on a static benchmark alone.
Outcome
The delivered system makes stable decisions under wildcard rules and reached a level close to strong human players in the client’s production version.
The approach also applies to other constrained decision problems with large state spaces, dynamic legal actions, delayed rewards, and multi-agent interaction.