We Rebuilt a 108-Card Game AI Three Times—From Handcrafted Search to PPO

Description and use

Building an AI for GuanDan required addressing combinatorial actions, imperfect information, and delayed credit. DanKS implements a two‑stage pipeline—candidate retrieval and ranking—with successive versions moving from handcrafted search to learned selection and finally to a memory‑aware actor‑critic trained with PPO.

Problem

Designing an AI for GuanDan, a four-player partnership card game with 108 cards, required handling combinatorial actions, imperfect information, and delayed credit while keeping the action space manageable.

Outcome

The project demonstrates a clear evolution from handcrafted search to learned candidate selection and finally to a memory‑aware actor‑critic trained with PPO, preserving each generation for research reuse.

Workflow

• 1. Decompose the hand and generate a bounded set of legal move candidates.
• 2. Rank the top‑K candidates using a policy that considers visible state, public history, seat context, and structural features.
• 3. In V3, augment the policy with card memory, coverage features, team‑belief signals, and train with PPO and generalized advantage estimation to handle delayed credit.

Evidence

We built it three times—moving from handcrafted structural search, to learned candidate selection, and finally to a memory-aware actor-critic trained with PPO.

Source

The retrieval stage is not just a speed optimization. Two legal moves may look similar on the table but consume very different parts of the remaining hand.

Source