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

Описание и применение

Создание ИИ для игры GuanDan потребовало решения проблем с комбинаторными действиями, неполной информацией и отложенной оценкой. В проекте DanKS реализован двухступенчатый пайплайн: генерация кандидатов и их ранжирование, с постепенным переходом от ручного поиска к обученному селектору и, наконец, к памяти‑осознанному актор‑критическому обучению с PPO.

Задача

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.

Результат

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.

Рабочий процесс

• 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.

Доказательства

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.

Источник

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.

Источник