Calix-L/DanKS

What it solves

DanKS is an AI agent designed to play GuanDan, a complex partnership card game characterized by hidden information and long-term strategic planning. It solves the challenge of making moves that aren't just locally optimal but preserve future options and coordinate with a teammate to win the hand.

How it works

The system uses a multi-stage pipeline to make decisions:

1. State Encoding: It encodes the current hand, public history, legal actions, and team context.
2. Structure-Aware Retrieval: Instead of evaluating every possible move, it identifies a compact set of meaningful candidate actions by examining the combinations left in the residual hand.
3. Policy Selection: A state- and candidate-conditioned Actor-Critic network ranks these candidates to select the best move.
4. Reinforcement Learning: The agent is trained using Proximal Policy Optimization (PPO) and Generalized Advantage Estimation (GAE) to connect current decisions to long-term outcomes.

Later versions (V3Pro) add inference refinements such as asset protection, equivalent-play rules, and verified endgame search for positions with 16 or fewer cards remaining.

Who it’s for

• AI researchers studying partnership games and hidden information.
• Developers interested in implementing PPO and Actor-Critic architectures for game AI.
• Game developers looking for an open-source web service framework to host card games.

Highlights

• Evolutionary AI Generations: Progresses from simple NumPy selectors (V1) to memory-aware neural networks trained with PPO (V3).
• Integrated Web Service: Includes two generations of open-source web game services (KSPlay) for deploying the AI in a playable browser environment.
• Human Data Integration: Provides access to the KSCB GuanDan dataset containing nearly 900 complete matches for studying human decision-making.
• Endgame Optimization: V3Pro implements specialized search and refinement rules for the final stages of the game.