Calix-L/DanKS

What it solves

DanKS is an AI system designed to master GuanDan, a complex four-player partnership card game. It addresses the challenge of managing a massive, combinatorial action space and the need for long-horizon credit assignment, where a move's value may only become apparent several turns later.

How it works

The system uses a pipeline that converts the game state into a compact policy decision through four main steps:

1. State Encoding: It processes the visible hand, action history, legal moves, and game context.
2. Structured Retrieval: It uses a budgeted decomposition search to find a representative set of candidate plays, summarizing their structure (length, sequences, suits, etc.).
3. Scoring: A shared encoder combines state and candidate features, allowing an actor to rank the candidates and a critic to estimate the state's value.
4. Self-Play Learning: It uses Proximal Policy Optimization (PPO) with Generalized Advantage Estimation (GAE) to learn from game trajectories, improving the AI's ability to favor actions that lead to long-term success.

Who it’s for

This project is for AI researchers and developers interested in game AI, reinforcement learning, and the application of PPO to games with large action spaces.

Highlights

• Three-Generation Evolution: The codebase includes V1 (structural retrieval), V2 (learned selection), and V3 (memory-aware policy learning with PPO).
• State-of-the-Art Performance: Achieves leading results against both rule-based and learning-based GuanDan baselines.
• Comprehensive Pipeline: Includes a full GuanDan rules engine, legal-action generation, and training infrastructure.
• Hardware Flexibility: Supports CPU, NVIDIA CUDA, and Ascend NPU acceleration, with optional C++ kernels for faster retrieval.