Taiwan Big Two AI: An Intelligent Card Game System Combining Heuristic Algorithms and Large Language Models

Taiwan Big Two AI (taiwan-big-two-ai) is an intelligent card game system that combines traditional heuristic algorithms with a multi-role research engine based on large language models (LLMs). It innovatively enables autonomous strategic game analysis and decision-making, providing a new paradigm for game AI.

Big Two, also known as "Pai Lao Er" or "Cho Dai Di", originated in China and is popular in East and Southeast Asia. It uses a 52-card deck (without jokers) and involves 3-4 players. The core mechanism is card-playing suppression (card types include single cards, pairs, straights, etc., with straight flushes being the highest). The first player to play all their cards wins. Its strategic depth lies in hand management, judging opponents' card types, retaining strong cards, and choosing the right time to attack, making it an ideal scenario for game AI research.

The project adopts a dual-engine architecture:

Heuristic AI Engine

Responsible for deterministic decisions and basic strategies, including card type recognition and validity checks, basic card-playing strategies, hand evaluation, and opponent modeling. Its advantages are fast response, interpretability, and stability, but it struggles with long-term strategies and psychological games.

LLM Multi-Role Research Engine

Includes four roles: aggressive (early risk-taking), conservative (risk control), balanced (balanced offense and defense), and analytical (opponent pattern recognition). The way of thinking is defined through system prompts, which is the core innovation of the project.

During key decisions, the engine consults the opinions of multiple roles (with reasons attached), and the meta-decision layer makes the final decision by synthesizing factors such as game state, opponent history, and remaining cards. Benefits:

1. Strategic diversity, avoiding single bias
2. Interpretable decisions, facilitating analysis and debugging
3. Dynamically adjust role weights to improve learning ability

Game State Representation

Structured description of the current table state, remaining cards of players, historical card-playing records, and hand details, balancing information volume and LLM context limits.

Prompt Engineering Strategies

1. System prompts define roles
2. Few-shot examples guide rules and strategies
3. Chain-of-thought requires reasoning process

Comparison with MCTS

Advantages: No training required, interpretable, flexible, low cost; Limitations: Decision quality is limited by the LLM's training knowledge.

• Game Developers: Provide new ideas for NPC design, balancing playability and strategic changes
• AI Researchers: A test platform for imperfect information decision-making, simulating real-world scenarios
• Ordinary Players: Play against AIs of different styles to improve game skills

1. Online Learning: Continuously improve strategies from human vs AI matches
2. Multi-Language Support: Adapt to card game variants in other regions
3. Visual Interface: Display the AI's decision-making process
4. Tournament Mode: Evaluate performance through AI vs AI competitions

taiwan-big-two-ai represents a new direction in game AI—combining heuristic rules with LLM reasoning to build intelligent opponents with "personality" and "thinking processes". This hybrid architecture can be applied to a wider range of strategic games and decision-making scenarios.