Taiwan Big Two AI: An Intelligent Poker Battle System Integrating Heuristic AI and Large Language Models

In the field of artificial intelligence research, games are important testbeds for evaluating AI capabilities. However, most studies focus on mainstream Western games, while traditional regional characteristic games are often overlooked. The Taiwan Big Two AI project fills this gap by applying modern AI technology to Taiwan's popular Big Two poker game, innovatively integrating traditional heuristic algorithms with the reasoning capabilities of large language models to build an intelligent battle system capable of autonomous strategic analysis.

Big Two is a widely popular poker game in Asia. Its rules are simple but have rich strategic depth: four players use a deck of cards without jokers, with various card types, and players must play a higher-ranked card than the previous one; the first to play all their cards wins. Its strategic complexity lies in hand management, information reasoning, dynamic game theory, psychological confrontation, etc., making it a 'solvable but non-trivial' AI testing platform.

The core of the system is a dual-engine architecture:

1. Heuristic AI Engine: Based on Monte Carlo Tree Search (MCTS), combined with evaluation functions considering factors like hand strength and card-playing efficiency, as well as optimization strategies for common card types. It features fast speed and strong interpretability.
2. Multi-Role LLM Research Engine: Includes roles such as tactical analyst and strategic planner. Through parallel analysis, opinion integration, and in-depth discussion for collaborative decision-making, it has autonomous research capabilities like hypothesis generation and evidence collection. The dual engines are integrated via intelligent routing: heuristic engine for simple scenarios, LLM activated for complex scenarios, and parallel comprehensive evaluation for key decisions.

Technical highlights include:

• Unified game state representation: structured data, natural language description, and visual display;
• LLM prompt engineering: clear role definition, context injection, few-shot reasoning guidance, and output constraints;
• Evaluation and training framework: self-play data generation, human-machine battle evaluation for practical testing, strategy analysis and improvement, and A/B testing for performance comparison.

The project has wide application value:

• Education: Helps learn strategies, understand high-level thinking, and demonstrate the application of game theory;
• Research: Explores the boundaries of LLM in incomplete information games, the effectiveness of multi-role architecture, and provides references for other game AIs;
• Expansion: Can be applied to other games like Dou Di Zhu, business decision analysis, educational tutoring, etc.

Taiwan Big Two AI demonstrates a new trend in AI technology: integrating different AI paradigms (heuristics and LLM) to build a powerful and interpretable system by complementing their advantages. For game AI researchers and LLM application developers, it is an innovative case worth in-depth study.