Neura Market
Paper logo

Paper

Free

GPT-4 agent for imperfect information games with Theory of Mind

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About Paper

Suspicion-Agent is an innovative AI agent that leverages GPT-4's advanced reasoning and knowledge retrieval to play imperfect information games (e.g., card games with hidden information). It employs Theory of Mind (ToM) to understand opponents' perspectives and intentionally influence their behavior. The agent uses prompt engineering to perform different functions and a planning strategy that adapts gameplay to various opponents, requiring only the game rules and observations as input. Without any specialized training or examples, Suspicion-Agent demonstrates strong performance in games like Leduc Hold'em, potentially outperforming traditional algorithms designed for imperfect information games. The paper makes game-related data publicly available.

Key Features

Leverages GPT-4 for knowledge retrieval and reasoning
High-order Theory of Mind (ToM) capacity
Prompt engineering for different functional modules
Adaptable planning strategy for various opponents
No specialized training or examples required
Only needs game rules and observations as input

Pros & Cons

Pros
  • Demonstrates strong high-order Theory of Mind, enabling understanding and influencing of opponents
  • Outperforms traditional imperfect information game algorithms in Leduc Hold'em
  • No need for specialized training or game-specific examples
  • Adaptable gameplay style through prompt engineering and planning strategy

Best For

Playing imperfect information card games like Leduc Hold'emAI research in decision-making under uncertaintyTesting Theory of Mind capabilities in LLMsDeveloping agents that adapt to opponent behavior without retraining

FAQ

What is Suspicion-Agent?
Suspicion-Agent is an agent that uses GPT-4's capabilities to play imperfect information games, leveraging Theory of Mind to understand and influence opponents.
What kind of games can Suspicion-Agent play?
It is demonstrated on three different imperfect information games, with quantitative evaluation in Leduc Hold'em.
Does Suspicion-Agent require training or examples?
No, it requires no specialized training or examples; it operates using only game rules and observations fed via prompt engineering.
Is the code or data publicly available?
The paper states that game-related data is made publicly available to encourage further research.