There's no Human in Charge: Playing Chef's Hat with a Large Language Model Based Agent

Generative artificial intelligence (AI), especially large language models (LLMs), has greatly improved the creation of digital content. This study examines the use of an LLM, specifically Llama-3-8B, in playing Chef's Hat, a multiplayer card game. Traditionally, reinforcement learning (RL) agents used in this game do not perform well against human players. Our goal was to see if the LLM could play the game effectively without additional training. We integrated the Llama-3-8B model into the Chef's Hat game and compared its performance with three existing agents: Random, DQL (Deep Q-Learning), and PPO (Proximal Policy Optimization). Using special instructions, we guided the LLM on how to play the game and make decisions. The results showed that the LLM-based agent performed as well as or better than the RL agents, making smart and strategic moves based on the game rules and current situation. This study highlights the adaptability and potential of LLMs in various competitive settings, suggesting they could be used as effective autonomous agents in human-robot interactions. Future work should explore combining LLMs with other types of agents, testing them in different games, and improving the instructions given to them.