Comparing AI and human decision-making mechanisms in daily collaborative experiments.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Game Theory

Background:

• The increasing capabilities of artificial intelligence (AI), particularly large language models (LLMs), raise questions about their potential to replace human decision-making.
• Understanding AI's performance in complex, collaborative environments is crucial for assessing its real-world applicability.

Purpose of the Study:

• To compare the decision-making performance of humans, LLMs, and reinforcement learning (RL) agents in a simulated multi-day commute game.
• To analyze AI's ability to learn, converge, and collaborate in a dynamic, interdependent decision-making scenario.

Main Methods:

• A multi-day commute decision-making game was designed to simulate collaborative choices with interdependent outcomes.
• Performance metrics included system-level results, convergence rates, individual decision dynamics, and decision mechanisms for humans, LLMs, and RL agents.

Main Results:

• LLMs demonstrated human-like learning from historical data and achieved convergence in individual commute decisions.
• In collaborative settings, LLMs exhibited limitations in perceiving others' actions and integrating physical knowledge.

Conclusions:

• LLMs show promise in mimicking human learning for individual decision tasks but require significant advancements for effective multi-agent collaboration.
• Further research is needed to enhance LLMs' social perception and grounding in physical realities for complex group decision-making.