Competition is vital in shaping all aspects of human society, including economics, social structures, and technology. Traditionally, studying competition has been reliant on empirical research, which is limited due to issues with data accessibility and a lack of micro-level insights. An alternative approach, agent-based modeling (ABM), advanced from rule-based to machine learning-based agents to overcome these drawbacks. However, these methods still struggle to accurately simulate intricate human behavior. This changed with the advent of Large Language Models (LLMs), which has enabled the creation of autonomous agents for social simulations.
Despite these developments, there is a gap in research on competition dynamics using LLM-based agents and this lack of understanding hinders our comprehension of competition across various fields. While empirical studies on competition have uncovered critical findings, they face limitations in controlling variables and collecting comprehensive data. The latest advancements in LLM-empowered-ABM have the potential to transform social simulations. However, existing studies tend to fall short in simulating complex competitive environments, thoroughly analyzing competitive behaviors, and tracking system evolution. Therefore, there is a need for more comprehensive studies using LLM-based agent simulations.
CompeteAI is a comprehensive framework developed by researchers from the University of Science and Technology of China, Microsoft Research, William & Mary, Georgia Institute of Technology, and Carnegie Mellon University to study competition dynamics between LLM-based agents. Using GPT-4, the researchers created a virtual town simulation with restaurant and customer agents. The restaurants compete to attract customers, who then act as judges by choosing restaurants and providing feedback. This setup allows the researchers to examine competitive behaviors and system evolution in detail.
The CompeteAI framework simulates a small-town environment with two competing restaurants and 50 diverse customers over a period of 15 days. Both the restaurants and customers are powered by GPT-4 LLM-based agents, with the restaurants managing their operations through pre-defined actions, and the customers choosing restaurants daily based on provided information. A comprehensive restaurant management system with APIs was developed to allow text-based LLM agents to interact effectively with the simulated environment and overcome implementation challenges. This system incorporates diverse customer characteristics and relationships to trigger authentic competitive behaviors.
When the researchers conducted experiments using the CompeteAI framework, they uncovered valuable insights into competition dynamics and the complex behavior of LLM-based agents. This revealed the use of classic market strategies and multi-factorial influences on customer decisions. The macro-level analysis showed interesting phenomena in the simulated competitive environment, including strategy dynamics, the “Matthew Effect,” and the effect of customer grouping on market outcomes.
In conclusion, the CompeteAI framework offers an innovative approach to studying competition dynamics. The study demonstrated the detailed behaviors of LLM-based agents, the relevance of economic and sociological theories, and the impact of competition on service quality in simulated environments. As this method offers valuable insights and promising results, it could be an invaluable tool for future interdisciplinary research into sociology, economics, and human behavior.
