Constructing LLM-based Agent Model and its Application to Multi-Agent Simulation

Abstract

There have been active attempts to construct agents that can work in conjunction with large language model (LLM) and apply them to various intelligent systems. In this paper, we describe an attempt to incorporate an agent model using LLM into multi-agent social simulation (MASS). In the implementation of MASS, a problem has been how to construct a computational model to simulate human behavior in the target social environment. Building a model to extract and reproduce the individual characteristics of a wide variety of people has been difficult, including implementation costs. We propose a method to generate a wide variety of likely behavior characteristics from LLM, a kind of collective knowledge, and to realize decision-making according to the surrounding environment. We construct a human flow simulation incorporating an agent based on the proposed method and verify the validity of the behavior.