A survey of multi-agent geosimulation methodologies: from ABM to LLM

This article is a review of multi-agent geosimulation methodologies, ranging from traditional Agent-Based Models (ABM) to Large Language Model (LLM) approaches. Drawing on two decades of research, the authors validate the Agent Reference Model (ARM), a formal framework specifying agent components: internal state (beliefs, goals, intentions, plans, history), internal dynamics, external state, and an interface to their environment. By comparing six multi-agent systems (MAS) methodologies (including a Beliefs-Desires-Intentions, BDI, based-methodology, GAIA and Prometheus, among others) with ARM, the study finds that none includes all the ARM concepts, while ARM encompasses all the features of these methodologies, confirming its generality. ARM has been integrated into the GALATEA simulation platform, which combines discrete event, continuous, and multi-agent simulation under the Discrete Event System Specification (DEVS) formalism. Furthermore, ARM is combined with geographic information systems (GIS) and database models through the Multi-Agent Geosimulation Infrastructure (MAGI) theory, which formalizes the environment, spatial layers, geo-referenced entities, and agent embodiment. This integration enables agents to perceive, act, and reason within realistic geographic spaces. The resulting reference model aims to explain interactions between agents, databases, and GIS as tools for modeling and simulating complex geographic systems. A literature review classifies existing work into cognitive frameworks, generic platforms, geographic automata systems, and generative agents using LLMs. A key finding is that LLMs can be integrated as agent components (perception, planning, action) following the ARM architecture. The authors conclude that ARM, as exemplified by their own implementation in GALATEA for GIS and by other existing systems, provides a formal specification for next-generation geosimulation platforms capable of modeling complex spatial systems with intelligent agents.