An autonomous multi-agent system for disaster simulation and crisis management using LLM orchestration and geospatial modeling.

Abstract

An adaptive multi-agent system (100) for disaster simulation and crisis response, comprising: a. a data acquisition module configured to receive unstructured disaster-related inputs from one or more external data sources, wherein the data acquisition module includes a large language model-based submodule for semantic extraction and is configured to transform the unstructured disaster-related inputs into a structured semantic representation; b. a knowledge base storing geospatial data, infrastructure data, and historical disaster data; c. a retrieval-based generation module communicatively coupled to the data acquisition module and the knowledge base, and designed to retrieve relevant data from the knowledge base and merge the retrieved data with the structured semantic representation to generate a context-aware scenario descriptor;d. a role-based multi-LLM orchestration module configured to assign different-sized language models to respective roles, including data extraction, agent behavior generation, and simulation state evaluation; e. an agent generation module configured to instantiate a multitude of heterogeneous agents from the context-aware scenario describer, each agent being equipped with demographic attributes, mobility restrictions, a vulnerability level, and one or more adaptive behavior parameters; f. a disaster propagation engine configured to model hazard propagation as a probabilistic, graph-based model; g. a simulation engine configured to coordinate the time-graded interaction between the multitude of heterogeneous agents and a modeled environment;and i.e. a geospatial visualization module and a decision support module coupled with the simulation engine and configured to provide real-time monitoring, scenario-based what-if analyses, and predictive decision support results;characterized in that the system (100) further comprises a closed-loop feedback adjustment module configured to receive a current simulation state from the simulation engine, forward the current simulation state together with retrieval-based reference material obtained via the retrieval-based generation module to a language model determined by the role-based multi-LLM orchestration module for the role of simulation state evaluation, and dynamically adjust one or more of the decision guidelines of one or more of the plurality of heterogeneous agents, propagation parameters of the disaster propagation engine, and global parameters of the system (100) depending on an output of said assigned language model.