Data-Driven Behavioral Agent Modeling for Effective Resident’S Evacuation Prediction on Flood Disasters

Abstract

We propose a novel methodology for modeling residents’ evacuation behavior during flood disasters, based on empirical data. Existing evacuation behavior models have not sufficiently evaluated prediction accuracy nor extracted characteristics of residents. Recently, people flow data during disasters has become increasingly available. In this study, we developed a data-driven agent model utilizing resident behavioral data under disaster events. As a result, our models demonstrated a marked improvement in predictive accuracy compared to conventional empirically-based models, especially by incorporating behavioral characteristics, like evacuation to higher ground. Specifically, we constructed both a “data-fitting typed agent model,” which reproduces individual residents’ evacuation decisions regarding occurrence, timing, and destination, as well as a “statistical decision-making agent model,” which incorporates psychological factors and regional characteristics. We quantitatively evaluated the predictive accuracy and compared distinctive features of both models. Experimental results confirmed that the proposed approach reduced prediction error by up to 14.8%.