Integration of Machine Learning Methods into Agent-based Simulations for Predicting Evacuation Time in Disaster Scenarios

Abstract

The behavior of people during an evacuation may have a significant impact on evacuation time, so it has been extensively studied using agent-based simulations. This paper aims to use machine learning for predicting agent evacuation time faster in advance, rather than waiting the entire simulation time. We use the well-known machine-learning polynomial regression as our prediction model, and linear regression and decision tree regression as our benchmark models. In order to generate a suitable dataset for training and validating our models, we automate the scenario-creation process from a single template scenario and the simulation output extraction process in the Vadere simulation framework. Our simulation experiments are carried out using the structure plan of the University of Münster's administrative building, with up to 100 agents located in a source room as individuals and in groups, attempting to find the shortest path to an exit. We significantly improve evacuation prediction using machine learning regression models in agentbased simulation experiments. Our polynomial regression model can predict evacuation time before the simulation begins, and the prediction results are close to the simulation results, with an average R2 score of 84%